University
stringclasses 19
values | Text
stringlengths 458
20.7k
|
|---|---|
ADE | Chapter 2 โ Anomaly detection and topological identification
Figure 2-3. Effect of the presence of a branch on a transient pressure head trace.
These figures show that the wave reflections induced by the presence of a topological
element affects the transient pressure signal significantly. The initial effect is either a
drop or an increase in pressure depending on the topological element. If the transient
wave is moving from a smaller diameter into a larger diameter (expansion), the
pressure in the pipeline drops while if a contraction or a branch (where all diameters
are the same) is present, the pressure in the pipeline increases. Figure 2-2 and Figure
2-3 reveal that the presence of simple topological elements can alter a transient
pressure wave in comparison with a single pipeline. In all three cases presented in
these figures, the maximum pressure recorded was larger than the initial transient wave
injected into the system. This behavior has been reported previously demonstrating
the importance of knowing the topology of the system before conducting condition
assessment tests (Wylie 1983; Karney and McInnis 1990; Ellis 2008). For instance,
Bohorquez et al. (2020b) analyzed the superposition of waves from a transient event
with different topological configurations showing that the pressure response can be
magnified if the transient wave propagates into a large hydraulic impedance section,
effectively accumulating head in a pipeline.
Different transient based methods have been applied to the identification of
topological elements in water pipelines, in particular for the identification, location
and characterization of branches. Time reflectometry methods have been implemented
through the analysis of the transient signal directly (Brunone et al. 2008; Meniconi et
al. 2017) or using a wavelet transform (Meniconi et al. 2011b) to confirm the existence
and status of a branch in a single pipeline. These applications have demonstrated that
the existence of a branch induces changes in the transient pressure signal that can be
used for its identification and characterization. Meniconi et al. (2011b) determined the
17 |
ADE | Chapter 2 โ Anomaly detection and topological identification
location of a branch (disregarding its status) by identifying the reflections from this
element in the first 2๐ฟ/๐ seconds of the wavelet transform after the transient test while
the status can be defined based on the transient signal between 2๐ฟ/๐ and 4๐ฟ/๐
seconds after the transient test. Although successful, these applications are based on
the visual inspection of the obtained signals and a manual analysis of the signals is
required.
Inverse Transient Analysis has also been used for the identification of branches and
the determination of pipeline lengths. Kim (2016) proposed a Genetic Algorithm (GA)
approach for the simultaneous location and characterization of a branch, a leak and a
blockage in a numerically modeled pipeline. A numerical model in the frequency
domain was selected to generate numerical transient pressure traces of possible
configurations of the system. This application of ITA proved successful in the location
and characterization of the branch in this pipeline demonstrating that these methods
can be used for topological identification. More recently, Capponi and Ferrante (2017)
proposed the use of the network admittance matrix method, a numerical method to
solve the transient flow equations in the frequency domain, to determine the lengths
of a maximum of three segments of pipelines in a single pipeline with a branch. This
numerical approach proved successful for tests where the length of the upstream
segment of the main pipeline was known. However, for cases where this segmentโs
characteristics were unknown and were included as an unknown in the optimization
process, the success of the ITA was highly dependent on the initial conditions of the
optimization algorithm. In general, the nature of the potential solution space is very
important in the success of an ITA application as optimization algorithms might select
local optima instead of the desired global optimum (Capponi and Ferrante 2017). In
addition, ITA methods can be computationally expensive as they require separate
transient numerical simulations for each trial considered in the optimization process
(Stephens et al. 2004) and often rely on assumptions and simplifications of the
transient numerical model for complex applications (Stephens et al. 2013).
A frequency response method has also been proposed for branch identification in
water pipelines. Duan and Lee (2015) obtained analytical expressions for the
frequency response function of a pipeline with a branch and used it in combination
with an optimization algorithm to detect inactive branches in a numerical pipeline.
This application proved successful in the determination of the branch location and
18 |
ADE | Chapter 2 โ Anomaly detection and topological identification
accurate in defining the length of the branch as long this length was not significant in
comparison with the main pipeline. Although this method does not require a
benchmark frequency response function, its accuracy is limited to certain system
characteristics and it is not fully automatic as an optimization process is needed to
solve the analytical expressions proposed.
2.3 Burst Detection
Water breaks or bursts can be significantly disruptive in water distribution systems;
therefore, locating and characterizing these abnormal events is vital for the correct
functioning of water supply systems. When a burst occurs in a pipeline, a negative
pressure wave travels along the pipeline in both directions, interacting with different
elements in the pipeline and its boundary conditions (Misiunas et al. 2005). For the
case of a reservoir-pipeline-valve system, Figure 2-4 presents the transient pressure
signal observed at the end of the pipeline for two different bursts happening at two
different locations along this pipeline. In both cases, the burst has been modeled as a
circular orifice that is open rapidly (Misiunas et al. 2005). This figure shows that the
negative transient wave generated by the burst is reflected in the system reservoir and
the time of arrival of this reflected wave can be used to determine the location of the
burst. In general, a burst located closer to the pipeline reservoir will induce a drop in
the pressure followed by a quick recovery while a burst located close to the
downstream end will induce a more prolonged low pressure before the reflection from
the reservoir arrives at the end of the pipeline. In addition, the severity of the burst can
be determined by the magnitude of the pressure drop.
In terms of detecting and locating a burst by analyzing the negative transient wave that
this abnormal event generates, some applications have been previously proposed. It is
important to highlight that these techniques address the burst location with passive
monitoring of the pressure behavior in the system. There is no generation of a transient
event via a valve closure, for example, as techniques proposed for anomaly detection
or topological identification.
19 |
ADE | Chapter 2 โ Anomaly detection and topological identification
(a) (b)
Figure 2-4. Effect of the occurrence of a burst on a transient pressure head trace.
Burst located at a) 30%, and b) 85% of the pipeline length.
Liggett and Chen (1994) introduced the idea of using event detection algorithms with
the capacity of detecting the sharp change in pressure generated by the occurrence of
a burst. These detection algorithms would work in conjunction with an ITA approach
to locate the burst by successively modeling these events at different nodes of the
system. Later, Misiunas et al. (2005) proposed a time reflectometry framework to
detect the occurrence of bursts using a two-sided cumulative sum algorithm to detect
the abrupt pressure changes and an offline analysis to determine the location of a burst.
This application proved successful, however, its accuracy depends on the relation
between the opening rate of the burst and the proximity of the burst to the boundaries
of the pipeline (Misiunas et al. 2005).
Wavelet transforms have also been applied to the passive detection of bursts in water
distribution systems. Srirangarajan et al. (2013) used multiscale wavelet analysis to
detect bursts and differentiate them from noise-induced peaks and additional elements
in the system. This approach was combined with a graph-based search algorithm for
the location of bursts based on the wave arrival at different pressure measurement
points. Although the proposed technique was successful in detecting the abnormal
events, the burst location procedure reported errors of up to almost 30% highlighting
the difficulty of locating bursts in complex systems. Finally, frequency domain
methods have been used to detect bursts combining wavelet transforms and the use of
spectrograms to denoise the raw signal and detect the occurrence of a burst (Zan et al.
2011). This method proved to be robust for the detection of abnormal events; however,
the location of the burst was not addressed.
20 |
ADE | Chapter 2 โ Anomaly detection and topological identification
2.4 Summary of Gaps
Considering the existing techniques for leak detection, topological identification and
burst detection using transient pressure signals, a series of gaps in these techniques
have been identified:
- The applicability of some of these techniques is dependent on the availability
of an exact numerical model of the analyzed pipeline. Often these models
might not be available or might be outdated.
- Techniques such as time reflectometry require manual analysis of the raw or
processed pressure signal. This manual interpretation can be difficult when
more complex systems are analyzed or pipelines subject to background noise
are inspected.
- Some techniques can be computationally expensive depending on their
application. These methods involve repetitive optimization procedures that can
easily become impractical, in terms of the amount of time required for the
analysis, if the analyzed problem is too complex. Moreover, if the same system
needs to be analyzed multiple times, procedures need to be conducted
independently increasing the computational effort.
- Methods involving optimization techniques are also sensitive to initial
conditions and may be susceptible to finding a local minimum instead of the
desired global optimum.
- Some existing techniques require partial or complete knowledge of the
behavior of the intact pipeline to determine some necessary parameters or for
comparison purposes. This requirement limits the application of these
techniques in existing pipelines where the intact conditions are not fully
known.
Although existing techniques have proven that using fluid transient waves for pipeline
inspection is possible, automatic techniques that do not require specific information
from the analyzed pipeline and can provide accurate and fast results are still needed.
21 |
ADE | Chapter 3
3 Artificial Neural Networks
In the last 70 years, Artificial Intelligence has emerged as the science and engineering
discipline aimed at making intelligent machines (McCarthy 2007). Developments in
Artificial Intelligence have focused on different ways to simulate human intelligence,
from rule-driven engines based on if-then statements (Bobrow and Stefik 1986) to
more complex applications such as computer vision (Schalkoff 1989; Lemley et al.
2017) and speech recognition (Waibel 1989). Applications of Artificial Intelligence
involving algorithms that allow computer programs to automatically improve through
experience are broadly defined as Machine Learning (Mitchell 1997). Machine
Learning algorithms adjust themselves in response to the datasets they are exposed to
(training datasets) by maximizing the likelihood of their predictions being correct in a
different, unseen, dataset (testing dataset).
Artificial Neural Networks (ANNs) are considered a subfield of Machine Learning
where the transformation of the data representation is conducted by a system inspired
by how neurons are connected in human brains (Cai et al. 2020). An ANN can be
viewed as a learned function, mapping a set of inputs to a set of outputs. The use of
ANNs has proven to be effective, computationally efficient, and robust, provided
enough information for its training exists (Caputo and Pelagagge 2003). In addition,
ANNs have the ability to incorporate new information into previously trained
networks provided that these results come from the same underlying distribution
(Shamir et al. 2010). This chapter presents an overview of ANNs to provide context
for this research. The following includes basic principles behind how ANNs function
and the two main architectures explored in the applications presented in this thesis. In
addition, a succinct literature review of previous applications of ANNs to pattern
23 |
ADE | Chapter 3 โArtificial Neural Networks
recognition, to the use of numerical data for ANN training, and the application of
ANNs to problems in water research is presented.
3.1 Artificial Neural Networks
The concept of using networks of artificial neurons to solve logic problems was first
proposed in 1943 (McCulloch and Pitts 1943). However, it was not until decades later
that their use became feasible when algorithms capable of adjusting the response of
the ANN to a training dataset were proposed (Dreyfus 1990). These algorithms
required extensive computational resources which impeded widespread use of ANNs
until recently, when the fast improving computational power and increase of data
availability has allowed the application of ANNs to more practical problems.
A basic architecture for an ANN is comprised of three elements: a series of inputs, a
hidden layer containing multiple neurons and a series of outputs. The neurons are
densely connected with the inputs by input links and each link is scaled with a real-
number value commonly known as weight. These weights transform and combine the
inputs in each neuron using an activation function embedded in that neuron producing
a real-number value as an output. The process of learning occurs by changing the
weights that connect the input and the neurons using an external stimulus in the form
of the training data containing examples of input-output pairs of the function to be
learned (Aggarwal 2018). ANN architectures can vary in terms of topology. If the
architecture of an ANN contains more than one hidden layer, it is considered a deep
network because the relation between the inputs and the outputs is less direct. The
creation, training and application of these ANN architectures, constitute the field
known as Deep Learning (Marcus 2018).
3.1.1 Fully Connected Dense Neural Networks
Dense neural networks, also known as multilayer neural networks, multilayer
perceptrons or feed-forward networks are neural networks that consist of a fully
connected input layer, multiple hidden layers and an output layer. The information
provided by the input layer feeds the successive hidden layers through the
corresponding weights in the forward direction until the output is predicted (Aggarwal
2018). The key feature of a dense neural network (referred to as a dense network in
this thesis) is that all the layers are fully connected meaning that a neuron in any layer
24 |
ADE | Chapter 3 โArtificial Neural Networks
of the network is connected to all the neurons in the previous and the next layer
(Haykin 1998). Figure 3-1 presents a diagram describing the architecture of a dense
network with two hidden layers.
Figure 3-1. Fully connected dense neural network.
As can be seen in this figure, each element of the input layer connects with each neuron
in the first layer through different weights. Similarly, each neuron of the first hidden
layer is connected with each neuron of the second hidden layer. Finally, each neuron
in the second hidden layer is connected to one of the three neurons that predict the
three outputs for this dense network. Dense networks can embed a lot of information
in their weights, given their large connectivity, and are capable of expressing very
complex functions to the point that, in theory, these networks are able to capture any
arbitrary function (Hornik et al. 1989). However, to accomplish this, a large number
of hidden layers or neurons in a layer would be required. This would translate in a
large number of weights to be trained creating the risk that a network over-fits to the
training data and does not generalize adequately to new data.
3.1.2 Convolutional Neural Networks
Different alternatives to conventional dense neural networks, without the principle of
full connectivity have successfully been proposed as part of the development of
machine learning algorithms. Biological neural networks are connected in ways
humans do not fully understand but in the few cases that the biological structure has
25 |
ADE | Chapter 3 โArtificial Neural Networks
been understood, significant advances have been achieved by designing ANNs using
those principles. One example of this strategy are convolutional neural networks. This
networks are inspired in the organization of the neurons in the visual cortex of cats
(Aggarwal 2018). Convolutional neural networks have proven successful in tasks such
as image classification (Ciresan et al. 2011), object detection (Zhu et al. 2015) and
signal processing (Kiranyaz et al. 2019). Most of the applications of convolutional
neural networks involve a 2-dimensional input (e.g. an image), however, signal
processing and pattern recognition applications, including those presented in this
thesis, often use 1-dimensional inputs. A diagram describing the architecture of a 1D-
convolutional network is presented in Figure 3-2.
Figure 3-2. Convolutional neural networks.
The primary difference between a convolutional network and a dense network is that
a convolutional layer consists of a set of learned filters. A filter is a learned template
in the form of a set of weights which map a small window of the input into a
corresponding window in the next layer of the network. To map an entire image, a
filter is slid across its entire input signal to produce a corresponding signal for the next
layer. In training, each filter learns a specific shape or pattern (feature) in its input that
is useful for the task at hand. In many convolutional neural networks there are multiple
convolutional layers, each with their own filters. These networks typically step down
in terms of the resolution of the signal in the deeper layers. This means that filters in
deeper layers capture progressively larger portions of the initial input signal. Another
characteristic of the convolutional networks is that the last few layers (usually three)
26 |
ADE | Chapter 3 โArtificial Neural Networks
are dense layer or fully connected layers where the information processed through the
convolutional layers is condensed into the outputs.
The connectivity characteristics of convolutional networks provide these networks
with the ability to capture local interactions between data points in the input which can
potentially work well in applications where there is potential to exploit locality of
features in the input data. Because the neurons in convolutional networks are not fully
connected, for a given number of input nodes, these networks have many less weights
and thus are less prone to over-fitting.
3.2 Artificial Neural Networks Training
The process of training an ANN is a multivariable optimization problem where the set
of weights that connects the layers of the ANN is computed to minimize an error
function between the real and the predicted outputs. Multiple optimization algorithms
have been used to train ANNs (Karaboga and Akay 2007; Kawam and Mansour 2012)
but one of the most widely used algorithms is Stochastic Gradient Descent (SGD).
This algorithm updates the weights by moving along the negative direction of the
gradient (calculated from a random batch of the complete training dataset) seeking for
the global minimum of the error function (Aggarwal 2018).
Although the SGD algorithm has a number of advantages in comparison to other
optimization algorithms such as fast execution times, it also has some disadvantages
related primarily to the possibility of finding a set of weights around local minima of
the error function. Considering that one execution of this algorithm can be completed
in a relatively short period of time (depending on the number of weights to train), the
iterative process runs through multiple executions of the SGD algorithm until the error
function drops below a defined threshold.
This iterative training approach is not considered a standard practice in the machine
learning field; however, this approach based on a 1+1 Evolutionary Strategy has
proven effective in previous applications (Beyer and Schwefel 2002). Because each
time that the SGD algorithm is executed the training dataset is divided differently to
achieve the final set of weights, each iteration will result in a different set of weights.
By preserving the best result achieved so far until the threshold is met or the maximum
27 |
ADE | Chapter 3 โArtificial Neural Networks
of iteration is reached, the best set of weights is retained at the end of the training
process.
3.3 Artificial Neural Networks for Pipeline Inspection using Fluid
Transients
Artificial neural networks have been used in a wide range of data-intensive fields,
including machine diagnostics (Samanta and Al-Balushi 2003), credit rating (Huang
et al. 2004), and facial recognition (Ming and Fulcher 1996), amongst many others
(Abiodun et al. 2018). However, to the knowledge of the author of this thesis, deep
artificial neural networks have not been previously used for the interpretation of
transient pressure traces for the inspection of pressurized water pipelines. This section
presents selected past applications of ANNs as examples to provide context to the
methods and applications presented in this thesis.
Mounce and Machell (2006) proposed the use of two artificial neural network
architectures (static ANN and time delay ANN) to detect the occurrence of bursts
using flow data at a DMA level showing potential for identifying changes in the flow
that corresponded to unusual fluctuations. Similarly, Mounce et al. (2010) proposed
the use of support vector regression models to predict time series data in a moving
time window and compared these series with measured data for the detection of
anomalies. The use of this supervised learning technique was applied to historical data
proving that 78% of the alerts corresponded to actual abnormal events in the system.
These applications confirm the potential of merging hydraulic measurements and
machine learning algorithms for the detection of system anomalies. However, the
frequency of sampling for these applications is not applicable to a transient analysis.
Other researchers have proposed techniques for the detection of leaks in liquid
pipelines using support vector machine models (SVM). Ni et al. (2013) showed that
SVM models can predict the occurrence of leaks more accurately than traditional
multi-layer perceptrons (fully connected ANNs with three layers) when applied to the
numerical model of an oil pipeline. Similarly, Li et al. (2018) proposed the use of SVM
algorithms to detect the occurrence of a leak on a moving window in combination with
a time reflectometry approach to determine the location of the leak. Both applications
proved successful when applied to numerically modeled pipelines, although the
28 |
ADE | Chapter 3 โArtificial Neural Networks
datasets used to train these models were small and difficult to reproduce in more
realistic conditions.
Deep artificial neural networks have not been used to interpret transient pressure traces
for condition assessment of pipelines, yet other applications of pattern recognition
have been reported. For instance, Kiranyaz et al. (2015) proposed the use of 1D-
convolutional networks to monitor and detect patient-specific electrocardiograms
(ECG). Using a combination of data from general databases and patient-specific ECG
records, a 1D-convolutional network has been successfully applied to the detection of
anomalous heartbeat based only on ECG measurements. This application highlights
the ability of 1D-convolutional networks in interpreting time series to identify and
classify anomalies.
Other relevant applications include the use of physical models for the training of an
ANN. Hajgatรณ et al. (2020) trained a dueling deep q-network with hydraulic
simulation data to optimize pump operations (defining pump speeds) using only
measured flow data at the time the method is applied demonstrating the capability of
deep learning techniques to solve problems in near real time. Similarly, Yu et al.
(2019) successfully used the numerical model of a benchmark building to obtain
vibration data for the training of a convolutional network to predict the location and
the severity of damage in the building. Although this application was not validated in
an experimental setting, it proved that is possible to train an ANN based on
information obtained from a physical model.
Previous studies have shown that deep artificial neural networks can be used to
interpret flow and pressure in water distribution systems for the detection and broad
location of abnormal events. In addition, applications in different fields have
demonstrated the ability of deep artificial neural networks to identify patterns in time
series. However, deep artificial neural networks have not been applied for the
identification, location and characterization of anomalies and abnormal events in
pressurized pipelines.
29 |
ADE | Chapter 4
4 Synopsis of Publications
The overall aim of this thesis is to develop and apply data-driven techniques for the
active and passive inspection of water pipelines using fluid transients and Artificial
Neural Networks (ANNs). This chapter discusses the relationship between the three
journal publications that have arisen from this research with the proposed aims and
summarizes their content and contributions as a way of introduction to the following
chapters. To discuss the connection between the journal publications and the research
aims, Figure 4-1 illustrates the contributions of each publication to the fulfilment of
the research aims presented in Figure 1-1.
The development of a framework to combine the use of fluid transient waves and
Artificial Neural Networks for the condition assessment of water pipelines (Aim 1) is
presented in Journal Publication 1. This framework has been described in all
publications; however, a more in-depth definition and analysis have been included in
Journal Publication 1 applied to the identification of the location and size of a junction
in a pipeline as a stepping stone in the fulfilment of the overall aim of this research.
Two novel techniques are proposed in this research that use Artificial Neural Networks
to interpret transient pressure signals obtained from a pipeline. The methodology for
the active inspection of pipelines (Aim 2) comprising the location of anomalies or
topological elements after the generation of an artificial transient event (Aim 2.1) is
described in Journal Publication 1 and Journal Publication 2. First, Journal Publication
1 explores the training and testing of a specific type of Artificial Neural Network for
the prediction of the location and size of junctions and leaks in a numerical pipeline.
31 |
ADE | Chapter 4 โ Synopsis of Publications
A more complete approach is then described in Journal Publication 2, where a
comprehensive methodology for the active inspection is presented with a novel
training framework deploying stochastic resonance to allow for the location of leaks
in pipelines where background pressure fluctuations are present and could interfere
with the ANN performance (Aim 2.2).A technique for the passive inspection of
pipelines for the detection of abnormal events (Aim 3) is included in Journal
Publication 3. This publication includes a methodology for the interpretation of a
continuously measured transient pressure signal (Aim 3.1) and a complete
methodology for the detection, location and characterization of bursts (Aim 3.2).
Finally, as can be seen in Figure 4-1, the three journal publications developed as part
of this research contribute to the validation of the new techniques (Aim 4). Journal
Publication 1 presents a numerical application of the active inspection methodology,
while Journal Publication 2 is focused on the experimental application of this
methodology. Regarding the passive inspection methodology, Journal Publication 3
addresses both the numerical and the experimental validation of this approach. A
summary of each journal publication is now presented.
Journal Publication 1 (Chapter 5) first describes the formulation of a framework to
use the transient pressure response of a pipeline after the generation of a controlled
small magnitude transient event in combination with ANNs for the detection of
different features in a single pressurized pipeline. Different elements of this
framework are described including the definition of an appropriate ANN architecture
to be applied in the recognition of feature-induced transient wave reflections and a
methodology for the generation of numerical transient pressure traces used as training
and testing samples. This framework is then applied to two separate cases: 1) the
location of a junction, defined as a change in the pipeline diameter and 2) the location
and size of a leak in a single pipeline, as an example of a pipeline anomaly. A
schematic representation of this framework applied to the active inspection of
pipelines included in Journal Publication 1 is presented in Figure 4-2.
A transient pressure trace signal is measured after the generation of a small magnitude
controlled transient event such as the closure of a side discharge valve (represented in
Figure 4-2 by the blue line at the top). This pressure trace is then downsampled (shown
in the figure with the red circles on top of the original pressure trace) and used as the
33 |
ADE | Chapter 4 โ Synopsis of Publications
input of a particular ANN. Finally, the downsampled transient pressure trace is then
processed through a particular ANN to obtain a prediction of the location and the size
of the analyzed feature.
Figure 4-2 Artificial Neural Networks used for active inspection of pressurized
pipelines. Adapted from Figure 5-3 in this thesis.
Results in this publication demonstrate that the trained ANNs are able to accurately
predict the location of the junction or the leak along the pipeline when presented with
numerical traces that have not been used for the training. The prediction of the size of
the junction (representing the two diameters in the pipeline on either side of the
junction) is almost perfect once a rounding process is conducted as in only one case
(out of 2,500) the final prediction of the diameters was inaccurate. For the case of the
leak, the size is on average predicted within 0.03 mm of its real size.
The main contribution of Journal Publication 1 is that for the first time, numerical fluid
transient traces obtained after the closure of a side discharge valve are interpreted by
ANNs to accurately locate topological elements or anomalies in a pipeline. The
developed framework and the results presented in this publication demonstrate that is
possible to train an ANN with numerical transient pressure traces to identify the
reflections caused by the presence of a junction or a leak in a single pipeline and
accurately predict their characteristics. In addition, the results of this publication
34 |
ADE | Chapter 4 โ Synopsis of Publications
indicate that by combining knowledge of the propagation of the transient wave after a
transient event (a key element in the definition of the framework), with the versatility
of performance of an ANN, an accurate and fast pipe inspection technique can be
developed.
Journal Publication 2 (Chapter 6) presents a comprehensive methodology for the
active inspection of pipelines based on the framework presented in Journal Publication
1. In cases where the transient pressure signal obtained from the analyzed pipeline
contains background pressure fluctuations, the framework presented in Journal
Publication 1 is not able to accurately locate anomalies such as leaks. This situation
arises because the pressure fluctuations are not reproduced by numerical transient flow
models affecting the accuracy of the anomaly location prediction of the ANNs. The
methodology presented in Journal Publication 2 enhances the ANN performance in
detection of leaks in pipelines via deployment of stochastic resonance. Stochastic
resonance is a phenomenon where the performance of a non-linear system is optimally
enhanced by the addition of a certain noise intensity (Harmer et al. 2002). This
phenomenon is applied in the active inspection methodology presented in Journal
Publication 2 through the creation of a set of ANNs that are trained with numerically
generated transient pressure traces containing artificial noise with different intensities
to determine the optimal noise intensity that enhances the 1D-convolutional neural
networks performance.
A diagram of a set of ANNs presented in this publication is shown in Figure 4-3. In
addition to train multiple ANNs with transient pressure traces containing different
noise intensities, multiple ANNs are trained with the same dataset to test the
robustness of the selected ANN architecture considering that different training
attempts using Stochastic Gradient Descent algorithms will produce a different set of
weights in each ANN. From the results obtained from the training of these ANNs and
its application to a pipeline, an optimum noise intensity has been determined.
35 |
ADE | Chapter 4 โ Synopsis of Publications
Figure 4-3. Set of ANNs used for active leak detection via deployment of stochastic
resonance.
This methodology has been applied in a real pipeline in the Robin Hydraulics
Laboratory of the University of Adelaide where 14 transient tests were conducted.
Results from the application of this methodology show that the addition of noise in the
transient pressure samples is fundamental for the enhancement of ANN predictions
for the location of a leak highlighting the existence of an optimum noise intensity to
obtain accurate and reliable results. The application of this methodology has allowed
for, the very accurate location and characterization of a leak in this laboratory pipeline.
The main contribution of Journal Publication 2 is the development of a comprehensive
methodology for the active inspection of water pipelines by adapting the framework
proposed in Journal Publication 1 to accurately locate leaks in single pipelines under
more realistic conditions where background pressure fluctuations are present in the
pipeline. This publication demonstrates that the deployment of stochastic resonance
assists in detecting leaks in water pipelines by showing the existence of an optimum
intensity of noise to be added to the numerical pressure transient traces for the training
of a set of ANNs. With the addition of noise in the training samples of an ANN, its
performance is significantly improved, to the point that consistent and accurate
predictions can be obtained.
36 |
ADE | Chapter 4 โ Synopsis of Publications
Journal Publication 3 (Chapter 7) describes a novel methodology for the
identification, location and characterization of the occurrence of bursts in a single
pipeline using 1D-convolutional neural networks for the interpretation of a continuous
transient pressure signal measured at one point along the pipeline. The main difference
of this methodology with the one presented in the previous publications is that given
that no artificial transient event is induced in the pipeline, this methodology focuses
on the passive and continuous inspection of the transient pressure signal. As shown in
Section 2.3, the transient pressure traces caused by the occurrence of a burst are more
prominent compared to the transient pressure traces obtained in the active inspection
technique for a pipeline with a leak (Journal Publication 1 and 2). This characteristic
makes these signals less sensitive to background pressure fluctuations and therefore
stochastic resonance principles are not applied in the passive inspection methodology.
However, background pressure in the pipeline is considered in this publication by
modeling different sinusoidal background pressure fluctuations before the occurrence
of a burst.
A diagram of the use of two different ANNs for the passive detection and identification
of a burst in a pipeline is presented in Figure 4-4. A sliding transient pressure time
window analysis is used in this methodology to successively analyze one time window
at a time. A burst detection ANN is used to classify each time window into three
possible categories to determine whether a burst has occurred. Once the burst is
detected, a burst identification ANN is able to predict its location and the size.
Journal Publication 3 includes a numerical application of the methodology proposed
where a sharp burst occurring along a 1000-m long numerical pipeline is considered.
This publication also presents an experimental validation of the proposed technique in
a real pipeline at the Robin Hydraulics Laboratory of the University of Adelaide. The
results from this validation indicate that the prediction of the location of the burst is
very accurate while the prediction of the burst size requires an additional step to ensure
its accuracy. To obtain the final burst size prediction, Journal Publication 3 includes
the application of a burst size adjustment procedure to obtain a more accurate final
prediction.
37 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
Foreword
This chapter presents the formulation of a framework to use the transient pressure
response of a pipeline after the generation of a small magnitude controlled transient
event in combination with ANNs for the detection of different features in a single
pressurized pipeline. This framework is applied to two condition assessment
problems: the location and characterization of a change in diameter in a single pipeline
(referred to as a junction) and a leak. To complement the information presented in the
publication and provide further context to this chapter, two elements are included here.
The importance of the selection of an appropriate ANN architecture for the location
of junctions is briefly described, followed by a description of the computational
resources necessary for the training of the ANNs.
The definition of an appropriate ANN architecture has proven to be a key
consideration for the creation of the framework presented in this chapter. Two
different ANN architectures were considered including a three-layer dense network
and a 1D-convolutional neural network. The characteristics and differences between
these two architectures have been discussed in Chapter 3 (Section 3.1) and in Section
5.3 of this chapter highlighting that while a dense network has been widely used in
previous applications, it lacks robustness and versatility when complex problems are
considered.
To demonstrate this, Figure 5-1 presents the error for the ANN prediction of the
location of 5,000 different junctions distributed along the total length of a numerically
modeled pipeline for two ANNs: a dense network (with 88,113 weights to train) and
a 1D-convolutional neural network (with 32,409 weights to train). This figure shows
that selecting a dense network as an ANN architecture is not robust enough to
accurately locate a junction in a numerical pipeline. Large errors are found for
junctions located in both extremes for the pipeline and no consistency is found in the
behavior of these predictions. This figure also shows that the junction location
predictions of a 1D-convolutional network are more consistent along the pipeline in
comparison to the predictions of the dense network and present fewer oscillations
41 |
ADE | Abstract
Condition assessment of water pipelines using fluid transient waves is a noninvasive
technique that has been investigated for the past 25 years. Approaches to identify
different anomalies and to identify elements of the topology of a pipeline have been
proposed but often require detailed modeling and knowledge of the system. On the
other hand, artificial neural networks (ANN) have become a useful tool in a range of
different fields by enabling a computer to solve a problem without being explicitly
programmed to do so, but rather by learning from a series of known examples. This
paper presents a new methodology that uses ANNs to predict the presence of features
in a pipeline. First, the location and characteristics of a junction have been predicted
as a way to identify elements of the topology of a pipeline followed by identification
of the location and sizing of a leak. The ANN characteristics and training approaches
have been determined for both the junction and the leak example. Results show that
the ANN that has been designed for this research is able to accurately predict the
location of a junction with an error in this estimation of 2.32 m (out of a 1,000 m long
pipeline) or less in 95% of the tested cases. The prediction of the two different
diameters on either side of the junction was extremely accurate with only one
misidentification of one of the diameters in the 5,000 tested examples. When the ANN
was trained and tested to locate and size a leak, the results were also successful. A
total of 95% of the tested examples located the leak with an error equal or less than
3.0 m (out of a 1,000 m pipe length) and the leak size was predicted with an average
absolute error of only 0.31 mm. The results presented in this paper demonstrate the
potential of combining the use of both fluid transient pressure waves and ANNs for
the detection of features in pipelines.
45 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
5.1 Introduction
Water is a vital resource to society and its continuous distribution for all types of use
is becoming a challenge, including time and spatial inequality in its availability, and
the lack of responsible management and poorly maintained infrastructure. In addition,
pipes in water distribution system networks are often underground, which complicates
their monitoring and maintenance. To overcome this, various noninvasive techniques
have been developed to monitor and detect faults in pipelines for accomplishing an
efficient and cost-effective maintenance. These methods include visual inspection
(Guo et al. 2009), electromagnetic methods (Wang et al. 2012), acoustic methods
(Juliano et al. 2013), ultrasonic, radiographic, thermographic methods (Zheng and
Yehuda 2013), and more recently transient-based inspection (Lee et al. 2008; Gong et
al. 2014a; Gong et al. 2018a).
Of these different techniques, transient-based methods have received special attention
in the past two decades given that they allow the inspection of large sections of a pipe
with a relatively simple set up (Gong et al. 2013c), and results can be obtained quickly
(Lee et al. 2006; Shi et al. 2017). These methods are based on the interpretation of the
effect that any feature in a pipeline has on the transient head trace, when a small
controlled artificial transient pressure event is generated. To detect faults by using
transient pressure signals, several methods have been proposed and can be organized
in three groups according to Colombo et al. (2009): (1) inverse transient techniques,
(2) frequency domain techniques, and (3) direct transient methods. However, whereas
each of these approaches has been moderately successful, they also have associated
disadvantages in terms of processing time or required knowledge of the analyzed
system.
The current paper presents an innovative transient-based technique that uses artificial
neural networks (ANNs) to identify topological elements such as junctions in water
pipelines networks and to locate and characterize leaks. Unlike previous studies, the
proposed technique is data-driven because it does not need any detailed information
with regard to the analyzed water network system. When the ANN is tested, it gives
successful results using just the transient head trace measured at the transient
generation location. Previous techniques of transient-based pipe condition assessment
47 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
have been model driven because they have always had to prepare a detailed simulation
model of the pipeline to be analyzed.
The technique proposed in this paper uses transient head data for the training of the
ANN that can be obtained either from historical data or from an available model and
is data-driven in its application because it only requires measured transient head data
to locate and characterize the desired features. In addition, the computational effort of
the proposed technique is concentrated in the ANN training stage, but once this stage
is complete, the technique can process and find leaks and/or topological features from
different transient head traces almost immediately, without the need of retrain the
ANN. Considering that this is the first reported joint application of fluid transients and
ANNs, the examples and the functioning of the technique are demonstrated only with
numerically derived data. However, the promising results of this first application
prove the promising potential of this approach and offer insights for future validations
(with experimental and field data) and applications in more complex systems.
The methodology proposed in this paper is applied to the location and characterization
of different features in a single pipeline. It is shown that the technique is successful in
accurately predicting the location of the feature within the pipeline and its size, and it
proves the potential of exploring the joint use of fluid transients and ANNs in more
realistic scenarios. The ANNs were trained with numerical data generated using a
method of characteristics (MOC) transient model and its performance has been tested
also with different sets of numerical data showing that the ANN is able to identify new
features that were not used in its training.
Two hydraulic systems have been considered. The first one is a pipeline with two
segment lengths of different diameter (referred to as a junction system), and the second
one was a single pipeline with the presence of a leak. The junction system was used
to determine the most appropriate structure and characteristics for the ANN given that
is a simpler system and the response of the transient head trace is easier to identify in
comparison to a leak in a pipeline. In addition, the identification of a junction was
selected as an example of topology identification because its location and the
diameters of the two segments are predicted.
The present paper provides a background in transient-based methods for identifying
topological elements and for locating leaks based on transient pressure traces,
48 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
highlighting certain limitations of each method that can potentially be overcome with
the use of ANNs. A summary of the key elements of ANNs that were used for the
present technique are then discussed, including the different configurations
considered. The hydraulic systems selected are defined next, and the ANN training
settings are described. Finally, results for the location and sizing of a junction (sizing
referring to the determination of the two diameters at the junction in the pipeline) and
for the location and sizing of a leak are presented. The proposed method is shown to
be accurate and general for the considered systems; however, in addition to some
concluding remarks, some challenges of the future application of this technique are
briefly discussed.
5.2 Background
Fluid transient waves are an interesting mechanism for detecting existing faults,
features in pipelines, and different elements of the topology of a water system
(Bohorquez et al. 2018). However, this technique is highly sensitive to multiple system
characteristics, and understanding what the pressure signal response should look like
when a specific fault is present in realistic cases is often challenging (Xu and Karney
2017). Different authors have applied techniques using fluid transients to detect
topological elements and to locate leaks in pipelines.
First, different techniques have been proposed to identify junctions, branches, and
partially closed valves in pipelines. Brunone et al. (2008) applied time reflectometry
to confirm the existence of a Y junction in a water main pipe in Italy. By interpreting
the arrival of the transient wave reflection at the Y junction when it arrives at the
measurement point, the precise position of the Y was computed accurately. Although
successful, this technique requires the visual analysis of the transient pressure trace to
determine the arrival of the reflected wave and depending on the system, can fail in
terms of both precision and reproducibility.
Meniconi et al. (2011b) developed a laboratory test to locate branches on a system
using a wavelet analysis and a reflection coefficient for estimating the size of the
branch. Their results showed that a visual inspection of the wavelet transform is a good
mechanism to locate branches, but the damping of the transient pressure waves
affected the use of the reflection coefficient due to the pipe viscoelasticity and
unsteady friction. In addition, to determine the condition of the branch (active or
49 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
inactive), this approach used a MOC model to compare with the measured pressure
traces.
Second, detection of leaks has received special attention given that water losses affect
the performance, diminishes the operational age of a water supply system, and
represents economic and environmental losses for a water utility (AL-Washali et al.
2016). The possibility of detecting single and distributed leaks by analyzing a transient
pressure trace has been analyzed by different authors that explored techniques that can
be divided in three groups: (1) inverse transient techniques, (2) frequency domain
techniques, and (3) direct transient methods (Colombo et al. 2009).
Inverse transient analysis (ITA) techniques are based on the calibration of a numerical
model in terms of the existence and characteristics of leaks to match the measured and
the numerical transient pressure trace(Colombo et al. 2009). The first application of
ITA and transient pressure was proposed by Liggett and Chen (1994) but a number of
researchers have used ITA for detecting leaks in pipelines by modifying the
optimization problem, the pressure measurements, and the selected algorithms (Covas
et al. 2001; Vรญtkovskรฝ et al. 2003b; Soares et al. 2011; Kim 2014; Capponi et al. 2017).
Despite the accuracy of ITA methods in the location of leaks under different scenarios,
they require a detailed numerical model of the analyzed system, they can report
discrepancies in leak size estimations (Covas et al. 2001), and an optimization model
needs to be executed each time a transient pressure trace is analyzed.
A second group of methods for locating leaks are the frequency domain techniques,
which are often associated with determining the frequency response function of the
system and comparing it with the one in a pipe without any anomalies(Gong et al.
2016a). A first numerical application of these techniques was proposed by Mpesha et
al. (2001) and since then, different approaches have been developed emphasizing the
theory, procedure, and application of this technique (Lee et al. 2005; Sattar and
Chaudhry 2008; Duan et al. 2010; Ghazali et al. 2010; Gong et al. 2013a; Gong et al.
2016a; Duan 2017). However, frequency domain techniques require extensive
measurements in the field and the frequency response function of the intact pipe,
which can be obtained numerically using a detailed model (Colombo et al. 2009).
A third group of methods has been defined by Colombo et al. (2009) as direct transient
methods. These methods aim to identify special features in the transient pressure trace
50 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
that are induced by the presence of the leak. This group includes techniques such as
time reflectometry (Brunone 1999; Lee et al. 2007a; Brunone et al. 2008; Guo et al.
2012; Lazhar et al. 2013), transient signal damping analysis (Wang et al. 2002; Nixon
et al. 2006; Brunone et al. 2018), and wavelet analysis (Ferrante and Brunone 2003;
Ferrante et al. 2007; Ferrante et al. 2009; Meniconi et al. 2011a; Brunone et al. 2013).
These methods have been successful in the detection and characterization of leak in
pipelines; however, the physical characteristics of the pipeline are required, and some
approaches depend on knowing the transient pressure trace of an intact pipeline. In
addition, experience is required to interpret the resulting pressure trace due to the
existence of vibrations, background transients, and instrument noise (Colombo et al.
2009).
Although the majority of previous techniques have achieved satisfactory results in the
task of detecting elements of the system topology or locating a leak using pressure
transients, these methods require information about the analyzed pipeline (physical
characteristics and resulting transient pressure traces after installation) or a detailed
numerical model (model-based techniques). In addition, to obtain results extensive
field tests may be needed, or a significant computational time is required for the
analysis of each transient pressure trace obtained from a test. Thus, there is an
advantage in using data-driven techniques that can provide accurate results fast and
with the potential of applications to different water pipelines configurations.
ANNs have had a broad set of applications in different fields, and they have become
a powerful tool in machine learning. The use of ANNs has proven to be effective,
computationally efficient, and robust, provided enough information for its training
exists (Caputo and Pelagagge 2003). In different applications, it has been
demonstrated that results from ANNs are robust and are not sensitive to background
noise in the analyzed system (Carlini and Wagner 2017; Mangal et al. 2019). In
addition, ANNs have the ability to incorporate new information to previous training
results provided that these results come from the same underlying distribution (Shamir
et al. 2010). In water-related research, ANNs have been successfully applied to a
variety of problems including water availability under climate change scenarios
(Swain et al. 2017), asset failure prediction (Harvey et al. 2014), water demand
prediction (Guo et al. 2018), cyber-physical attacks location (Taormina and Galelli
51 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
2018), and burst detection in water distribution systems (Romano et al. 2014), among
many others.
Nonetheless, the use of ANNs for the identification of elements in the topology of a
system, for leak detection and in general, for pipeline condition assessment using fluid
transients, have not been previously explored to the knowledge of the authors. Only
one application was reported by Belsito et al. (1998) where the location of leaks was
predicted using ANNs in liquefied gas pipelines for different leak sizes. Pressure
signals were used in their research as input information for the ANN, but transient
waves were not considered. The results showed that the ANN was able to locate the
leak accurately in more than 50% of the cases, for a leak size of 1% of the base flow.
The research presented in the current paper is the first combined use of ANN and fluid
transient waves to identify topological aspects and to detect and characterize leaks in
water distribution system pipelines.
5.3 Artificial Neural Networks
An ANN implements a mathematical function (model) from n inputs to m outputs, and
this function is represented by a mathematical graph of connections and nodes linking
the input to the output. An example network is shown in Figure 5-2. The inputs to the
ANN are a vector of numerical values; these values are transmitted through the links
of the graph to activation functions, which represent neurons. All links in the graph
have an associated weight which is used to scale the value traveling on that link. Each
activation function transforms the sum of the weighted values it receives, to an output
value that is then propagated through the network. Thus, the input values are
transformed by traversing the weighted links and the activation functions in the graph
until they reach the output links.
To produce the desired behavior, an ANN needs to be trained. The training process
modifies the weights associated with each of the links in the network to improve the
accuracy of the model represented by the network. In theory, through modification of
weights alone it is possible for a network of at least three layers, with enough links, to
approximate an arbitrary function (Cybenko 1989). However, deep networks (with
more than three layers) tend to be more practical to train, more versatile, and less prone
to over-fit (Urban et al. 2016). The extent to which this approximation succeeds is
determined by the interaction between the network architecture and its training
52 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
process. In practice, training the network is a process of mathematical regression
where a gradient search is used to adjust the weights in the network to minimize the
error between the actual outputs of the network and the desired output. To be useful
in the desired application domain, a network will approximate the required function to
a high level of accuracy on both the data it was trained with and any new test data that
it is presented with.
Figure 5-2. Artificial neural networks (ANNs) overview.
In the past decade, the technology used to train, specify, and implement ANNs has
advanced rapidly. As a result, modern ANNs present the designer with a very broad
range of design decisions for the ANN relating to topology, scale, activation functions,
regularization strategies, and training methodology. As a general rule, the designer of
an ANN has to use a network design that captures the behavior of the desired function
without having so many weights (parameters) as to over-fit the data used in the training
process.
In the application described in this paper, the functions that are approximated by the
ANNs transform an input in the form of a transient head time series into continuous
scalar values representing the location and the sizes of features in a pipeline as the
output, as shown in Figure 5-3. Thus, in this application, the input data (the transient
head time series) may be quite large, and the resulting relationships to be learned are
then quite complex. Two architectural options for the ANN have been explored in this
paper. The most general available ANN structure was evaluated first, which
corresponds to a fully connected dense network. Such networks are, in theory (Hornik
et al. 1989), able to capture arbitrary functions. Subsequently, the use of a 1D-
convolutional network was explored. A-priori convolutional networks are known to
53 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
be easier to train and capture structural invariants in their input signal more reliably
than fully connected networks (Lecun et al. 1998).
Figure 5-3. An ANN applied to pipe condition assessment.
A dense network, an example of which is shown in Figure 5-4(a), connects each
neuron in a layer to every neuron in the subsequent layer. Dense networks embed a lot
of information in their weights and can express quite complex functions. However, the
large number of weights in dense networks create the risk the network will over-fit the
training data and not generalize adequately to new data.
a) b)
Figure 5-4. Architectural options of ANNs: (a) dense network; and (b) 1D-
convolutional network.
The second architectural option used in this paper is a 1D-convolutional network
[Figure 5-4(b)]. These networks link each neuron in a layer to neurons in the
corresponding neighborhood in the subsequent layer. Convolutional networks capture
local interactions between data points and can work well in applications where there
54 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
is potential to exploit spatial locality of features in input data. For a given number of
input nodes, convolutional networks have many less weights and thus are less prone
to over-fitting. Additionally, these networks have been successful in problems that
require a deep configuration with a faster and easier training when compared with a
dense network.
5.4 Hydraulic Systems Configuration
An ANN requires data as a training set to determine its characteristics and the
functions behind its performance before being used for predicting a result. To generate
the input data for detecting a junction or a leak (depending on the case), repetitive
numerical simulations of the transient response of a system to a valve closure have
been conducted using a MOC numerical simulation for transient behavior while
changing the location and characteristics of the analyzed feature. This section
describes the systems used for obtaining the input data for: (1) identifying junctions
as a topological element, and (2) for detecting the presence of leaks.
5.4.1 Junction Model
The system considered for applying an ANN to identify topological elements is a
junction in a single pipeline with two different diameters as described in Figure 5-5.
The pipeline is connected at the upstream end to a reservoir with a fixed head ๐ป and
0
on the downstream end to a side discharge valve. The length of the pipeline is
fixed (๐ฟ ), the length of the upstream segment of the pipeline of diameter (๐ท ) is
๐ 1
defined as ๐ฅ, and the length of the downstream pipe segment of diameter (๐ท ) is
2
then (๐ฟ โ๐ฅ).
๐
Figure 5-5. Junction system description.
Steady-state conditions of the system were fixed for all the transient simulations. An
initial head and an initial velocity in the pipeline (at the valve) were defined as ๐ป =
0
55 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
70 m and ๐ = 0.15 m/s, respectively. Only steady-state friction was considered by
0
using a DarcyโWeisbach friction factor, ๐, with a pipeline roughness height of ๐ =
0.01 mm. The total length of the pipe was as assumed to be ๐ฟ = 1,000 m.
๐
The detection of the junction includes the prediction of both its location (by predicting
the length of the upstream pipe segment) and the combination of diameters on either
side of the junction in the pipeline. To accomplish this using the ANN, input training
data included simulations of 10 different combinations of diameters on either side of
the junction. These combinations were defined according to the Australian/New
Zealand standard for ductile iron pipes with cement mortar lining and are presented in
Table 5-1. Different wall and cement mortar lining thicknesses were considered for
the different diameters. It is important to highlight that in the 10 combinations of
diameters, five correspond to flow going from a larger to a smaller (๐ท > ๐ท )
1 2
diameter, whereas five correspond to flow going from a smaller to a larger
diameter (๐ท < ๐ท ).
1 2
Table 5-1. Outer diameters combinations for detection of junctions (Standards
Australia 2014).
Combination ๐ท /๐ท ๐ท (๐๐) ๐ท (๐๐)
1 2 1 2
1 1.25 750 600
2 1.50 750 500
3 1.2 600 500
4 1.33 600 450
5 1.11 500 450
6 0.9 450 500
7 0.75 450 600
8 0.83 500 600
9 0.67 500 750
10 0.80 600 750
As it was mentioned previously, generation of the ANN input data (transient head
traces at the valve after its closure) to train and test the ANN was accomplished by
running multiple numerical simulations of the system using a conventional MOC. For
each of the diameter combinations, the length of the upstream segment of the pipeline
was changed along the complete length of the pipeline to simulate different locations
of the junction. The number of generated locations, selection of these locations, the
time step, and length of each reach used in the MOC are described later in this paper.
56 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
5.4.2 Leak Model
A single pipeline was also selected for locating and sizing a leak as is shown in Figure
5-6. The location of the leak is determined by a distance, ๐ฅ, measured from the
upstream reservoir, and the total length of the pipe is defined as ๐ฟ . The diameter of
๐
the pipeline (๐ท) was fixed for all the numerical simulations, and the diameter of the
leak was defined as ๐ท .
๐ฟ
Figure 5-6. Leak system description.
The steady-state head and velocity were the same as the junction system. A ductile
iron pipe with cement mortar lining is considered with an internal pipe diameter of
727.5 mm, a ductile iron pipe wall thickness of 4.76 mm and a cement mortar lining
thickness of 12.5 mm. A steady-state flow of 62.35 L/s results from the initial velocity
of 0.15 m/s. The total length of the pipe is 1,000 m, and a steady-state Darcy-
Weisbach friction factor was calculated for an assumed pipeline roughness height of
๐ = 0.01 mm.
Considering that detection of the leak includes its location and size, different leaks
needed to be modeled. For all sizes, the leak was defined as a circular orifice with
diameter (๐ท ) that varied in diameter between 13 and 58 mm. This diameter range was
๐ฟ
selected to account for the flow through the leak in comparison with the steady-state
flow in the pipeline. Table 5-2 shows the list of the circular orifice diameters
considered to represent the leak including the ratio (as a percentage) of the diameter
of the leak to the internal diameter of the pipeline (727.5 mm).
Generation of the transient head trace variation data for the training and testing was
conducted using a MOC by changing the size of the leak randomly with a precision of
1 mm. The location of the leak was also modified in each simulation, and a summary
57 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
of the approach for generating the locations of the leaks is explained in the next
section.
Table 5-2. Leak diameters considered for training and testing the ANN.
D D /D D D /D D D /D D D /D
L L L L L L L L
(mm) (%) (mm) (%) (mm) (%) (mm) (%)
13 1.79% 25 3.44% 37 5.09% 49 6.74%
14 1.92% 26 3.57% 38 5.22% 50 6.87%
15 2.06% 27 3.71% 39 5.36% 51 7.01%
16 2.20% 28 3.85% 40 5.50% 52 7.15%
17 2.34% 29 3.99% 41 5.64% 53 7.29%
18 2.47% 30 4.12% 42 5.77% 54 7.42%
19 2.61% 31 4.26% 43 5.91% 55 7.56%
20 2.75% 32 4.40% 44 6.05% 56 7.70%
21 2.89% 33 4.54% 45 6.19% 57 7.84%
22 3.02% 34 4.67% 46 6.32% 58 7.97%
23 3.16% 35 4.81% 47 6.46%
24 3.30% 36 4.95% 48 6.60%
From each training set of location and size, the transient head trace at the closed valve
was obtained for a duration of 2.5 s in both hydraulic systems which corresponds to
more than the first period of reflections for the pipeline system (2๐ฟ/๐), and to less
than the complete cycle of the transient pressure wave (4๐ฟ/๐). The selection of this
time corresponds to the fact that transient waves contain information on the complete
pipeline and the different features in the first (2๐ฟ/๐), and that energy dissipation
effects are not significant.
5.5 Artificial Neural Network Training Settings
This section summarizes the ANN and input dataset configuration that was required
to develop a satisfactory application of ANNs for the location of junctions and leaks.
Different settings needed to be defined including the type and characteristics of the
ANN, and the characteristics of the input head data in terms of their nature, time
resolution, and size.
Two different ANN architectures were considered, a dense network and a 1D-
convolutional network. Preliminary tests for the determination of the location of a
junction, which are not included in this paper for brevity, showed that the prediction
in the location of this feature with a 1D-convolutional network were smoother and
more uniform for junctions located across the analyzed pipeline with an average error
58 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
of 1.14 m, whereas the results of the dense network predicted junction locations with
errors up to 30 m. In addition, convolutional networks have shown successful
performance in predicting multiple outputs in fields such as image segmentation (Raza
et al. 2017). This ANN architecture was then used for both the location of the junction
and the leak.
The 1D-convolutional network used in this paper has been designed by defining a set
of characteristics for it, by training the ANN for the location of a junction and by
testing it to analyze its performance. These characteristics were defined to ensure that
the final ANN was successful in finding the location of the junction, without over-
fitting to the training data. The final set of characteristics of the 1D-convolutional
network include a network with: (1) the use of leaky rectified linear unit (Leaky ReLU)
as an activation function, (2) three convolutional layers of size 1,200, 600, and 300,
(3) 10 filters in each layer, and (4) three dense layers of size 21, 9, and 2 (or 3
depending on the analyzed feature). With this configuration, 32,409 weights have been
trained.
To train and test the ANNs, multiple locations for the considered features (junction or
leak, depending on the case) were selected for generating the input data. The number
and the characteristics of these examples were defined in terms of the distribution of
these locations (random or uniform locations along the pipeline), the time resolution
of each transient head trace (original resolution or down sampled), and the number of
locations along the pipeline (ANN input data size). To define these characteristics,
preliminary training and testing runs were developed for the junction model and then
applied to the leak model.
For determining the junction locations distribution along the pipeline (to create the
ANN input dataset of transient head traces), two options were considered: generating
transient head traces for junctions at either uniform spacing or randomized spacing
along the pipeline. A random distribution of the location of the junction proved to be
more efficient when compared with a uniform distribution given that the ANN
predictions of the preliminary tests were more accurate in terms of the median of the
error and the maximum errors.
On the other hand, depending on the number of junctions considered to generate the
input transient head traces, the distance between locations varied between 0.1 and 2 m
59 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
and the numerical results of MOC needed to reflect different transient head traces for
each location. To achieve this, the time resolution of the MOC is required to match
the spatial separation of each location to apply the method along the characteristic
lines. This implies that, for instance, for a spatial resolution of the junction location of
0.1 m and a total simulation time of 3 s, 10,000 junction locations would be necessary
to cover a 1,000 m pipeline and each of these locations would have 20,000 head values.
In this case, the complete input dataset would include 200 million head values.
Considering the potential size of the ANN input dataset, a timewise down-sampling
process was explored. To evaluate the performance of this process, two input datasets
were created: one with transient head traces with 26,800 head values (for 10,000
different junction locations) and one with a uniform down sampling from the first
dataset to 1,200 head values, for the same 10,000 different junction locations. This
indicates that the down-sampled transient head traces preserve the information from
the original junction location even though they only contain 1,200 head values. Two
1D-convolutional networks were then trained and tested 20 times, and the obtained
results were analyzed in terms of the distribution of the average and the maximum
absolute error for all the junctions across the pipeline.
Results from this sensitivity analysis are presented in Figure 5-7. This figure shows
that the median average absolute error is slightly larger for the down-sampled input
dataset; however, the total variation of these errors is smaller. In addition, the
maximum absolute error is notably better when the down-sampling process is carried
out. This behavior is explained primarily because the training process of the ANN is
easier when fewer weights are required to describe the same head variation traces.
Therefore, by down sampling the transient head trace variation, the size of the input
data decreases without significantly compromising the information contained within
it.
Finally, the number of training and testing examples (ANN input sample size) was
defined. The selection of this parameter is directly involved with the computational
effort required to generate the input data and subsequently, to train the ANN. The most
convenient input sample size was determined separately for the location of the junction
and the leak because the transient head deviations induced by the presence of a leak
60 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
are more subtle than for a junction, resulting in the necessity of using more training
examples.
Figure 5-7. Input data down sampling for junction location. Box whiskers plots of:
(a) average absolute error; and (b) maximum absolute error.
For the location of a junction, four different sample sizes were tested: 500, 1,000,
5,000, and 10,000 locations along the 1,000 m pipe. After performing the preliminary
training and testing of the ANN, a sample size of 5,000 was selected as the most
appropriate. It provided accurate results for the location of the junction with
significantly less computational effort (the training process took 248 min) because the
junction only needs to be located each 0.2 m for a sample size of 5,000 and not each
0.1 m as for a sample size of 10,000. Using a sample size of 5,000 indicated that to
simulate 2.5 s, 13,400 head values would be calculated with a โt = 1.866ร10โ4 s.
After the down-sampling process, 1,200 head values were obtained with a โt =
0.0021 s.
For the location of a leak, more samples were required. Four different input data size
sets were tested: 5,000, 10,000, 25,000 and 50,000. Results of preliminary predictions
made with ANNs trained with 25,000 and 50,000 were significantly better than the
first two sample sizes. However, given the results of these tests it was not possible to
choose a preferred sample size; therefore, results are presented for two ANNs, one
trained with 25,000 locations and the other with 50,000. In both of these sample sizes,
the separation between leaks was 0.2 m (preserving the ฮt described above for the
location of the junction). Therefore, five (or 10 depending on the case) different
locations were randomly generated each 0.2 m.
61 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
5.6 Results
Once the different settings for training an ANN were defined as described above,
results for the prediction of the location and size of a junction and a leak have been
obtained. For both cases (junction location and leakage), the results are presented in
terms of the training and testing location errors, followed by testing of the size errors
(diameters of both pipes on either side of the junction in the first case and diameter of
the leak in the second case). In addition, a distribution of the error, represented as a
percentage exceedance plot, for the location of the features is presented.
5.6.1 Junction Location and Sizing
Results for the topology identification application by locating and sizing a junction
along a pipeline are presented first. Based on the hydraulic system configuration
described in Figure 5-5, the location of a junction was predicted using the ANN by the
length of the upstream segment of the pipe ๐ฅ. Sizing of pipes on either side of the
junction model refers to the determination of the two diameters associated with the
two segments of the pipe. Those diameters are included in the combinations described
in Table 5-1. The ANN used was a 1D-convolutional network using a sample size of
5,000 (half for training and half for testing) with examples generated randomly (in 0.2
m intervals) and down sampled to 1,200 time steps.
Location errors are shown in Figure 5-8(a) for the 2,500 examples used for the training
stage of the application. The average absolute error for the training dataset was 0.79
m with a maximum error of 7.21 m when the junction is located at 987 m downstream
(within 13 m of the end of the pipe) of the reservoir (the total length of the pipe is
1,000 m). From this figure, it is possible to observe that near the extreme ends of the
pipe, results tend to present larger errors in comparison to the results in the central part
of the pipe. This behavior is due to the effect that those extreme locations have on the
transient head trace at the measurement point. When close by, the reflections from the
junction interact with the reflection at the reservoir (if the junction is close to the
upstream end) or with the initial transient head rise (if the junction is close to the
downstream end of the pipe). For these locations, instead of having a head drop (or
increase depending on the ratio of diameters ๐ท /๐ท ), the presence of the junction
1 2
causes a short spike at the beginning of the trace or close to a time of 2๐ฟ/๐ s that
62 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
makes it difficult for the ANN to use the same weights that represent this behavior in
the rest of the pipeline.
Figure 5-8. Location errors for junction location: (a) training dataset errors; and (b)
testing dataset errors.
On the other hand, testing location errors for the 2,500 examples used for this stage of
the application are presented in Figure 5-8(b). By comparing this figure with Figure
5-8(a) it is possible to conclude that the ANN trained is not over-fitted because the
error behavior is similar for the training and the testing datasets. Over-fitting is one of
the key issues when working with ANNs because depending on its parameters and the
input data, the ANN can perform satisfactorily for the training set but poorly for new
unknown data as the test dataset. The testing dataset had an average absolute error of
0.81 m (slightly larger than the average absolute error for the training dataset) and a
maximum error of 8.3 m when the junction was located at 420 m from the reservoir.
Errors in the estimation of the values of both diameters of the pipe junction (๐ท and
1
๐ท ) are shown in Figure 5-9. In general, errors in diameter were extremely small,
2
reaching a maximum of 27.1 mm for ๐ท and only 7.25 mm for ๐ท for the worst
1 2
estimation. Average absolute errors were 0.31 and 0.27 mm, respectively. However,
when predicting diameters, it is important to consider that pipe diameters cannot take
on continuous values due to commercial production restrictions; therefore, results
from Figure 5-9(a) were rounded to the closest diameter in the list defined in Table
5-1 and are shown in Figure 5-9(b). By performing this rounding procedure, only in
one diameter prediction case (out of the 5,000 diameter predictions, two per each
testing example) was the diameter not correctly identified.
63 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
a) b)
Figure 5-9. Junction diameter errors for testing dataset: (a) before rounding process;
and (b) after rounding process.
Finally, Figure 5-10 presents the percentage exceedance associated with the absolute
error in the junction location prediction. This type of plot is useful to analyze the
distribution of errors in detail. The percentage exceedance can be interpreted as the
proportion of time that the junction location surpassed a certain error size. For
instance, according to Figure 5-10, in 5% of the testing examples, the junction location
was predicted with an error of 2.32 m or larger. For reference, the maximum location
error for the training and the testing datasets is also shown in the figure.
Figure 5-10. Percentage exceedance for absolute junction location error.
The distribution of the errors shows that the extreme values of error (larger than 3 m)
are relatively rare in occurrence, and these correspond to only 2% of the total testing
64 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
examples. In addition, the similarity in the distribution of error for the training and the
testing datasets is validation of the adequate performance of the 1D-convolutional
network to predict the location of a junction in a single pipeline.
5.6.2 Leak Location and Sizing
Results for the prediction of the location and size of a leak are presented in this section.
As described above, the preliminary tests developed for determining the most adequate
input training sample size for the location of the leak were inconclusive between using
25,000 and 50,000 examples for the training and testing of the ANN. Therefore, results
are presented for both input sample sizes in Figure 5-11.
Figure 5-11. Location errors for leak location: (a) training dataset errors with 25,000
training examples; (b) testing dataset errors with 25,000 training examples; (c)
training dataset errors with 50,000 training examples; and (d) testing dataset errors
with 50,000 training examples.
Training errors for an input sample size of 25,000 are shown in Figure 5-11(a),
whereas the training errors for a sample size 50,000 are shown in Figure 5-11(c). The
average absolute error for the ANN when the input sample size was 50,000 is 1.09 m
in comparison with 1.24 m that was obtained when the input sample size is 25,000. In
contrast, the maximum value for the error is smaller for 25,000 examples (10.13 m
65 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
compared with that of 32.29 m for the 50,000 sample size). In the training process,
both sample sizes present reasonable results; however, when both ANNs are used for
predicting, results differ for the two sample sizes.
Figure 5-11(b) and Figure 5-11(d) present the leak location errors for the testing
dataset obtained from the ANN trained with the considered input sample sizes. By
comparing these two figures, it is possible to note that using an input sample size of
25,000 leads to larger errors both for the maximum and for the average absolute error.
In addition, errors larger than 5 m, for example, are more frequent as seen in the
figures. These results show that the ANN obtained with a sample size of 25,000 was
over-fitting to the training examples because it is unable to predict correctly new data
when more testing examples are used.
Based on these findings, the ANN obtained from an input sample size of 50,000 was
selected because it can predict the location of the leak with an average error of 1.15
m, and it shows a similar behavior when compared with the training errors (showing
that less over-fitting issues are present). The maximum error for the 25,000 examples
used for the testing (which corresponds to a total input dataset of 50,000 because half
of the data is used for training) was 98.21 m; however, this misleading result was
predicted for an example where the leak was located only 1.01 m from the reservoir.
Those errors were expected given the fact that the effect on the transient head trace of
a leak located close to either of the extreme ends of the pipe can be difficult to
distinguish from the reflections at the boundary conditions. In addition, it was
observed that for leaks close to the reservoir or close to the closed valve, the down-
sampling process in time could cause the loss of information on the reflection from
the leak. However, in a real application, if the leak was located close to either of the
extremes, it would be manually detected in the process of obtaining the transient data.
In general, the time down-sampling process does not affect the accuracy of the
technique because even though some information is lost in this process, the ANN is
sensitive to the changes in pressure that are not lost in the down sampling.
Errors for the leak size are presented in Figure 5-12, only for the ANN obtained from
an input sample size of 50,000. This figure shows that errors in the prediction of the
leak size are highly satisfactory because most of them are within the range of โ0.2, 0.2
mm. The maximum error is 1.11 mm when the leak is located only 61 cm away from
66 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
the reservoir, and the average absolute error is 0.03 mm. Considering that in a real
pipeline, leak size could take any continuous value, a rounding process was not
performed as was done for the diameter size in the junction sizing. In general, results
from Figure 5-12 show that the prediction of the leak size using ANNs is successful.
Figure 5-12. Leak size errors for testing dataset.
Figure 5-13 presents the percentage exceedance for the absolute leak location error for
both the training and the testing dataset. To improve the analysis of the data, the first
and the last 12 m (which would correspond approximately to two segments of pipe in
the field) were eliminated before compiling the figure because it allows us to see the
behavior of the error once the extreme and unrealistic errors are discarded. By doing
this, the maximum error for both the testing and the training is about 12 m, and errors
larger than 4 m are only found in 1.89% of the examples. Likewise, in 95% of the
examples, the ANN prediction for the location of the leak in the pipe is at least or more
accurate than 3.0 m, which represents 0.3% of the total length of the 1,000-m-long
analyzed pipe. Finally, a similar distribution of location errors for the training and the
testing datasets demonstrate the adequate performance of the ANN when it is
predicting locations in new examples.
67 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
Figure 5-13. Percentage exceedance for absolute leak location error
5.7 Conclusions
A novel technique to identify topological elements such as junctions and to detect,
locate, and size leaks has been proposed in this paper using ANNs as a tool based on
the generation of a transient event. Two ANN architectures have been compared, and
a 1D-convolutional network proved to predict more accurately the location of a
pipeline junction in comparison with a dense network. The transient head data required
for the training and testing of the designed ANNs have been obtained numerically
using the MOC by changing the position of the analyzed feature randomly along the
length of the pipeline. A time down-sampling process has been conducted to reduce
the time resolution of the transient head traces, which reduced the computational time
and enhanced the performance of the ANN in predicting the location of the features.
The number of training examples required to obtain accurate results has been defined.
For the location and sizing of a junction, 2,500 examples are enough to train an ANN
with accurate results. However, for the location and sizing of a leak, 25,000 examples
are necessary to train (in a 50,000 sample dataset for training and testing) the ANN
because the effect of a leak in the transient head variation trace is more subtle and is
more difficult for the ANN to learn from the input data.
68 |
ADE | Chapter 5 โ Fluid transients and ANN for Leak Detection and Topology Identification
The numerical application of the technique shows that an ANN can accurately predict
the location of a junction with an error in the location smaller than 2.32 meters in 95%
of the examples tested with a virtually perfect prediction of the diameter sizes on each
side of the pipeline junction. In addition, for the location and sizing of a leak, the
trained ANN can also accurately predict the leak location with an error smaller than
3.0 mm in 95% of the examples tested with an average absolute error of 0.03 mm in
the prediction of the leak size. These results demonstrate the outstanding potential of
using machine learning techniques and fluid transient waves for the location of
topological elements and anomalies in pipelines. The approach proposed is fast,
accurate, and data-driven because no previous information of the system or a hydraulic
transient numerical model is required for the testing stage; only a transient head
variation trace is needed.
Considering that this is the first application of fluid transient waves and ANNs, the
proposed approach has been tested in simple hydraulic systems with data obtained
numerically; however, some challenges arise when the same technique is applied to
more complex situations, and the impact of these challenges will be addressed in future
field tests of the performance of this technique. Detection of topological elements and
anomalies such as leaks is, in essence, a complex problem. Pipelines and water
distribution systems have different elements that can affect the transient head signal
obtained in the field, and changes in background conditions can make the detection of
anomalies challenging. However, the background fluctuations are typically more
gradual with lower frequency content allowing them to be differentiated from other
anomalies. In addition, for subtle anomalies (such as small leaks), the deteriorating
condition of the pipeline itself can hide the transient head response of the anomaly.
Considering these challenges, it would be expected that the accuracy of the use of the
proposed technique would be reduced when applied to real pipelines. Nonetheless,
ANNs are highly amenable to retraining on real data as it is harvested, and ANNs are
robust to noise. The results here offer promise that the use of ANNs can be extended
to field settings.
The application presented in this paper demonstrates the promising potential of the
proposed technique, yet it should be expanded to the analysis of hydraulic systems
with different dimensions and configurations, and validated with laboratory and field
data to explore its usefulness and accuracy in comparison to other existing techniques.
69 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Foreword
This chapter (Journal Publication 2) presents a comprehensive methodology for the
active inspection of pipelines based on the framework presented in Chapter 5. There
are cases where the transient pressure traces obtained from the analyzed pipeline
contain background pressure fluctuations, where this framework is not able to
accurately locate leaks. This is because these fluctuations are not reproduced by
numerical transient flow models. To address this, the deployment of stochastic
resonance is used to enhance ANN performance for leak detection. This chapter also
presents the results of the application of this methodology to the detection of a leak in
a pipeline in a laboratory setting.
The principles of stochastic resonance are applied in this methodology are used for the
ANN training datasets containing numerically generated transient pressure traces with
the addition of different noise intensities. By training ANNs with these datasets, an
optimal noise intensity can be determined. The deployment of stochastic resonance
has proven to be fundamental to obtain more robust ANN predictions to accurately
detect leaks in real pipelines. To complement the information presented in the
publication and provide additional context for this chapter, two elements are included
here. The importance of developing a more robust methodology to detect anomalies
in pipelines is described by presenting the performance of the ANN proposed in
Chapter 5 to interpret a transient pressure trace obtained in a laboratory setting. These
results are compared with the leak location prediction of the final set of ANNs
presented in this chapter. In addition, the computational resources required for the
training of the ANNs used in this methodology are also reported.
Following the framework proposed in Chapter 5, five ANNs were trained to predict
the location of a leak in the laboratory pipeline described in Section 6.4.1. The
distribution of the predicted locations for 14 experimental laboratory based conducted
tests is presented in Figure 6-1 in purple. This figure shows that the use of the ANN
architecture and the training framework described in Chapter 5 does not provide with
73 |
ADE | Abstract
Water losses through leakage represent a significant problem for asset management in
water distribution systems. The interpretation of fluid transient pressure waves after
the generation of a transient event has been used as a technique to locate and
characterize leaks for more than two decades but existing approaches are often both
model-driven and limited to the existing knowledge of the system. A new, data-driven,
technique using artificial neural networks (ANN) has recently been proposed to detect,
locate and characterize anomalies in water pipelines by interpreting the patterns that
different anomalies induce in transient pressure traces. However, the application of
this technique in more realistic conditions (e.g. in the presence of background pressure
fluctuations) has previously proven challenging because these conditions are not
reproduced by the numerical models and affect the accuracy of the anomaly location
prediction of the ANNs. To address this, one alternative to enhance the response of
any non-linear system includes the introduction of artificial noise, a phenomenon
known as stochastic resonance. This paper harnesses this approach by finding the
optimal artificial noise intensity to be introduced into the training dataset for a set of
convolutional neural networks. In this paper, the enhanced detection of leaks in
pressurized pipelines via deployment of stochastic resonance is demonstrated. The
methodology has been applied to a real pipeline in a laboratory at the University of
Adelaide where 14 transient experimental tests were conducted. Results have shown
that the addition of noise to the transient pressure head training samples significantly
enhances the ANN predictions for the leak location highlighting the existence of an
optimum noise intensity to obtain both accurate and reliable results. When trained with
the optimum noise intensity, the ANNs were able to locate leaks with an average error
of 0.59% in terms of the actual location (in a 37.24 m long pipeline) demonstrating
the promising potential of developing techniques based on ANNs to detect leaks and
anomalies in water pipelines.
77 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
6.1 Introduction
Population growth and urban expansion are a challenge for water distribution systems
(WDSs) since these systems are responsible for the supply of a vital resource to
society. In recent years, major cities have faced a serious water supply crisis(Ahmadi
et al. 2020). One major challenge in addressing these crises is the detection of water
losses and pipeline repair, which has received attention considering that the percentage
of water losses can reach values of 35% in cities such as Kolkata (Mukherjee et al.
2018). Different methodologies have been used to estimate, monitor, detect and
pinpoint the location of leaks as part of water losses management strategies
(Mutikanga et al. 2013). One of these methodologies includes the use of fluid
transients for leak detection that usually involves the generation of a transient event
that travels along the pipeline allowing its inspection in a way similar to the
functioning of radar and sonar techniques (Puust et al. 2010).
Fluid transient based techniques have proven successful in the detection, location and
characterization of leaks in pipelines using the information that can be retrieved from
transient pressure data. However, in most cases, existing techniques are model-driven.
Such model-driven approaches usually require extensive and accurate numerical
modeling, a priori estimation of certain pipe parameters assuming an intact or original
condition, or they require long processing times to obtain an estimate of the leak
characteristics. These limitations motivate the need for data driven techniques that can
quickly interpret transient pressure data obtained from a test and locate leaks
accurately. A new technique merging the use of fluid transients and Artificial Neural
Networks (ANNs) has recently been proposed to locate leaks and changes in pipeline
diameters in pipelines following a transient event (Bohorquez et al. 2020a). In
addition, this technique has also been adapted to detect the occurrence of bursts
(Bohorquez et al. 2021). Although these applications have demonstrated the potential
of using ANNs to interpret transient pressure traces, further experimental and field
validation is needed to test the performance of this technique under more realistic
conditions.
This paper presents an important stepping-stone in developing a general technique to
use ANNs for leak detection in pipelines using fluid transients. The performance of
ANNs for the detection and location of leaks in water pipelines is shown to be
79 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
enhanced via deployment of stochastic resonance. In the following, a background in
transient-based methods for leak detection is provided, focusing on recent
applications. A summary of the new methodology for the detection of leaks in
pipelines under more realistic conditions using a set of ANNs trained with different
noise intensities is then presented. This methodology is split into two main stages:
model development and model application. Finally, the proposed methodology is
applied to a series of experimental transient tests conducted in a laboratory setting.
The results demonstrate that, by training a group of ANNs with pressure transient
traces with the optimum noise intensity, the accuracy of the leak location predictions
can be significantly enhanced, thus providing more robust predictions.
6.2 Background
Transient-based leak detection techniques have been in development for more than
two decades (Jรถnsson and Larson 1992; Liggett and Chen 1994). Different approaches
have been explored and can be classified into three main groups: inverse transient
techniques, frequency response techniques and direct transient methods (Colombo et
al. 2009). More recently, frequency response methods have been combined with
enumeration techniques for leak detection in pipelines with branches and loops by
separating the effect of these known elements on the frequency response of the system
and employing a GA-based optimization to find the leak characteristics (Duan 2017).
This method has proven successful for a numerical application and it has shown the
potential of transient-based methods for operation in more complex systems. Meniconi
et al. (2019) examined the influence of the pipeline initial flow conditions on transient
pressure traces after the generation of a transient event for the visual detection of a
leak in the pipeline (as an example of a direct transient technique). This study
concluded that, depending on the location of the transient generator device, the
transient measured signal can be more sensitive to the initial conditions. If the
generator is located close to the water source, the transient pressure signals obtained
are almost indistinguishable. In contrast, locating the generator close to the end of the
pipeline can produce different transient traces for the same leak in terms of the initial
pressure rise (Meniconi et al. 2019).
Matched-field processing (MFP) has been explored as a frequency domain technique
that can obtain satisfactory results even in noisy environments and it has been
80 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
developed for elastic (Wang and Ghidaoui 2018) and viscoelastic pipelines (Wang et
al. 2019). Advantages of this method include its robustness under noisy conditions or
uncertainty in the wave speed. It has proven effective in both numerical and
experimental conditions when the noise has been assumed to be white noise with a
zero-mean Gaussian distribution (Wang et al. 2019). Other recent applications have
shown that frequency-based techniques can detect leaks under realistic background
noise scenarios using the paired impulse response function obtained from two
measurement points in the system (Zeng et al. 2020). Although significant advances
have been achieved in transient-based methods in recent years, most of the existing
techniques still require testing under perfect conditions (without any leaks), detailed
numerical modeling, significant computer resources or extensive preprocessing of the
signals.
A different group of techniques for leak detection in pipelines has proposed the use of
machine learning algorithms to process the available information from a particular
pipeline system. Some of these techniques have used surrogate features of the pipeline
to predict the most likely location of a leak (Geem et al. 2007) or to predict the
remaining lifetime of a pipeline (Zangenehmadar and Moselhi 2016). However, more
recent techniques have proposed the combined use of machine learning algorithms and
hydraulic measurements in the pipeline. Romano et al. (2014) used different self-
learning artificial intelligence, statistical analysis and Bayesian inference tools for the
detection of burst at a DMA level in real water distribution systems using wavelets for
the denoising of the obtained signal before its analysis. Roy (2017) proposed the use
of pressure fluctuations with hybrid dense ANNs for the location of leaks by
classifying the status of the system to characterize a normal and abnormal condition
in the pipeline. Mujtaba et al. (2020) introduced the use of adaptive thresholds to detect
the occurrence of leaks in gas pipelines using pressure and mass measurements at the
beginning of the pipeline as inputs for the machine learning model and the potential
mass flows at the end of the pipeline as the output of the model for the comparison
with measured data.
Bohorquez et al. (2020a) presented a methodology that uses the transient pressure trace
after the generation of a transient event and convolutional neural networks (CNNs) to
determine the location and the size of a leak in a water pipeline. This merging of
pressure transient traces and CNNs has been demonstrated in a numerical application
81 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
and has great potential, given that this technique is data-driven and can provide
immediate results for the characteristics of a given anomaly. Nonetheless, challenges
can arise when pipelines under more realistic conditions are analyzed. Background
pressure fluctuations due to system operations such as changes in demand or unknown
system components, are not reproduced by numerical models but have an impact on
the transient pressure traces.
Although, to date, no applications using noise to improve ANN performance have
been reported in assessing the condition of water pipelines, several related strategies
have been applied in other fields. Previous approaches have proposed the introduction
of noise during the ANN training in the ANN training samples (Rifai et al. 2011), in
the activation functions (Ikemoto et al. 2018), in the ANN weights (Goodfellow et al.
2016), or in the direction of update of the ANN weights (Neelakantan et al. 2015). The
most popular approach has been the introduction of noise directly into the ANN
training samples to enhance modelsโ robustness and reduce overfitting (Bishop 1995).
Rifai et al. (2011) demonstrated that the error of a multilayer perceptron for document
recognition can be reduced by adding a Gaussian distributed noise in the input layer
regardless of the standard deviation of the noise. Fukami et al. (2020) applied the same
concept to different ANN architectures that were trained to estimate laminar wakes in
a fluid field from limited measurements demonstrating that for the analyzed
architectures, the addition of noise in the training samples improved the performance
of the ANNs when tested in noisy input measurement environments. Nonetheless, if
the magnitude of the noise deviation was too big, the ANN performance was
compromised.
These past applications demonstrate the potential of the use of noise during the training
of an ANN. However, few studies have identified the potential of using an organized
framework to introduce noise in the training of an ANN. The phenomenon where the
performance of a non-linear system (i.e. in this case an ANN) is optimally enhanced
by the addition of a certain noise intensity is known as stochastic resonance. The
concept was proposed for the first time by Benzi et al. (1981) as a non-linear
cooperative effect in which periodic signals can be greatly amplified by large
environmental fluctuations; however, other researchers rapidly extended this to
include any non-periodic signals (Collins et al. 1995). In a non-linear system, it can be
shown that there exists a nonzero value of noise that gives an optimal response to the
82 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
system (Harmer et al. 2002). Stochastic resonance has been observed and applied in
multiple fields (Benzi et al. 1982; Luchinsky et al. 1999; Wang and Santamarina 2002;
Cheng et al. 2020) with only a few applied to the training of an ANN.
Ikemoto et al. (2018) introduced a noise-modulated neural network as an application
of stochastic resonance by perturbing the threshold units in the activation functions
with different noise intensities (described by its standard deviation). Their application
in benchmark artificial problems demonstrated that by adding noise to the threshold
units, the standard deviation of the mean squared error (MSE) decreased as the
standard deviation of the noise increased regardless of the structure of the neural
network in terms of the hidden units. The advantages of using stochastic resonance in
areas related to the development of new technologies such as signal processing (Feng
et al. 2019) or time series analysis (Falanga et al. 2020) is an active research area.
However, previous applications using stochastic resonance in ANNs have been limited
to artificial and numerical benchmark problems in computer science and no
applications have been reported for anomaly detection problems in real infrastructure
such as the detection of leaks in water pipelines.
6.3 Methodology
The methodology developed in this research paper to detect leaks in pipelines using a
set of ANNs is outlined in Figure 6-2. This methodology is divided into two stages:
model development (Stage 1 in Figure 6-2) and model application (Stage 2 in Figure
6-2). The leak detection model development stage should be carried out first and can
be repeated regularly to account for new transient pressure information data collected
from the system. The model application stage comprises the processes required to
analyze a transient pressure head trace to determine the location and the size of a leak
in the pipeline.
83 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-2. Model development and application of leak detection methodology.
6.3.1 Leak Detection Model Development (Stage 1)
The first stage of the proposed methodology is the development of a leak detection
model. This stage comprises the training of a set of ANNs that can locate and size a
leak when the analyzed pipeline experiments background pressure fluctuations. An
appropriate ANN architecture needs to be designed and transient pressure head traces
are numerically generated for the training of these ANNs. The five steps presented in
Stage 1 in Figure 6-2 summarize the development of the leak detection model. It is
important to highlight that these ANNs do not constitute a metamodel of the transient
flow pressure response to the closure of a valve in a pipeline with a leak. These ANNs
are trained to identify the transient pressure wave reflections created by the existence
of a leak in the pipeline.
6.3.1.1 ANN Architecture Definition
The first step in the leak detection model development is the definition of an
appropriate ANN architecture. Bohorquez et al. (2020a) concluded that 1-D
convolutional networks with three convolutional layers, had the potential to identify
leaks in numerically modeled pipelines. However, it has been found that a more robust
architecture is required for an application of this technique in pipelines under more
realistic conditions. The design of this new architecture considered different
84 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
alternatives, including variations from the architecture proposed in Bohorquez et al.
(2020a) to a 1-D convolutional network with 5 layers and increasing filters in those
layers.
Results from this design process are not shown here for brevity but the resulting ANN
architecture includes: a) four convolutional layers, b) use of a Leaky Rectified Linear
Unit (Leaky ReLU) as the activation function, c) 20 filters that increase in the last
convolutional layer and d) three dense layers of size 14, 6 and 2. The resulting number
of weights for the 1-D convolutional networks used in this work depends on the size
and number of filters, and the downsampling frequency selected in Step 1.3 in Figure
6-2. Eq. (1) presents the total number of weights for a 1-D convolutional network
where the first term represents the weights in the convolutional layers (๐) and the
second term represents the weights in the dense layers (๐)
๐ ๐
๐ = โ [((๐ค รโร๐ )+1)ร๐ ]+โ [(๐ ร๐ )+๐ ]. (1)
๐โ1 ๐ ๐ ๐โ1 ๐
1 1
In this equation, ๐ค and โ are the width and height of the filters, ๐ is the number of
๐
filters in the convolutional layer ๐ and ๐ is the number of neurons in the dense layer ๐.
๐
For the first dense layer (i.e. ๐ = 1), ๐ depends on the dimensions of the input layer
๐โ1
defined by the downsampling frequency thus affecting the total number of weights for
the ANN to learn. In general, a larger input layer provides the ANN with more
information regarding the transient pressure head trace, but the training of the ANN is
harder because there are more weights to define.
6.3.1.2 Transient Pressure Head Samples Generation
The leak detection model development stage includes the training and testing of a set
array of ANNs (Step 1.2 in Figure 6-2). To train these ANNs, numerical transient
pressure head data or available recorded transient data can be used. For the application
presented in this paper, numerical transient pressure traces have been used for the
ANN training. Figure 6-3 presents the hydraulic configuration of the single pipeline
that has been used to generate the numerical transient pressure traces. The pipeline is
supplied by a reservoir with an upstream head ๐ป , it has an internal diameter ๐ท and a
0
total length ๐ฟ . At the downstream end of the pipeline, there is a side discharge valve
๐
that is initially open with a flow ๐ . The transient pressure head data is obtained from
0
85 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
a measurement point (๐) at the same location as the side discharge valve. The specific
characteristics of the pipeline that has been analyzed in this paper are presented in
Table 6-1.
Figure 6-3. Single pipeline with a leak system configuration.
A leak, modelled as a circular orifice with a diameter ๐ท , could be present at any point
๐ฟ
along this pipeline. As part of this step, a range of leak sizes must be defined for the
generation of the training data. This range can be defined based only on the diameter
of the leak, the flow that is going through the orifice when the leak is active, or
previous knowledge of the system on past detected leaks. For the application shown
in this paper, the leak size range was defined based on the available orifice diameters
in the laboratory associated with the pipeline experimental apparatus and is presented
in the Results section.
Each sample of the ANN input dataset is a transient pressure head trace generated after
the closure of the side discharge valve with a leak present at a specific point along the
pipeline. To form the complete ANNs training and testing dataset, different leak
locations and sizes were considered. A total of 50,000 different transient pressure head
traces were generated using the Method of Characteristics (MOC) at randomly
selected locations from 5,000 segments along the pipeline and a random leak size. The
MOC can be applied by defining a spatial (โ๐ฅ) and time (โ๐ก) resolution that is
consistent with the wave speed of the pipeline following Eq. (2). This means that for
a desired spatial resolution, a specific time resolution needs to be selected for a
pipeline with wave speed ๐
๐ = โ๐ฅโโ๐ก. (2)
For any pipeline that is analyzed using this methodology, leaks need to be generated
at 5,000 different locations. Therefore, the time resolution required to guarantee that
each transient pressure head trace is different would be very small. For instance, for a
86 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
pipeline that is 1,000 m long and has a wave speed of 1,000 m/s, the average separation
between transient pressure head traces would be 0.2 m, resulting in a required time
resolution of 0.2 ms. This means that to generate a transient pressure head trace for
2๐ฟ/๐ seconds, there would be 10,000 pressure head values per trace. If these complete
traces were used for the training of the ANNs, the total required parameters to train
one ANN (following the architecture defined in Step 1.1) would be 454,000 according
to Eq. (1). The potential size of this input dataset shows that a downsampling process
is necessary because for this example, the total input dataset would contain a total of
500 million pressure head values.
For the application in a numerically modeled pipeline described in Bohorquez et al.
(2020a), the transient pressure head traces were obtained after modelling the sudden
closure of a side discharge valve. However, a more realistic approach should consider
that regardless of the closure method (i.e. a mechanical actuator, a solenoid activated
valve or any other device), the injected transient wave is not completely sharp. As part
of Step 1.2 in the leak detection model development, the closure curve of the side
discharge valve should be obtained and incorporated into the MOC modelling. This
can be achieved by running preliminary tests in the analyzed pipeline to characterize
this curve.
6.3.1.3 Transient Pressure Head Downsampling
As was mentioned above, the potential size of the input dataset when using the MOC
for the generation of the transient pressure head traces can be very large. Therefore, a
time-wise downsampling process is conducted in Step 1.3 of Figure 6-2. Previous
research has shown that ANNs trained with downsampled data have improved
performance and are more computationally efficient (Bohorquez et al. 2020a). This is
partially because ANNs trained on downsampled data have fewer weights and thus are
less prone to overfitting. In addition, the use of downsampled data could potentially
reduce data transfer requirements on applications of this methodology in the field.
Depending on the analyzed pipeline, Step 1.3 includes the selection of the sampling
frequency to which transient pressure head traces are going to be transformed into.
This selected frequency will influence the final number of weights that the ANNs will
be trained on as the frequency defines the size of the initial layer in the ANN.
87 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
6.3.1.4 Noise Characteristics Definition and Application
Bohorquez et al. (2020a) demonstrated the potential of using ANNs for leak detection
in pipelines using fluid transient pressure waves. However, the application of this
technique to pipelines under more realistic conditions such as those containing
background pressure fluctuations, has proven a challenge because the generated
numerical pressure transient head traces cannot replicate these conditions. To address
this issue, this paper introduces the deployment of stochastic resonance applied to the
training of a set of ANNs. The addition of multiple noise intensities to the transient
pressure head input dataset enhances the robustness and the performance of the trained
ANNs when the optimum noise intensity is applied. The addition of noise in training
datasets has been explored in other numerical applications involving artificial
intelligence in different fields. It has been found that the addition of a noise distribution
to the input of a model can translate into a better response of the model output (Murray
and Edwards 1994; Rifai et al. 2011; Fukami et al. 2020). Considering this, Step 1.4
of the leak detection model development comprises the definition of the noise
distribution and the selection of noise intensities to be added to the numerical transient
pressure head traces samples (obtained in Step 1.3).
The transient pressure head noise has been characterized by a Gaussian distribution
with zero mean and a standard deviation ๐, similar to the concept presented by Duan
(2017). The magnitude of the standard deviation has been defined with respect to the
magnitude of the pressure drop in the transient pressure head trace when the smallest
leak (from the range defined in Step 1.2) is present in the pipeline. To illustrate this,
Figure 6-4 presents a generic example of two transient pressure head traces.
The continuous blue line represents the transient pressure head trace after the closure
of a side discharge valve for an intact pipeline. The dash-dotted blue line denotes the
transient pressure head trace when a leak is present in the pipeline and where a
transient event has been generated. The initial pressure head increase after the closure
of the valve is the same in both cases but differences arise when part of the transient
wave reflects from the leak. The bigger the leak present in the pipeline, the bigger the
drop in pressure will be (Bohorquez et al. 2018; Meniconi et al. 2019; Wang et al.
2019). A second y-axis is included on the right-hand side of Figure 6-4 to present the
differences between both transient pressure head traces (red line). This line shows that
88 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
there is a difference โโ during the first 2๐ฟ/๐ seconds after the closure of the side
discharge valve.
Figure 6-4. Pressure head difference between transient pressure head trace for an
intact pipeline and a pipeline with the smallest leak.
Considering this, if the selected noise intensity has a standard deviation larger than
this difference, it is expected that the ANNs will not perform well in identifying small
leaks. In this case, the noise added to the transient pressure head trace would hide the
transient wave reflections from the leak. A total of ๐ noise intensities are selected in
Step 1.4 and the standard deviation for each intensity is defined in Eq. (3) as a
proportion of โโ where ๐ is the multiplier for noise intensity ๐ โ {1,โฆ,๐} and ๐ can
๐ ๐
be any number larger than zero
๐ = ๐ รโโ. (3)
๐ ๐
The selection of the number of noise intensities ๐ will depend on the knowledge of the
background pressure fluctuations present in the analyzed pipeline. In addition, the
computational resources available should be considered when selecting this variable
because each additional noise intensity will represent more ANNs to be trained (as
will be explained in Step 1.5). The array of multiplier values [๐ ,โฆ,๐ ] depends on
1 ๐
the analyzed pipeline and the expected background pressure fluctuations, however, it
is expected that if large values of ๐ are selected, the ability of the ANN to identify
๐
certain leak sizes will decrease.
89 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Once the number of noise intensities and the array of standard deviations ๐ have been
๐
defined, multiple random transient pressure head traces are created. In this paper, five
transient pressure head traces are created for each of the samples of the input dataset
obtained from Step 1.3 by adding the noise to the original transient pressure head trace.
This allows the ANNs to be exposed to different transient pressure head traces that
correspond to the same leak location and size but with different values for the pressure
noise. Thus, the input dataset for each noise intensity has 250,000 different transient
pressure head samples.
6.3.1.5 Leak Detection ANNs Training and Testing
The last step of the leak detection model development (Step 1.5 in Figure 6-2) is the
training and testing of a set of ANNs with the architecture defined in Step 1.1 using
the ๐+1 input datasets obtained from Step 1.4 (including the original dataset without
any noise in the samples). A diagram presenting the set of ANNs to be trained is shown
in Figure 6-5. Each leak detection ANN receives as input one transient pressure head
trace and should be able to predict the correct location and size of the leak only based
on this information.
Figure 6-5. Leak detection set of ANNs. Each group of ANNs are trained with
samples with different noise intensities.
90 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
This diagram shows that for each noise intensity (๐ ) including a no-noise
๐
scenario (๐ ), ๐ leak detection ANNs are trained using its corresponding dataset.
0
Considering that the training of these ANNs is conducted using Stochastic Gradient
Descent algorithms, every time the ANN is trained a different final set of weights is
obtained. Similar to applying Genetic Algorithms starting from different random
number seeds, training ๐ leak detection ANNs can assist in testing the consistency of
the ANN predictions. A good set of ANNs should provide very similar results when
testing with the same data despite having different ANN weights. The number of
possible ANNs to train for each noise intensity will depend on the availability of
computational resources.
Each input dataset is then randomly divided into two groups of equal size: a training
dataset and a testing dataset. For the training process, smaller groups of data are
selected one at a time to find values for the ANN weights and then validated with the
rest of the training data. This is known as batch training and it allows the ANN to learn
from smaller groups of data to avoid overfitting (Nakama 2009). Considering that the
separation of the input dataset into a training and a testing dataset is random and that
batches for each training trial are different for each ANN, the resulting weights are
different.
Once the training process is complete, the ANNs are tested with transient pressure
head traces that have not been exposed to. These predictions are then compared to the
real location and sizes of the testing samples. An ANN that has been successfully
trained should present with a similar distribution of errors in the training and the
testing stages.
6.3.2 Leak Detection Model Application (Stage 2)
The second stage of the methodology presented in this paper is the leak detection
model application (Stage 2 in Figure 6-2). This stage includes a number of different
steps that are necessary to process real measured transient pressure head data from a
valve closure test in a pipeline with a leak to obtain a prediction of its location and
size. A six-step process is described in Stage 2 in Figure 6-3 and it is divided into two
sub-stages: pre-processing and analysis. This section explains how each step may be
carried out for any analyzed pipeline when results from multiple valve closure tests
91 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
are available. However, the same procedure could be conducted if only one test result
is available.
6.3.2.1 Generator Pressure Fluctuation Reduction
The first part of the leak detection model application refers to the pre-processing of
the tests that have been conducted. In this pre-processing stage, the first step
corresponds to the reduction of pressure fluctuations caused by the transient generator
device in the recorded transient pressure head signal (Step 2.1 in Figure 6-2).
Preliminary applications of the leak detection model, not shown in the paper for
brevity, demonstrated that the performance of the set of ANNs was not satisfactory for
different transient tests under the same conditions. To understand the reasons for this
apparent inconsistency in the leak detection model predictions, a vulnerable region
detection analysis was conducted.
Vulnerable region detection analysis has been previously used in computer science to
evaluate the performance of a classifier machine learning model to small perturbations
in different regions of an image. This type of analysis have shown that deep neural
networks are vulnerable to changes around the object of interest (Shu and Zhu 2019).
Vulnerable region detection analysis is closely related to the study of adversarial
examples for deep neural networks where imperceptible perturbations (localized or
distributed in the image) can disrupt the predictions of the models (Szegedy et al. 2013;
Akhtar and Mian 2018).
For the ANNs developed in this paper, this analysis included the successive testing of
the ANNs with perturbed transient pressure head traces. These traces were obtained
by applying different magnitudes of perturbation to each point along the original
transient pressure head trace. Figure 6-6 presents the distribution of the predicted leak
location error (at the top of Figure 6-6) after applying successively a single 0.1 m
perturbation in turn along a transient pressure head trace measured in the laboratory
(shown at the bottom of Figure 6-6). The distribution of errors was obtained after
testing the perturbed samples with five ANNs trained with the same noise intensity.
92 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-6. Leak location error when a perturbation of 0.1 m is applied at only one
point along a laboratory transient pressure head trace.
This figure shows that perturbations at the first 60 points of the transient pressure head
trace induce a considerably larger error in the distribution of the leak location
predictions. These 60 points correspond to the steady state pressure head before the
valve closure. When the perturbation is applied after the valve closure, the ANN
predictions are more consistent, although errors are also present. This analysis
demonstrated that there are features in the steady state segment of a laboratory
transient pressure head trace that induce errors in the ANN performance. Thus, a more
in-depth analysis of measured transient pressure head traces has been conducted.
Figure 6-7(a) presents an example of a transient pressure head trace obtained in the
laboratory (with characteristics presented in Table 1 and equivalent to the system
presented in Figure 6-3). In this figure, two segments of this trace are enlarged
preserving the same scale. Subplot a) shows the background pressure fluctuations
before the valve closure and subplot b) shows the background pressure fluctuations
after the dissipation of the transient event created by the valve closure. Clear
differences in these pressure fluctuations are visible before and after the transient
event. The background pressure fluctuations before the transient event are more
prominent in magnitude. This is due to the interaction that the transient generator
(open valve) has with the pipeline itself that adds to the interaction that the leak orifice
has with the pipeline. The background pressure fluctuations induced by the transient
93 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
generator have been reported by Gong et al. (2018b) where leaks were simulated in
real pipelines with the opening of a standpipe connected to an air valve or a fire
hydrant.
A background pressure fluctuation reduction step has been included to reduce the
background pressure fluctuations caused by the combined effect of the transient
generator and the leak to only the fluctuations due to the presence of the leak. To
accomplish this, the distribution of pressure fluctuations needs to be studied in more
detail. Figure 6-7(b) presents the distribution of the pressure head for the segments of
the complete transient trace highlighted in Figure 6-7(a). Step 2.1 in the proposed
methodology includes fitting both pressure head series to a probabilistic distribution
(in this case a normal distribution), which in the examples presented show a reasonable
agreement. The distribution of pressure head after the transient test will normally have
a larger mean value because there is a reduction in the total flow in the pipeline. The
parameters of the normal distribution for the background pressure fluctuation before
the transient generator closure are denoted with a ๐ as subscript and the fluctuations
corresponding to the pressure after the transient generator closure are denoted with the
subscript ๐.
a) b)
Figure 6-7. Background pressure fluctuation analysis. a) Background pressure head
fluctuation: i) before the transient event and ii) after the transient event. b)
Distribution of pressure head before and after the transient event.
The procedure for Step 2.1 consists of transforming the background pressure
fluctuations before the transient event to have a similar distribution to the pressure
94 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
fluctuations after the generator closure. Based on the parameters of the distributions
obtained, a new probability function can be defined as a normal distribution with mean
๐ and standard deviation ๐ . This means that the new normal distribution preserves
๐ ๐
the mean value of the background pressure fluctuations before the transient test but its
standard deviation is modified to match the background pressure fluctuations after the
transient test. For any value of pressure from the measured transient pressure head
trace before the transient event (โ ) a Z-score (๐ ) is computed using ๐ and ๐ . With
๐ก ๐ ๐ ๐
this value, a modified value for the pressure (โโฒ) is found using Eq. 4. This is then
๐ก
repeated for each value of pressure before the generator closure to obtain a transient
pressure head trace with reduced background pressure fluctuations
โโฒ = ๐ ร ๐ +๐ (4)
๐ก ๐ ๐ ๐.
6.3.2.2 Transient Pressure Head Traces Shifting and Trimming
Once the background pressure fluctuations induced by the transient generator have
been reduced, the transient pressure head traces are shifted and trimmed at Step 2.2
(see Figure 6-2). This is conducted to match the conditions used for the generation of
the numerical transient pressure head samples at Step 1.2 of the model development
stage. Vertical shifting of the transient pressure head traces obtained from
measurements might also be required if the conducted tests had a different initial
pressure. This shifting includes the computation of the difference between the mean
pressure before the transient event and the steady state pressure used in the training of
the ANNs and the transformation of the transient pressure head traces by adding or
subtracting this difference. If the steady state pressure of the transient tests are
significantly different, some variation in the transient response of the system can be
expected, as reported by Meniconi et al. (2019). Trimming the time extent of the
transient pressure head traces includes the selection of the length of interest from the
complete trace. The specific length to trim the traces will depend on the selected
characteristics in Step 1.2 but, in general, the objective is to include some pressure
information before the transient event and at least 2๐ฟ/๐ seconds after the valve closure
to cover the complete length of the pipeline.
95 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
6.3.2.3 Transient Pressure Head Traces Downsampling
To capture the reflections from anomalies such as leaks, high-frequency pressure
transducers are required for realistic applications (Nguyen et al. 2018; Zeng et al.
2020). In addition, this sampling frequency will depend on the pipeline dimensions
and material. Step 2.3 in Figure 6-2 includes the downsampling of the measured
transient pressure head traces to the frequency selected in Step 1.3. This is necessary
to match the measured transient pressure head trace with the input layer of the ANNs
(as presented in Figure 6-5). This downsampling process has been applied in different
fields as signature recognition and can be carried out using different interpolation
methods including linear, polynomial or spline interpolation (Martinez-Diaz et al.
2007).
6.3.2.4 Leak Detection ANNs Application
The reduction of background pressure fluctuations and the shifting, trimming and
downsampling of the measured transient pressure head traces complete the pre-
processing stage. A second stage comprises the following three steps: analysis of the
measured transient pressure head traces using the available ANNs (Step 2.4 in Figure
6-2), selection of a final prediction for the leak location and size (Step 2.5 in Figure
6-2) and a verification process (Step 2.6 in Figure 6-2).
The leak detection ANNs application step includes the analysis of all the recorded and
pre-processed transient pressure head traces using the ANNs trained in Step 1.5.
Therefore, multiple leak location and size predictions will be obtained from this step
depending on the number of noise intensities (๐) selected in Step 1.4, the number of
ANNs trained per noise intensity (๐) and the number of available transient tests
results (๐). For each transient test (๐), a box whisker plot can be created that
summarizes the distribution of the predicted leak locations. This distribution is created
from the results of ๐ leak detection ANNs for each noise intensity (๐) and the ANNs
trained with transient pressure head samples without any added noise. The analysis of
the transient tests through ANNs trained with different noise intensities constitutes the
application of stochastic resonance. Therefore, it would be expected that an optimum
noise intensity is identified and a final prediction is selected in the following step of
the methodology.
96 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
6.3.2.5 Leak Location Prediction Selection
According to Harmer et al. (2002), the addition of some noise to a non-linear system
can enhance its response. However, there is a point at which the addition of too much
noise prevents further improvement. As mentioned above, this phenomenon is known
as stochastic resonance. One of the objectives of this paper is to demonstrate the
application of this concept to the use of ANNs to detect leaks in pipelines under more
realistic conditions such as background pressure fluctuations. The selection of a leak
location (and size) prediction is included as a separate step in Figure 6-2 (Step 2.5)
because it involves the analysis of the distribution of predictions obtained in Step 2.4.
The first part of this analysis is related to the scatter of the leak location predictions
for the ANNs (๐) of a particular noise intensity (๐). If stochastic resonance is
relevant, it would be expected that the predictions of the leak location in ANNs trained
with larger noise intensities would be more consistent and closer to the real location
of the leak until the optimum noise intensity is reached. To test this, box whisker plots
can be created using the predictions from the available transient tests to evaluate the
effect of adding noise to the training samples of the ANNs.
On the other hand, it is expected that ANNs trained with noise intensities that are too
large would not perform well on the training. This is because the reflections from small
leaks would be combined with the added Gaussian noise and the overall performance
of the ANNs would decrease. To measure this, for each group of ANNs corresponding
to each noise intensity (๐) including the ANNs trained without any noise, the root
mean squared error (RMSE) is computed for both the ANN training and testing
datasets. If the resulting RMSE for a particular noise intensity exceeds a predefined
threshold or the training and testing RMSE are considerably different, this noise
intensity would be considered too large.
By analyzing the scatter of the leak location predictions and the ANN RMSE for
training and testing, one group of ANNs (๐) trained with the optimum noise intensity
(๐ ) can be selected. Using these ANNs, a final prediction for the location and size
opt
of the leak in the pipeline using the ๐ร๐ available predictions can be obtained. This
can be done by computing the median leak location for each transient test available
and then analyzing the distribution of those predictions. If for each transient test the
97 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
median location predictions are clustered around one possible location, only that
prediction will be assessed in the verification process. However, if two or more
clusters are identified, the multiple leak location predictions would be used in the last
step of the methodology.
6.3.2.6 Leak Detection Verification
The last step of the application stage considers the use of a numerical transient model
of the analyzed pipeline to verify if the predicted leak characteristics match the
measured trace to a reasonable degree of accuracy (Step 2.6 in Figure 6-2). This does
not represent a verification of the complete methodology but a confirmation step
including a potential refinement of the ANN predictions for a particular pipeline.
Using the same MOC numerical model used in Step 1.2 of the model development,
new transient pressure head traces can be obtained using the predicted leak
characteristics (size and location) obtained in Step 2.5. These numerically generated
traces are then compared with the pressure head measured to assess its similarity using
the normalized root mean squared error (NRMSE). Differences would be expected
between these transient pressure head traces due to different elements present in the
pipeline that are not included in the numerical model. In addition, the fact that the final
prediction is obtained from a distribution of predictions can also cause differences
between these pressure head traces. However, a threshold can be defined to decide if
the ANNs predictions are accurate enough and a final prediction has been reached.
Preliminary analysis showed that cases in which the ANNs predictions are not within
the defined threshold are due to a discrepancy in the predicted leak size, in a similar
way to what it has been reported by (Bohorquez et al. 2021) for the detection of bursts.
Thus, a potential leak size correction has been considered in this methodology through
the generation of additional numerical transient pressure head traces covering the
range of possible leak sizes to find the one that produced a trace with the lowest
NRMSE.
6.4 Results
The proposed methodology for leak detection in pipelines as described above has been
applied to a series of tests conducted in the Robin Hydraulics Laboratory of The
University of Adelaide. The objective was to demonstrate the feasibility of using
98 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
ANNs to detect the location and size of leaks in pipelines under more realistic
conditions. This section outlines the characteristics of the analyzed pipeline and the
transient tests followed by a description of the application of the methodology for both
stages following the steps presented in Figure 6-2.
6.4.1 Laboratory Tests
The pipeline in the laboratory has the configuration shown in Figure 6-3. The pipeline
is connected at both ends to pressurized tanks. An inline valve has been closed at the
downstream of the pipeline to allow flow only through a solenoid valve installed right
before the end of the pipeline. The characteristics of the pipeline are shown in Table
6-1. A circular orifice of size 2.2 mm has been installed located 28.52 m downstream
of the source tank to simulate a leak.
Table 6-1. Pipeline characteristics.
Characteristic Units Value
Length of pipe (๐ฟ ) (m) 37.24
๐
Internal diameter of the pipe (๐ท) (mm) 22.14
Wave speed of pipe (๐) (m/s) 1305
Wall thickness (๐) (mm) 1.63
๐ฟ /๐ time (s) 0.029
๐
The transient event to detect the leak is generated by the fast closure of the solenoid
valve with a closure time of 5 ms. The pressure has been measured with a PDCR 810
pressure transducer with a 10 kHz sampling rate. A total of 14 transient tests were
conducted with the same configuration under similar initial conditions. The pressure
head traces measured for the 14 tests at the downstream end of the pipeline are
presented in Figure 6-8 where each line represents a different test. The initial pressure
head at the end of the pipeline was set to between 20.0 and 23.9 m. The pressure head
was measured from 0.2 s before the valve closure and for a total of 3 s (although Figure
6-8 shows the pressure changes only until 1 s).
99 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-8. Laboratory transient pressure head traces. Each series represents a
different test with a different initial pressure.
Using the results from these 14 tests and the known characteristics of the pipeline, the
ANN leak location methodology presented in Figure 6-2 has been applied to this
system.
6.4.2 Model Development
First, a leak detection model has been developed for the pipeline described in Table 1
following the steps described in Stage 1 of Figure 6-2. The 1-D convolutional ANNs
that were created followed the architecture previously described with four
convolutional layers, 20 filters and three dense layers (Step 1.1). A total of 50,000
numerical transient pressure head traces were generated with a MOC numerical model
by modelling 10 leaks at random locations within each 7.45 mm interval along the
pipeline. Each of these 10 transient pressure head traces had a different randomly
selected diameter varying between 0.4 and 3.5 mm. The total simulation time was set
as 0.09 s which corresponds to 3.15๐ฟ/๐ seconds, ๐ฟ/๐ seconds before the closure of
the valve and 2.15๐ฟ/๐ seconds after to account for the effects of the valve closure
curve in the computed pressure head. To obtain different transient pressure traces, the
time resolution of the MOC numerical model needed to be at least 0.006 ms.
Therefore, the total size of the ANNs input dataset before the downsampling process
is 788 million transient pressure head values (Step 1.2) where each trace has almost
16,000 head values.
100 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
According to Step 1.3 in Figure 6-2, the obtained input dataset was then downsampled
to a selected downsampling frequency of 5 kHz. This frequency has been selected
considering the dimensions of the pipeline and the potential number of weights to train
in the resulting ANNs. A smaller downsampling frequency would create a very small
ANN that would not be able to learn enough information from the transient pressure
head traces. Smaller downsampling frequencies can be selected for larger pipelines
with larger ๐ฟ/๐ characteristics. The resulting number of weights for the leak detection
ANNs following Eq. (1) is 13,868.
After the downsampling process, the input dataset contains 8.55 million transient
pressure head values for the 50,000 traces. This dataset was used in Step 1.4 to create
additional ANN input datasets with the addition of noise in the transient pressure head
traces. Following the definition of noise intensity presented above, the smallest leak
drop (โโ in Eq. (3)) corresponding to the smallest leak considered was 0.1238 m. Six
different noise intensities have been considered in this step and the selected values of
๐ and the derived standard deviations (๐ ) are presented in Table 6-2. These noise
๐ ๐
intensities have been selected considering that the objective was to obtain ANNs with
the ability to find leaks across the complete defined leak size range, without
significantly decreasing performance with the addition of noise.
Table 6-2. Selected Standard Deviation for Gaussian Noise Distribution
Standard Resulting Standard
Deviation Deviation ๐
๐
Multiplier ๐
๐
0.05 0.0062
0.10 0.0124
0.25 0.0310
0.50 0.0619
1.00 0.1238
1.50 0.1857
The information presented in Table 6-2 was used to generate six additional input
datasets. Each dataset contains a total of 250,000 transient pressure head traces given
that five traces have been created for each for the original numerical traces. Five ANNs
were created for each defined noise intensity and five ANNs using the original training
dataset, with no noise included. Each group of five ANNs have the same architecture
but different resulting weights considering Stochastic Gradient Descent algorithms
have been used in its training. As it was explained in Section 6.3.1.5, using these
101 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
training algorithms is similar to applying Genetic Algorithms using different random
number seeds. The resulting set of 35 ANNs were trained and tested simultaneously
using GPU computer cores on the University of Adelaideโs High Performance
Computer (HPC), Phoenix. The training process was conducted for a maximum of 24
hours or less if the desired threshold of accuracy had been achieved.
Figure 6-9 presents the percentage exceedance associated with the absolute average
error in the location of leaks. This plot summarizes the results of the training and the
testing of seven of the 35 ANNs where each plot (a-g) corresponds to a different noise
intensity ANN. Only one plot per noise intensity is included because the distribution
of the errors obtained during training and testing was consistent across the five ANNs.
Two series are included in each of the plots of Figure 6-9. The blue solid line
corresponds to the distribution of the absolute average leak location error for the
samples used for the ANN training. The pink dotted line presents the leak location
error for the samples used during the ANN testing. The percentage exceedance can be
interpreted as the proportion of the total trained or tested samples where the average
leak location surpassed a certain error size. An average error in the predictions is
presented because in some cases two or more traces with the same leak location and
size have been used either for the training or the testing.
It is important to observe that the maximum percentage shown in the figure is 10% (x-
axis). This means that 90% of the time that these ANNs are used with numerical
transient pressure head traces, the absolute average leak location error that is obtained
is smaller than the minimum absolute average error visible in these plots. In addition,
the y-axes in Figure 6-9(a-e) are presented at the same scale to facilitate its analysis.
It can be seen that as the standard deviation for the Gaussian distributed noise
increases, the absolute average leak location errors also increases due to the noise
added to the training and testing samples. From this figure is also evident that the
ANNs trained and tested with transient pressure head traces without any noise
performed better than the rest. However, for all the considered noise intensities, 90%
of the time the absolute average leak location error is 0.12 m or smaller which points
to a successful result from the training of these ANNs.
102 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-2). The pre-processing of the obtained transient pressure head traces (Steps
2.1-2.3 in Figure 6-2) started with the reduction of the background pressure
fluctuations due to the flow through the solenoid valve installed at the end of the
pipeline. Following the process described at Step 2.1, two 0.2-second segments were
analyzed in each of the 14 measured pressure head traces before the solenoid valve
closure and at the end of the 3-s recorded signal. Two normal distributions were
obtained from the pressure fluctuations before and after the solenoid valve closure for
each transient test. An average standard deviation before the transient test of 0.0392
m and an average standard deviation after the transient test of 0.0097 m were obtained.
An example of the resulting transient pressure head traces after the background
pressure fluctuation reduction step is presented in Figure 6-10.
In this figure is possible to see that the background pressure fluctuation reduction
process does not change the transient pressure head traces dramatically, as no
differences are evident when a 20 m scale is used for the y-axis. However, when a
different scale is analyzed (in the red subplot) clear differences in the pressure
fluctuations are noticeable after the transformation of the pressure before the transient
events. This step allows for a reduction in the background transient pressure head
fluctuations allowing for an improved application of the leak detection ANNs.
Figure 6-10. Results of background pressure fluctuation reduction.
The resulting transient pressure head traces were further transformed to complete the
pre-processing described in Figure 6-2. First, the 14 measured transient pressure head
traces were shifted to be aligned to one initial average steady state pressure head. As
104 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
shown in Figure 6-8, the initial pressure head of each test was slightly different within
a 3.9 m range. All traces were aligned to an average steady state pressure head of 21.16
m. This value corresponds to the initial pressure head considered for the generation of
the numerical transient pressure head traces in Step 1.2. The resulting shifted traces
were also trimmed to select only the segments of the transient pressure head of interest
corresponding to ๐ฟ/๐ seconds before the closure of the solenoid valve and 2.15๐ฟ/๐
seconds after this closure. The resulting transient pressure head traces are presented in
Figure 6-11.
It is important to observe that since each transient test had a different steady state
pressure head, the initial pressure head increase after the solenoid valve closure is also
different in every test. This is due to the small differences in the resulting flow in the
pipeline given different initial pressures, in a similar way as reported by Meniconi et
al. (2019). However, the transient pressure head traces were not further transformed
to test the leak detection ANNs performance to predict accurate leak locations under
these conditions. The last step of the pre-processing stage included the downsampling
of the measured transient pressure head traces to a 5 kHz frequency to match the traces
to the dimensions of the input for the leak detection ANNs.
Figure 6-11. Transient pressure head traces to process through leak detection ANN
(each series corresponds to a laboratory test).
The second part of the leak detection model application involved the analysis of the
transient pressure head traces (Steps 2.4-2.6 in Figure 6-2). All the pre-processed
105 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
transient pressure head traces were analyzed using a total of 35 trained ANNs in Step
1.5 (five ANNs per noise intensity level including the ANNs trained with samples
without any noise). The distribution of leak locations predictions is shown in Figure
6-12. A seven-color scale has been used in this figure to illustrate the distribution of
leak location predictions on each noise intensity defined in Step 1.4 and the ANNs
trained with samples without any noise. In addition, Figure 6-12 presents an indication
of the end of the pipeline (37.24 m) and in light blue the location of the leak in the
pipeline (at 28.05 m).
This figure shows the very large range of the leak location predictions when the ANNs
have been trained without any noise in the transient pressure head traces. Except for
two outliers in the predictions for traces #13 and #14, none of the leak location
predictions are within the physical limits of the pipeline. Therefore, these predictions
are not visible in the figure. This result demonstrates the challenges of applying ANNs
for the detection of anomalies in pipelines under more realistic conditions (Bohorquez
et al. 2020a). Since these ANNs have been trained with theoretical numerical samples
with perfect data, the predictions when the analyzed transient pressure head traces
have background pressure fluctuations result in illogical predictions for the leak
location.
Figure 6-12. Predicted leak location for sets of ANNs trained with samples with
different noise intensities.
106 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-12 also presents the significant influence that the addition of noise in the
numerical transient pressure head traces for the ANNs training has in the resulting
distribution of leak location predictions. The addition of a Gaussian distributed noise
with a standard deviation of 6.2 mm (๐ ) has a significant effect in the resulting leak
1
location predictions. Most of these predictions can now be found within the physical
limits of the pipeline (yellow series in Figure 6-12). It is important to note that a
background pressure fluctuation with a standard deviation of 6.2 mm is significantly
smaller in magnitude in comparison to the real background pressure fluctuations
observed in this pipeline. However, its introduction in the ANN training dataset has
proven to be highly effective in improving the obtained leak location predictions. This
finding aligns with previous authors findings that state that the addition of noise can
beneficial for ANN performance (Fukami et al. 2020).
Despite the clear advantages of applying Gaussian distributed noise, the results
presented in Figure 6-12 also demonstrate that the addition of noise with a very small
standard deviation is not enough for a satisfactory prediction of the location of the
leak. This highlights the importance of deploying stochastic resonance to determine
the optimum noise intensity that should be introduced in the ANN training samples
(Harmer et al. 2002). Figure 6-12 demonstrates that as the noise intensity (๐ )
๐
increases, the distribution of the leak locations are more compact and are, in general,
closer to the real leak location. Predictions from ANNs trained with noise intensities
๐ and ๐ (see Table 6-2) are within the length of the pipeline but vary considerably
2 3
between the different transient tests conducted. Leak location prediction errors
obtained from the last three noise intensities (๐ ) range between 2 and 3.8 m with a
4โ6
couple of predictions outside the physical length of the pipelines for ๐ .
4
Although most of the transient tests allow for a similar distribution of predictions for
a particular noise intensity, transient test #1 and #12 resulted in more scattered leak
location predictions. Leak location predictions for transient test #1 are less satisfactory
because this test had a more prominent difference in the steady state pressure head.
Thus, the difference in the resulting initial pressure head increase after the closure of
the solenoid valve is more prominent as can be seen in Figure 6-11. On the other hand,
even though transient test #12 does not present with any particular differences in
comparison with the other transient tests, it has produced less consistent results for all
the noise intensities. These results point out that there might have been additional
107 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
background noise during this test. Considering this, conducting multiple tests provides
more information that the ANNs can process instantaneously and allows for a more
confident prediction of the leak location.
A perfect distribution of leak location predictions would imply that all the ANNs
trained for a particular noise intensity predict the correct leak location. However, given
that each ANN has a different set of resulting weights after the training process, this
result would be very hard to accomplish. Therefore, the effectiveness of ANNs should
be measured on their ability to produce consistent predictions with a reasonable degree
of accuracy for field applications of this technique.
To further analyze the results obtained from the leak detection ANNs, Figure 6-13(a)
presents the distribution of the absolute median error in the predicted leak location for
each group of ANNs trained with different noise intensities. The median leak location
prediction of each transient test in Figure 6-12 has been extracted and the error
between this prediction and the real leak location has been computed. The distribution
presented in blue in this box plot is obtained from the 14 median leak location errors.
This distribution is presented as an absolute value to demonstrate the applicability of
stochastic resonance as it has been reported previously (Ikemoto et al. 2018).
The absolute median error in the leak location for the ANNs trained without any noise
(i.e. noise standard deviation of zero in Figure 6-13(a)) is not visible in the scale of the
plot because all of the predictions are outside the length of the pipeline. Similarly, this
plot demonstrates that the addition of a very small noise distribution in the training
samples (ฯ=6.2 mm) drastically improves the performance of the ANNs. The resulting
distribution of absolute mean location errors oscillates between 1 and 8 m. However,
an 8 m error is still not acceptable for the location of a leak in a 37.24 m long pipeline
(which represents a 21.48% error). As the noise standard deviation increases, it is clear
that the distribution of the absolute median error narrows, in concordance with the
concept of stochastic resonance. Absolute mean location errors vary between 0.02 and
1.09 m (0.05โ2.93% error) for the largest noise standard deviation considered.
108 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-13. Leak location prediction selection. a) Left axis: Absolute median
predicted leak location error for different noise intensities (standard deviation)
obtained from laboratory tests. The error value for standard deviation = 0 is well
above the maximum extent of the vertical axis (234.27 m). Right axis: Training and
testing ANNโs root mean square error obtained using numerical samples. b) Leak
location error range during training and testing using numerical samples.
Analyzing only the distribution of the absolute median location errors in Figure
6-13(a), it would seem logical to select the predictions of the ANNs trained with the
largest noise intensity. However, the optimum noise intensity should be selected also
by consideration of the performance of the ANNs during training and testing. Figure
6-14(a) presents on the right-hand y-axis the distribution of the RMSE for the training
(in light green) and the testing (in black) of the ANNs for each noise intensity (shown
in Table 6-2). The RMSE has been computed using the leak location error of each of
the 125,000 samples used for the training or testing of the ANNs (or 25,000 for the
109 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
case of the ANNs trained without any noise). A satisfactory ANN training process will
have low values of RMSE and similar RMSE magnitudes in both the training and the
testing.
Figure 6-13(b) presents error plots of the RMSE (in circles) and the complete range of
errors for the leak location prediction (whiskers). These figures show that ANNs
trained with samples with large noise intensities result in larger values of RMSE and
significantly larger ranges of possible leak location errors. Both of these metrics are
considerably larger for the last two sets of ANNs (corresponding to ๐ = 0.1238
and 0.1857) with significantly different results for the training and the testing of these
ANNs. These results point to a certain level of overfitting in these ANNs that is also
visible in Figure 6-9(f) and Figure 6-9(g).
Although these results were presented as part of the model development stage (Step
1.5 in Figure 6-2), they are relevant in the model application stage for the leak location
prediction selection step. The final leak location prediction should be a robust
prediction (in terms of consistency amongst the conducted tests) and be the product of
a reliable set of ANNs. For this reason, it can be concluded that the optimum noise
intensity for this application of the proposed leak detection model is obtained when
the noise has a Gaussian noise distribution with a standard deviation of 6.2 mm (๐ ).
4
The median leak location prediction for this group of ANNs was 28.74 m and the
median predicted leak size was 2.32 mm. These predictions represent a 0.58 % error
in the location of the leak and a 5.52 % error in the size of the leak.
The last step of the model application consists of verifying the accuracy of the obtained
prediction. This step comprises the generation of a numerical transient pressure head
trace with the characteristics of the final prediction obtained in the previous step and
its comparison with the measured transient pressure head traces. This comparison is
presented in Figure 6-14 using test #18 as an example.
110 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
Figure 6-14. Comparison between transient pressure trace numerically modelled
from the selected ANN prediction and laboratory transient pressure trace.
A reasonable match between these two traces is observed in this figure demonstrating
the successful prediction of the location and size of the leak using a set of ANNs. The
NMRSE has been computed between these two transient pressure head traces
obtaining a value of 2.06% demonstrating again the accuracy of the methodology
proposed.
6.5 Conclusions
This paper has proposed a new comprehensive technique for the location and
characterization of leaks in pipelines using fluid transients and ANNs. This
methodology has proven successful when applied to pipelines under more realistic
conditions in the presence of background pressure fluctuations. A full methodology
that is divided into two stages (model development and model application) has been
presented. The model development stage includes the design and training of ANNs
capable of identifying leaks in transient pressure head traces with different noise
intensities. The model application stage describes the pre-processing and analysis
steps for any pressure transient measurements obtained from a transient event caused
by the closure of a valve in a pipeline.
111 |
ADE | Chapter 6 โ Fluid Transients, ANN and Stochastic Resonance for Leak Detection
This technique has been applied to a laboratory pipeline where 14 transient events
were generated with the closure of a side discharge solenoid valve. One circular orifice
was installed in the pipeline to simulate a leak. A leak detection model with 35
different ANNs was developed for this pipeline where six different noise intensities
were considered with standard deviations between 6.2 mm and 0.186 m. The
application of the leak detection model to the available transient tests has demonstrated
the significant importance that the addition of noise has in the performance of the
ANNs for the prediction of the location of a leak in the pipeline. For the ANNs trained
using numerical transient pressure head traces with no noise added, the distribution of
the leak locations was beyond the actual extremities of the pipeline in most cases.
However, as the noise intensity increases, the distribution of the leak location
predictions narrows around the real leak location (as shown in Figure 6-12).
The results obtained in this paper demonstrate that the deployment of stochastic
resonance assists in detecting leaks in water pipelines. With the addition of noise in
the training samples of an ANN, its performance is significantly improved, to the point
that consistent and accurate predictions can be obtained. To select the optimum noise
intensity for the presented laboratory application, a combined analysis of the
distribution of the predicted leak locations and the RMSE of the training and testing
of the ANNs has been conducted. Results from that analysis have shown that the
optimum noise intensity was found when the standard deviation was 6.2 mm.
The final leak prediction has been determined for the laboratory pipeline only 0.74 m
away from the real leak location. This prediction corresponds to an error of 0.59%
with a very accurate prediction of the leak size. The results obtained in this paper
demonstrate that the use of ANNs trained with numerical samples with the addition of
noise is a promising technique for leak detection in pipelines under more realistic
conditions. Although expected differences are evident between the available
numerical model and the measured transient tests, an accurate prediction of the
location and size of the leak was obtained. This represents an important stepping-stone
in developing a fully automated methodology for leak detection in pipelines using
transient waves. However, more research is needed to analyze the performance of this
technique under noisy conditions such as demand consumption and in more complex
systems such as pipeline lopped networks.
112 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
processor and 16 GB of memory using the ANN training procedure described in
Section 3.2 taking between 3 and 15 hours until the training threshold was met.
The passive inspection methodology presented in this chapter relies on the
identification of the negative transient pressure wave caused by the occurrence of a
burst. Since the exact moment of the occurrence of this abnormal event is unknown, a
methodology for the analysis of a continuous transient pressure trace is required. This
chapter proposes a novel time window analysis that is incorporated in the proposed
methodology and is presented in Figure 7-1. This analysis includes the processing of
one time window at a time to determine if abnormal conditions are evident in the
pipeline until time windows containing the complete set of reflection are found (time
window 4). The classification of these time windows is carried out using a
convolutional neural network incorporating the activation function Softmax
(Goodfellow et al. 2016).
Figure 7-1. Sliding time window analysis.
To train an ANN to detect the occurrence of an abnormal events, multiple transient
pressure traces representing different bursts have been numerically generated,
processed and used. The total number of time windows obtained from this process can
be very large. If this complete dataset of transient pressure time windows was used for
the training of an ANN, this process would be too computationally expensive.
Therefore, only a portion of this dataset needs to be selected. A preliminary analysis
indicated that a careful selection of a fraction of the complete dataset is required. This
demonstrates that the use of ANNs for the inspection of pipelines requires an
understanding of the transient flow behavior in a pipeline to obtain satisfactory results.
116 |
ADE | Abstract
The occurrence of bursts in water pipelines can not only prevent the system from
functioning properly, but it can also produce significant water loss that disrupts
activities in urban areas. Therefore, the detection and location of bursts in water
distribution systems is a vital task for water utilities. Various techniques currently exist
to detect the occurrence of these events, but there is a need for a permanent monitoring
method that can detect and identify anomalous events quickly and accurately. This
paper presents a new technique that uses artificial neural networks (ANNs) to detect
and identify bursts in pipelines by interpreting the transient pressure waves that a burst
causes along pipelines. The technique is divided into two stages: a model development
stage and an application stage. The model development stage includes the generation
of transient pressure traces and the training and testing of two different ANNs to (1)
detect burst occurrence and (2) identify burst location and size. The application stage
includes the processing of a potentially continuous transient pressure trace, analysis
by the previously trained ANNs, and then the verification of the results using a
transient flow forward numerical model. A numerical application demonstrates the
principles of the technique and the potential for merging the use of fluid transient
waves and ANNs. The technique has also been validated in the laboratory, indicating
that the prediction of the location of the burst is very accurate while the prediction of
the burst size requires an additional step to ensure its accuracy
119 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.1 Introduction
Water transmission and distribution pipelines are critical infrastructure for modern
cities. Owing to the sheer size of these pipelines and the fact that most of them are
buried underground, the health monitoring and maintenance of this infrastructure is
challenging. In addition, some water transmission pipelines cover long distances
through remote areas that are not easily inspected on a regular basis. To monitor these
systems, different noninvasive techniques have been developed to identify events that
may put the functioning of a pipeline at risk. These techniques include visual
observations (Thomson and Wang 2009), acoustic monitoring (Shimanskiy et al.
2003; Muggleton et al. 2006; Juliano et al. 2013), thermographic infrared inspection
methods (Fahmy and Moselhi 2010), ground-penetrating radar methods (Hunaidi and
Giamou 1998), and remote sensing (Agapiou et al. 2016; Martins et al. 2019).
However, these techniques are time-consuming, do not provide permanent monitoring
of the pipelines, or have a short inspection range along a given pipeline. Therefore,
there is a need for a permanent monitoring method capable of identifying anomalous
events and, potentially, their associated characteristics in near real time.
Hydraulic-based techniques have been developed based on the understanding of the
movement of a fluid along a pipeline and are typically related to the measurement and
analysis of two hydraulic variables: flow (or velocity) and pressure in the pipeline.
These techniques can include volume-based methods coupled with alert systems
(Mounce et al. 2003), pressure and flow analysis using statistical detection (Puust et
al. 2008; Li et al. 2014; Lee et al. 2016; Wu et al. 2018b), system state estimation
analysis (Andersen and Powell 2000), and transient-based methods (Misiunas et al.
2005). However, while each of these approaches has been moderately successful, they
also have associated disadvantages. Some of these methods are only applicable for the
detection of leaks, are not able to pinpoint the location of the abnormal event, or
require an accurate numerical model of the pipeline, or the resolution of the data limits
their ability to detect and locate anomalies in a timely manner.
Among the hydraulic-based techniques, transient-based methods have received
attention because they provide for the inspection of a long section of pipe using only
pressure measurements (Wang et al. 2002; Brunone et al. 2013; Gong et al. 2014a).
These methods are based on the interpretation of the effect that the occurrence of an
121 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
anomalous event has on a measured transient pressure trace. The interpretation of the
transient pressure trace can be conducted by visual analysis or by processing the
transient pressure trace to identify reflections and detect the occurrence of abnormal
events (Misiunas et al. 2007; Srirangarajan et al. 2011; Rashid et al. 2014; Rathnayaka
et al. 2016). Nonetheless, a visual analysis cannot be conducted in real time and the
processing techniques available require extensive numerical processing that makes a
real-time monitoring system difficult.
This paper presents a new transient-based technique that employs artificial neural
networks (ANNs) to interpret rapid changes in a transient head trace to identify, locate,
and characterize bursts in water pipelines. Unlike current transient-based techniques,
the proposed approach is data-driven and relies on no detailed information regarding
the analyzed pipeline for the interpretation of the transient pressure trace in near real
time. In addition, in contrast to many existing approaches, the proposed technique does
not use an ANN as a metamodel of the transient phenomenon but instead uses an ANN
as a tool to interpret the measured transient head traces to identify bursts. This
technique analyzes short time segments of a potentially continuous transient head trace
in previously trained ANNs through two different processes. A first ANN analysis
detects whether a particular transient pressure head time window contains abnormal
changes that could have been produced by the occurrence of a burst. This initial
analysis also determines whether the information contained in the time window is
enough to locate and characterize the potential burst. The second ANN analyzes the
detected abnormal transient head time windows from the first ANN to predict the
location and the characteristics of the occurring burst. Finally, the technique uses a
transient flow numerical model to verify that the predictions of location and size of
the burst are coherent and are similar to the measured transient head trace.
The approach proposed in this paper works with transient pressure head data to carry
out the training of the ANNs. These pressure head data may be obtained from recorded
data, from an available numerical model, or from a combination of these two sources.
Once the ANNs have been trained, the use of this technique is data-driven in that it
only requires measured transient head data to detect and identify a burst (by
determining its location and size). Additionally, it can be applied in near real time
considering that the computationally expensive processes are concentrated in the
122 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
training of the ANNs and not in the processing of new transient pressure head traces,
which can be tested almost immediately without retraining the ANNs.
One example hydraulic system is considered in this paper to demonstrate the
functioning and performance of the proposed technique. A single pipeline with the
potential presence of a burst at any point along its length is analyzed. A numerical
system with this configuration is used to present the operation of the technique and to
train the ANNs. In addition, a laboratory validation is included to demonstrate the
promising potential of the technique.
This paper includes a background in hydraulic and transient-based methods for
identifying and locating bursts in pipelines. The example hydraulic system is described
next, and the methodology to apply the proposed technique is then presented by
explaining the two stages involved in the process. The results for the training, testing,
and application of the proposed technique are presented for the numerical system.
Finally, the experimental validation is described. The proposed technique is shown to
be successful in analyzing a potentially continuous transient head trace to identify,
locate, and characterize the occurrence of a burst in a pipeline.
7.2 Background
The detection and location of bursts in water distribution systems is a complex task
for water utilities. Various authors have proposed techniques for locating bursts in
water pipelines using hydraulic-based methods. A first group of techniques includes
the use of flow and pressure measurements with statistical methods to detect the
occurrence of abnormal behavior in a system (Wu and Liu 2017). Ye and Fenner
(2011) proposed the use of an adaptive Kalman filtering process to predict flow (or
pressure) in a system at a district meter area (DMA) level. This statistical
characterization of a dynamic system is able to model the normal hydraulic parameters
that are compared to measured data to detect the occurrence of bursts. Similarly, Ahn
and Jung (2019) proposed a hybrid statistical model that combines statistical process
control with two univariate methods to enhance the performance of the burst detection
technique in terms of the detection probability, the rate of false alarms, and the average
detection time. Other authors have proposed the use of statistical risk functions (Cheng
et al. 2018) and principal component analysis (Palau et al. 2012) to detect bursts in
transmission mains and in DMAs. A different approach was described in Wu et al.
123 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
(2018a), where a clustering-based method was used to identify bursts using only 1 day
of historic measured data without using statistical methods to model the expected
normal conditions of the system. Although these techniques are successful in detecting
the occurrence of bursts, they are unable to accurately pinpoint the location of bursts
and their range of effectiveness is often limited to a DMA level.
A second group of techniques uses supervised learning techniques coupled with
hydraulic measurements to send an alert regarding the occurrence of a burst. Mounce
and Machell (2006) used two ANN architectures (static ANN and time delay ANN) to
detect the occurrence of bursts using flow data at a DMA level. The use of ANNs
showed potential for identifying changes in the flow that corresponded to unusual
fluctuations of this hydraulic variable. Mounce et al. (2010) proposed the use of
support vector regression models to predict time series data in a moving time window
and compare these series with measured data for the detection of anomalies. The use
of this supervised learning technique was applied to historical data proving that 78%
of the alerts corresponded to actual abnormal events in the system. Similarly, Romano
et al. (2014) proposed a fully automated data-driven methodology at a DMA level
using all the pressure and flow measurements available. This approach combined the
use of an ANN for the short-term forecasting of hydraulic values and statistical
processes to determine whether an abnormal event had occurred. The results obtained
showed the potential of data-driven technologies for near real-time incident
management in water distribution systems.
Other authors have proposed the use of artificial immune systems not only for
detection but also for an approximate localization of a burst. Tao et al. (2014) proposed
the use of pressure data every 10 min and was able to detect and localize a burst in
48% of the cases in a real network. More recently, Huang et al. (2018) proposed the
use of a random forest classifier to detect bursts in real time by analyzing successive
time windows (every 15 min) of flow data at a DMA level. These contributions
demonstrate the use of a supervised learning technique for the detection of bursts in
pipelines; however, all the applications to date have been focused at a DMA level and
using SCADA flow (or pressure) data, which are often available in intervals between
1 and 15 min.
124 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
A different group of techniques involves the use of fluid transients to detect the
occurrence of bursts in water systems. These techniques are able to detect and locate
anomalies in pipelines such as (Wang et al. 2002; Lee et al. 2007a; Lee et al. 2007b;
Capponi et al. 2017), blockages (Rubio Scola et al. 2017), or wall deterioration (Gong
et al. 2013c) and have obtained accurate results. Liggett and Chen (1994) proposed an
inverse transient algorithm to detect leaks using the method of characteristics (MOC)
and the Levenberg-Marquardt method to find friction and leak parameters in a system.
These authors proposed, for the first time, the use of transient pressure measurements
to locate pipe ruptures using an event algorithm with the capacity to approximately
pinpoint burst locations to locate additional nodes in the numerical model of systems
and conduct inverse transient analysis. Misiunas et al. (2005) used the principles of
time domain reflectometry for detecting and locating abrupt pipeline breaks using a
single pressure measurement point in a pipeline. A system of continuous monitoring
of the pressure at a high sampling frequency, coupled with a cumulative sum test and
prefiltering techniques, was used to detect changes in the data. In addition, an offline
analysis of a short time window was conducted to interpret the pressure changes to
determine the burst location. The results of this research have demonstrated that
transient pressure signals can be used to detect bursts in water systems.
More recently, Srirangarajan et al. (2011) described the use of a wavelet-based
multiscale analysis combined with a focusing algorithm and a graph-based search
algorithm to detect and locate a burst event. This technique showed the potential for
application of transient pressure techniques if the characteristics of the system are
known and presented the first application in a network layout. Other authors have
proposed the use of embedded and distributed event processing algorithms to detect
transient events in a system and stroke-based transient recognition algorithms to
classify transient events as bursts (Hoskins and Stoianov 2014). Although several
authors have proposed techniques that use transient pressure signals to detect and
localize bursts in pipelines, the existing techniques require offline processing that can
delay the detection of bursts or require specific prior knowledge about the systems.
The analysis of transient pressure signals using supervised learning algorithms has not
been widely explored. Bohorquez et al. (2020a) presented a technique that uses ANNs
to predict the presence of different features (leaks and junctions) in a pipeline after the
generation of a controlled transient event. The obtained results demonstrated the
125 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
potential of combining the principles of transient-based techniques with ANNs to
interpret the pressure traces. A similar application was proposed by Perera and
Rajapakse (2011) for the identification of transient faults in power transmission
networks using hidden Markov models and probabilistic neural networks (PNNs) to
classify transients as faults of normal switching events. Considering the current
literature, this paper presents a methodology that for the first time merges ANNs and
fluid transient waves to detect, locate, and characterize bursts in water pipelines.
7.3 Hydraulic System Configuration
The proposed methodology has been applied to a single pipeline under the assumption
that a burst can occur at any point along its length. The pipeline, of diameter ๐ท, is
connected at the upstream end to a reservoir, and at the downstream end there is a
closed inline valve (Figure 7-2). The length of the pipeline is ๐ฟ , and the location of
๐
the burst is characterized by a distance ๐ฅ measured from the upstream end of the
pipeline. The bursts are modeled as circular orifices of diameter ๐ท .
๐ต
Figure 7-2. Single pipeline with a burst system description.
Two pipelines with the configuration presented in Figure 7-2 were analyzed. A
numerical pipeline was considered with the characteristics presented in Table 7-1. In
this case, the head at the reservoir is defined as ๐ป = 55 m at the beginning of the
0
simulation, and a sinusoidal fluctuation of this head is considered to model gradual
changes in pressure head that can be observed in pipeline systems. Steady-state friction
was considered using a Darcy-Weisbach friction factor ๐with a pipeline roughness
of ๐ = 0.01 mm. The different burst locations and sizes that have been considered are
described in what follows, as part of the methodology (Steps A.2 and A.3 in Figure
7-3) of the proposed technique.
126 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
Table 7-1. Numerical pipeline characteristics.
Characteristic Units Value
Length of pipe (๐ฟ ) (m) 1,000
๐
Internal diameter of pipe (๐ท) (mm) 587
Wave speed of pipe (๐) (m/s) 1,111
๐ฟ /๐ time (s) 0.9
๐
A second pipeline was analyzed as part of the experimental verification of the
proposed technique. The characteristics of this pipeline are shown in Table 7-2 and
explained in detail in the experimental results section of this paper.
Table 7-2. Experimental pipeline characteristics.
Characteristic Units Value
Length of pipe (๐ฟ) (m) 37.24
Internal diameter of pipe (๐ท) (mm) 22.14
Wave speed of pipe (๐) (m/s) 1,290
Wall thickness (๐) (mm) 1.63
๐ฟ /๐ time (s) 0.029
๐
7.4 Methodology
The effect that a burst has on a transient head trace is characterized by a sharp drop in
the head that propagates in both directions away from the burst location (Misiunas et
al. 2005; Bohorquez et al. 2018). The proposed methodology for detecting and
identifying bursts is presented in Figure 7-3. Two stages have been included in this
diagram. The first stage comprises the ANN model development (Stage A in Figure
7-3), which is carried out first and can be updated regularly depending on the
availability of new pressure head data. The application stage (Stage B in Figure 7-3)
includes the required steps to process and interpret a continuous transient pressure
head trace using the ANNs for detecting, identifying, and verifying the occurrence of
a burst.
127 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
Figure 7-3. Burst detection and identification methodology.
7.4.1 Model Development Stage
To establish a near-real-time monitoring system for a pipeline, the first stage of the
proposed methodology includes the development of the burst detection and
identification model. This model is composed of two trained ANNs that can predict,
first, whether a particular analyzed transient head trace contains abnormal head
fluctuations corresponding to the occurrence of a burst in the pipeline and, second, the
location and size of this burst. As explained in this section, this model does not
represent a metamodel of the transient flow phenomenon that occurs in a pipeline after
a burst. The design, training, and testing of these ANNs are focused on interpreting a
measured transient head signal to detect the occurrence of a burst.
The steps described in this stage are required only once to set up a monitoring system
in a pipeline, which means that the computationally expensive processes are not
required for the application stage. However, if new transient pressure head data
become available (from numerical modeling or historical measured data), the ANNs
can be retrained to include any new information regarding the transient behavior of
the pipeline.
128 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.4.1.1 ANN Architecture Definition
The first step to developing a model capable of detecting and identifying bursts is the
definition of the ANN architecture (Step A.1 in Figure 7-3). Previous applications of
ANNs with transient fluid head traces have shown that a dense network (the most
general and widely used ANN structure) is not able to adequately capture the changes
in pressure due to the presence of different elements in a pipeline (Bohorquez et al.
2020a). Considering this, a one-dimensional (1D) convolutional network architecture
was chosen since these networks have fewer weights and are less prone to overfitting.
In addition, convolutional networks have shown successful performance in predicting
multiple outputs in fields such as image segmentation (Raza et al. 2017). These 1D-
convolutional networks are designed to perform satisfactorily both in the detection of
the occurrence of a burst and in the identification of the burst location and size. This
design was developed by modifying different characteristics of the ANN architecture,
including the number of convolutional layers, type of activation function, number of
filters in each layer, and training batch size.
The final configuration of the designed 1D-convolutional networks includes (1) a
maximum of seven convolutional layers, (2) the use of leaky rectified linear unit
(Leaky ReLU) and Softmax as activation functions, (3) a maximum number of 12
filters that increase in each convolutional layer, (4) a training batch size of 50 samples,
and (5) three dense layers of maximum sizes of 21, 9, and 3. With these characteristics,
the designed ANNs have between 26,808 and 81,250 weights to be trained.
7.4.1.2 Transient Head Trace Sample Characteristic Definition
The model development stage includes the training and testing of two different ANNs
(Steps A.7 and A.8 in Figure 7-3). The data used for these processes are transient head
data that can be obtained from recorded data, from an available numerical model, or
from a combination of these two sources (Step A.2 in Figure 7-3). In this paper,
numerically generated data were used for ANN training and testing to demonstrate the
application of the proposed technique. In the context of this paper, a sample is the
transient head trace that would be observed if a burst (with specific characteristics)
occurred at a specific location along a pipeline.
129 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
To train and test the ANNs, the spatial distribution and the characteristics of the
samples were defined to cover a range of values for the potential burst locations and
sizes. For the pipeline described in Figure 7-2, the transient head traces were generated
by modeling a burst located at random distances on average every 0.2 m along the
pipeline. This spatial distribution of samples was selected considering that generating
samples with a larger spacing decreases the effectiveness of the ANN training and
choosing a smaller spacing does not provide significantly better results, but it does
result in a larger computational effort (Bohorquez et al. 2020a). In addition, the use of
random distances was introduced to avoid the risk that the ANNs would learn only
from regularly spaced burst locations. Considering the previously described pipeline,
if the burst location is changed every 0.2 m, the total number of samples is 5,000.
Each one of these samples was assigned a random burst size. As was described in
Figure 7-2, the burst was modeled as a circular orifice with a diameter ๐ท . The range
๐ต
of possible burst sizes was defined considering the head drop that each burst size can
cause in the transient head trace. The smallest burst size considered (๐ท = 17 mm)
๐ต
causes a 1-m pressure head drop, and the largest burst (๐ท = 88 mm) causes a 20-m
๐ต
pressure head drop at the burst location.
7.4.1.3 Transient Head Trace Sample Generation
Various authors have proposed the use of numerically simulated data to train ANNs
that may be used to analyze real-time data (Perera and Rajapakse 2011; Zhou et al.
2019). In this paper, the transient pressure head samples for the ANNs training (Steps
A.7 and A.8) were generated using the Method of Characteristics (MOC). This
transient fluid calculation method transforms the two hyperbolic partial differential
equations that govern the behavior of unsteady flow into four ordinary differential
equations in order to obtain the variation of flow and head at different points along a
pipeline at a given time. Two of these ordinary differential equations describe the
relation between the time step (โ๐ก), the spatial resolution of the calculation (โ๐ฅ), and
the wave speed of the transient wave in the pipeline (๐). Only steady-state friction
was considered in the transient numerical modeling. The effects of unsteady friction
were neglected because the transient head trace samples only cover a maximum of
3.5๐ฟ/๐ seconds after the occurrence of the burst, and during these first seconds, the
130 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
effect of unsteady friction is not significant, as has been previously reported (Gong et
al. 2014a; Zhang et al. 2018).
Considering that the selected spatial separation between bursts is on average only 0.2
m, the selected time step needs to match this spatial resolution. The wave speed in the
selected pipeline is 1,111 m/s, therefore the required time step is at least โ๐ก = 1.799ร
10โ4 s. The total simulation time included a period of time before the occurrence of
the burst and at least 3.5๐ฟ/๐ seconds after the burst occurrence in order to capture the
first cycle of reflections of the burst wave in the pipeline and to perform the burst
occurrence consistency test (explained below at Step B.5). Thus, the total simulation
time was 4.04 s. This means that each of the 5,000 generated transient head traces had
more than 20,000 head values.
7.4.1.4 Downsampling and Processing
The potential size of the input data for the ANN training and testing when the sample
spatial distribution and the required time step are considered as described in the
previous step is approximately 100 million head values. Considering this, a timewise
downsampling (Step A.4 in Figure 7-3) was applied to the samples in the input dataset
since this has proven successful in the training and testing of ANNs with the ability to
interpret transient head traces (Bohorquez et al. 2020a). The downsampling frequency
selected was 256 Hz to match the potential sampling frequency in the field considering
existing technology.
Further processing of the downsampled transient head trace is required because the
proposed technique is intended to work in near real time. A sliding time window
concept (Mounce et al. 2010; Huang et al. 2018) is applied by partitioning each
transient head trace into time windows (moving one data point forward at a time) that
could contain the first period of transient wave reflections of the occurrence of a burst.
To accomplish this, the length of each time window must be at least 2๐ฟ/๐ seconds. In
this paper, this length was selected as 2.5๐ฟ/๐ seconds. Considering the downsampling
frequency, the length of each time window, and the total length of the transient head
trace (8.5๐ฟ/๐), each trace was transformed into 1,328 time windows. Therefore, the
5,000 transient head traces were transformed into 6.9 million time windows, each one
with a length of 577 head values (corresponding to 2.25 s).
131 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.4.1.5 Time Window Sample Classification and Selection
Once the partitioning of the transient head traces is complete, the resulting time
windows can be classified into three categories depending on the contained head
information, as shown in Figure 7-4. The first category is defined as Normal Head
Condition, or Category N, since these time windows only contain part of the assumed
slow sinusoidal head variation before the burst. Figure 7-4(a) shows the variation of
the head in two different scales to show the part of the sinusoidal head variation for
this particular time window. The second category is defined as Abnormal Head
Condition with Incomplete Information for Identification, or Category Ab-I. The time
windows in this category capture the initial head drop due to the burst, but the burst
wave reflection at the upstream reservoir is not included, as shown in Figure 7-4(b).
The last category is defined as Abnormal Head Condition with Complete Information
for Identification, or Category Ab-C. The time windows in this category [as presented
in Figure 7-4(c)] contain a complete reflection of the transient wave created by the
burst at the boundary conditions of the pipeline. Considering this, at Step A.5 of the
model development stage, this classification is used to characterize the available time
windows. For the pipeline described earlier, there are 3.5 million time windows in
Category N, 1.1 million time windows in Category Ab-I, and 2.3 million time windows
in Category Ab-C.
It is important to recognize that the total number of available time windows in the
three categories is very large. With the purpose of reducing the required time for the
ANNs training and facilitating the data management, only 20 time windows per
category were randomly selected for each of the 5,000 locations of the modeled bursts.
By conducting this selection, the total number of time windows was reduced to
approximately 300,000. Different numbers of selected time windows were assessed in
terms of the final ANN performance, and it was found that 20 time windows per
location provides good performance with a reasonable computational time.
132 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
Figure 7-4. Time window classification.
7.4.1.6 Burst Detection ANN Training and Testing
The methodology of the proposed technique includes the use of two different ANNs.
First, a burst detection ANN was trained to analyze each time window and allocate it
to one of the three categories described previously (Step A.7). The input dataset for
the training and testing of the burst detection ANN contains the 300,000 time windows
selected at Step A.6. This input dataset is then randomly divided into two groups, one
of which is used for ANN training and the other for ANN testing (50% training and
50% testing). As presented in Figure 7-5, the burst detection ANN receives as input
one time window at a time with the information of the corresponding category and
uses gradient search to find the best combination of weights to describe the training
dataset. The output of the burst detection ANN is the probability of each time window
133 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
of belonging to each category and the final category for each time window is assigned
as the category with the highest probability. This process is conducted using the
Softmax activation function (Goodfellow et al. 2016) which returns three integers (two
zeros and a one corresponding to the selected category). Once the training process is
complete, the testing dataset is processed on the burst detection ANN to obtain a set
of predicted categories, which can be compared to the correct category of each time
window.
7.4.1.7 Burst Identification ANN Training and Testing
The second ANN in the proposed methodology is referred to as the burst identification
ANN. This ANN was trained to analyze only time windows previously classified as
Category Ab-C. The training and testing of this ANN are included in Step A.8. The
input dataset for this ANN was defined as half of the complete time window dataset
allocated to the last category at Step A.5 (1.2 million time windows). Similarly to the
burst detection ANN, the input dataset was randomly divided into two groups for the
training and testing processes. The output of the burst identification ANN is the
location and the size of a burst, based on the interpretation of a particular time window,
as shown in Figure 7-5.
Figure 7-5. Proposed ANNs (N = Normal, Ab-I = Abnormal Head Condition with
Incomplete Information for Identification, and Ab-C = Abnormal Head Condition
with Complete Information for Identification).
134 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.4.2 Application Stage
The second stage of the technique presented in this paper (Stage B in Figure 7-3)
comprises the proposed steps that can be carried out to process and interpret a
continuous transient head trace in order to detect and identify bursts. The steps
presented in this section include the processing of a measured signal to be interpreted
by the model created in the model development stage of the technique and could
potentially be applied in near real time to a transient head measured signal.
7.4.2.1 Transient Pressure Head Data Retrieval
Once the burst detection and identification model has been created and validated, it
can be applied to a continuous transient head trace obtained from a given measurement
source. This transient head trace can be obtained from new and different numerical
simulations, measurements from a laboratory setup, or pressure head measurements in
a real pipeline. In either of the last two cases, a high-frequency pressure transducer is
required for capturing the head variations on a real-time basis. The selected sampling
frequency depends on factors such as the pipeline wall properties, the wave speed, and
the downsampling frequency selected at Step A.4. In general, the transient head
measurements should have at least the same frequency that was selected for the
training of the ANNs. The data retrieval system (Step B.1) should also include data
acquisition, processing, and communication modules to process the initial data
obtained from the pressure transducer.
7.4.2.2 Sliding Time Window Selection
As presented in the model development stage, the analysis of the transient head trace
is conducted by partitioning this trace into short time windows that are shifted one
point at a time as the head measurements are obtained. A particular time window must
be at least 2๐ฟ/๐ seconds long in order to capture the first set of wave reflections caused
by the occurrence of the burst. Step B.2 comprises this partitioning and the isolation
of one time window at a time to be analyzed following the subsequent steps of the
application stage.
135 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.4.2.3 Burst Detection Analysis
Once one time window of the transient head trace has been isolated, it is necessary to
determine whether this time window contains any abnormal head variation that could
have been caused by a burst. This burst detection analysis is composed of two different
steps. The first step (Step B.3 in Figure 7-3) includes the transformation of the current
time window to match the sampling frequency that was selected for the training of the
ANNs (Step A.4). As shown earlier, the downsampling frequency is selected based on
the sampling frequency available for the head measurements or an estimation of the
expected and required sampling frequency in the field.
The downsampled time window obtained from the continuous transient head trace can
then be analyzed in the burst detection ANN obtained in Step A.7 (at Step B.4). The
result from this ANN is a prediction of a category to characterize the current time
window. There are three possible results of this analysis. If the burst detection ANN
predicts that the current time window only contains normal head fluctuations, it is not
necessary to continue the analysis, and the next time window can be selected
(returning to Step B.2). This condition is defined as โNormal Conditionโ and in an
alert system can be represented by the color green.
If the burst detection ANN predicts that the current time window belongs to Category
Ab-I, then the condition of the pipeline is now defined as abnormal and can be
represented by the color orange. In this case, the analysis continues to the next time
window because for the predicted category it is known that this time window does not
contain enough information to locate and characterize the burst. Lastly, if the burst
detection ANN predicts that the current time window belongs to Category Ab-C, the
analysis process continues to the burst identification analysis module. In this case, the
condition of the pipeline is changed to โAbnormal ConditionโPossible Burstโ and may
be represented by the color red.
7.4.2.4 Burst Identification Analysis
Of all the possible time windows analyzed in the application stage, the only ones that
are analyzed in the burst identification module are those that are classified as Category
Ab-C. The main objective of this module is to determine the location and size of the
previously detected burst.
136 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
Once one time window has been defined in Category Ab-C, a consistency test is
conducted (Step B.5 in Figure 7-3). This consistency test determines how many
windows have been classified in this same category immediately before the current
time window. The main objective of this step is to provide a technique with more
robustness to possible misinterpretations from the burst detection ANN. For instance,
a particular time window might have been classified as Category Ab-C owing to
normal fluctuations of the head in the pipeline, but the following time windows are
again classified as Category Ab-I. In this case, the process continues until an invariant
classification in Category Ab-C is obtained. The consistency test is considered
complete once at least ๐ time windows are continuously classified in this category,
where ๐ corresponds to the number of time windows that cover ๐ฟ/๐ seconds for the
analyzed pipeline.
If the consistency test is completed, the analysis process continues to Step B.6, where
the previously trained burst identification ANN (at Step A.8) is used to predict a
possible location and size for the detected burst. It is important to mention that not just
the last time window is used in this step, but the complete batch of time windows that
have been included in the consistency test. Given that ๐ time windows are used to
obtain a prediction for the location and characteristics of the burst, ๐ different
combinations of predicted locations and sizes are obtained. Different statistical
measures were considered to achieve a final prediction, and the median was selected
because it is less sensitive to possible outliers in the prediction (potentially present due
to misinterpretations of the burst identification ANN).
7.4.2.5 Burst Verification Analysis
It is important to highlight that the proposed technique merges existing knowledge of
the impact of bursts in the transient head traces in a pipeline with the use of ANNs to
rapidly and more accurately interpret these traces. However, the use of ANNs is also
complemented by the use of transient numerical forward models to reinforce the
robustness to the predictions. This is the case for the burst verification analysis. This
analysis is comprised of two steps: a comparison of the measured transient head trace
and a numerical transient head trace generated using the burst identification ANN
predictions, and then a potential burst size prediction adjustment.
137 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
At Step B.7 (Figure 7-3), a numerical simulation of the transient head trace caused by
a burst with the predicted characteristics in the previous step is conducted. This
transient head trace is then compared with the measured transient head trace using a
metric such as the normalized root mean square error (NRMSE) to evaluate their
match. If the prediction from the burst identification ANN is accurate enough, the
NRMSE between both traces should be under a predefined threshold. In this case, an
alert is raised for the pipeline, including the predicted location and size of the burst.
For the examples presented in this paper the NRMSE threshold was defined as 2.0%.
In cases where the NRMSE between the two obtained transient head traces is above
the threshold, a potential correction in the burst size prediction is conducted. Multiple
tests demonstrated that when the predictions of the burst identification ANN are not
accurate, most of the time it is due to an error in the prediction of the burst size. To
correct for this, at Step B.8 different transient head traces are generated covering the
complete range of possible burst sizes to find the burst size that causes a transient
pressure trace that is very similar to the analyzed transient head trace (using again
NRMSE) and thus adjusts the final prediction. If the value of the comparison metric
improves after Step B.8, an alert will be raised with the original burst location
prediction and the adjusted burst size prediction. However, if after conducting the
burst size adjustment the NRMSE is still large, this means that none of the burst
characteristics was predicted accurately, and the analysis will continue with the next
time window at Step B.2.
The complete methodology described in this section was assessed using the numerical
pipeline described earlier in this paper and was validated with experimental data. The
results for these two applications are presented in the following sections.
7.5 Numerical Results
The methodology presented in Figure 7-3 was applied to the system described in
Figure 7-2 and Table 7-1. A burst detection ANN and a burst identification ANN were
designed, trained, and tested to detect and identify bursts. This section presents the
results for both the model development stage and the application stage.
138 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
7.5.1 Model Development Stage
Considering the multiple steps involved in the model development stage (Step A.8 in
Figure 7-3), samples of numerical transient head traces were generated at random
distances on average every 0.2 m along the pipeline to complete 5,000 transient head
traces in total. Once all the transient head traces were generated, these were
downsampled to a sampling frequency of 256 Hz and divided into time windows of
length 2.5๐ฟ/๐ seconds. A total of 6.9 million time windows were obtained (from Step
A.5) and 20 time windows per burst location were selected for the training and testing
of the burst detection ANN.
At Step A.7, a burst detection ANN was trained and tested using this input dataset.
Figure 7-6 shows the classification performance of the burst detection ANN into two
of the three possible categories. This accuracy is presented in terms of the percentage
of time windows that are misclassified at each considered burst location. In each
figure, the results are presented for the training and testing processes separately. It is
important to note that the total accuracy of classification in Category Ab-I is 96.07%
for the testing dataset and 99.17% for the classification in Category Ab-C.
Figure 7-6. Percentage of misclassified time windows for (a) Category Ab-I
(incomplete); and (b) Category Ab-C (complete)
Figure 7-6(a) presents the percentage of time windows misclassified in Category Ab-
I. As can be seen in this figure, the behavior of the results is similar in the training and
testing of the burst detection ANN, showing that there is no overfitting of the ANN to
the training data. In addition, it is possible to observe that when the burst is located
very close to the upstream end of the pipeline, the burst detection ANN has difficulty
139 |
ADE | Chapter 7 โ Fluid Transients and ANN for Burst Detection and Identification
classifying a time window as Category Ab-I. This is explained by the fact that the
transient head response of a burst located at this end only contains a quick head drop
that, after which it almost immediately recovers. Figure 7-6(b) shows the percentage
of misclassified time windows for Category Ab-C. This figure demonstrates that the
burst detection ANN does not present any significant difficulties in correctly
classifying a time window in this category in either the training or testing steps of the
development of the ANN.
Plots of the results of the accuracy in classifying a time window of Category N are not
presented because, for the training and testing of the burst detection ANN, the
accuracy in this category was 100%. At Step A.8, the burst identification ANN has
now been trained and tested. Results for this step are presented in Figure 7-7.
Figure 7-7 shows the median error in the estimation of the burst location [Figure
7-7(a)] and size [Figure 7-7(b)], depending on the actual location of the burst (burst
position). The median error was selected to present these results because several time
windows corresponding to the same burst location were used for the training and
testing purposes, and the use of this metric makes it possible to analyze an estimation
of the distribution of the predictions.
Figure 7-7. Median error in estimation of (a) burst location; and (b) burst size along
the pipeline.
These two figures demonstrate that the burst identification ANN does not overfit to
the training dataset given that the testing dataset predictions follow the same trend.
The maximum median errors in the prediction of the burst location were found when
the burst was located at end of the pipeline. However, 99% of the median location
prediction errors are smaller than 6.33 m (0.63% of the pipeline length). On the other
140 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.