University
stringclasses 19
values | Text
stringlengths 458
20.7k
|
|---|---|
Chalmers University of Technology | 4. Results: Business Understanding
volume parts may need to book more than one truck to carry all ordered parts from
material suppliers, and therefore these separated shipments increase the risks of de-
viation for all of them arriving on time. The hazardous parts or parts required to
clear custom result in more process during transportation, and these extra processes
may endure the fluctuation of processing time along with the delivery.
Demand for parts refers to the number of parts that requires to be shipped by LSP
from material suppliers. Different level of demand directly determines the way of
transportation. The high demand of a part could achieve the full truck load (FTL)
from one supplier, and simply adopt DDT transportation. That is different from
when there is less demand for a part. It brings in the order from a material supplier
is limited truck load (LTL) which has to go through cross-docking or milk run in
order to fully utilize the capacity and realize the cost benefits of shipments. Waiting
time at the cross-docking point or material suppliers throughout the milk run is a
high-risk factor of lead time deviation. This information of transportation solution
is available in Volvo. Besides, the receiving quantity of one spare part and the ac-
cumulated number of all parts in one shipment are available.
The value for one shipment could associate with the on-time delivery positively.
Since the higher value for one shipment, the more attention and priority it receives.
This attention and priority could help the shipment being processed earlier and get
rid of the extra waiting time. This record of value for one shipment is available at
Volvo.
Deviation related to LSPs
The second aspect belongs to LSP. LSP evaluation and historical delivery perfor-
mance could be associated with the deviation of TLT.
Similar to material supplier evaluation, LSP evaluation result is obtained in order
to evaluate the performance of LSP and this information can be used to predict the
delivery performance of LSP. At Volvo, SM perform LSP evaluation and generate
a final score for each LSP. The content of evaluation is including pickup and deliv-
ery precision, administration, deviation reporting in real time and communication.
However, not like SEM and logistics audit results for material suppliers which are
logged into the database, the scores of LSP are scattered in each evaluation report of
LSP and not integrated into databases. Therefore, this information is not likely to
be considered into the prediction model. The same situation exists in the LSP his-
torical delivery performance records. Lack of information also brings the difficulty
of estimating the ability of LSP handling uncontrollable disruptions of environment
and society such as labour shortage and storms.
Deviation related to supply chain
The third aspect is the supply chain including partners’ performance and the traffic
and weather information alongside the route.
46 |
Chalmers University of Technology | 4. Results: Business Understanding
Delivery precision of material suppliers could relate to the deviation of lead time.
Material supplier could order transportation booking earlier than the schedule to
get rid of the finished stock and cause the early arrival of orders. On the contrary,
waiting at material suppliers happens when LSP arrive at material suppliers but
suppliers are not ready to dispatch the orders. This situation could be prevented if
the material supplier communicate and update the delay information proactively. If
not, the delay at material suppliers is a key factor causing a delay in transportation.
The transportation in the company currently does not link with the corresponding
delivery precision from previous material supplier. Extra time is also generated from
missing documents from material suppliers such as proof of collection. The depar-
ture time also determines the arrival time at CDC, and the arrival time beyond the
operational hours at CDC will need to wait overnight to be unloaded. However, the
departure time is not accurately recorded in the company’s system.
Likewise, waiting at warehouse results from lack of capacity in the warehouse to
receive and unload the freight. This waiting time all relate to the scheduling issues
since other actors in the supply chain share the resource of the warehouse. The
capacity of the warehouse, however, is not integrated into the database either and
link with previous transportation at Volvo.
Thecountry ofmaterialsupplierscoulddeterminetheconditionoftransportationby
having different roads quality, geology. Besides, the political and economic situation
differs from country to country. Traffic and weather information all the way to the
destination of CDC also directly affect the transportation lead time. The country
information is available, while the weather and traffic information are not existing
at Volvo.
47 |
Chalmers University of Technology | 5
Results: Data Understanding and
Preparation
In this chapter, a data mining goal is first generated based on previous business
goal. Then the result of features selection is presented based on the result of relevant
factors of deviation.
5.1 Data Mining Goal
To realize this business goal, we need to set up the data mining goal accordingly. To
choose between regression and classification model, the regression generate contin-
uous value for lead time deviation, while the classification could give the outcome
of three classes, namely early, on time, and late. In terms of predictive capability,
the classification model predicts whether there will be deviation while the regres-
sion model can give more information on deviation including how much deviation
there will be. However, it could be more difficult to have a reliable result to be
the numerical values, considering the distribution of lead time deviation with the
majority of the case being on time which corresponds to the deviation to be 0. This
high portion of 0 could distort the result of regression, since it is difficult to learn
from not enough instances with different distribution of days of the deviation. In
comparison, classifying the output could accumulate a lot of instances to learn for
each class. The goal of data mining is thereby to generate two machine learning
models for predicting deviation in material supplier lead time and truck arrival time
respectively by testing various classification machine learning algorithms and eval-
uate their performance.
49 |
Chalmers University of Technology | 5. Results: Data Understanding and Preparation
5.2 Features Selection
To represent each factor related to lead time deviation, a list of relevant and avail-
able data is collected and examined for the selection of features. There are three
cases occurred in this process. Firstly, there are data that can directly represent
the factors such as using order hits to represent demand. Secondly, some factors do
not have data to directly represent them or the corresponding data are not avail-
able in the company, but there are some data may represent these factors indirectly.
For example, the material supplier’s prioritization for Volvo could be represented
by Volvo’s sales level spend on that supplier. However, there are some factors that
could not either find suitable data for indirect representation, such as historical
delivery precision performance and evaluation results of LSP. They are currently
scattered in different excel sheet for each LSP and not logged into the database.
The relationship between the factors and features is illustrated at Table 5.1. After
manually linking these two phases of transportation and material supply, previous
deviation of material suppliers is available, and the delivery performance from ma-
terial suppliers could affect the success of pick up for LSP.
The description and characteristics of features for SLT and TLT are presented in
Table 5.2 and Table 5.3 separately. Among data type, the number within parenthe-
ses represents the dimension of each categorical variable.
50 |
Chalmers University of Technology | 5. Results: Data Understanding and Preparation
Table 5.2: Current available features for SLT model
Source Features Description Data type
Segmentation Parts segmentation of parts in terms of Categorical variable
(adapted) function groups and life cycle, ‘slow A’, (11)
‘fast A’, ‘slow B’, ‘fast B’......
Spare
Vital Code The importance of the parts for 1- Prior- Categorical variable
Part
itized parts 2- Service Parts 3- Consump- (4)
tion Parts 4- Non-vital parts
Prepack The requirement whether a part needs to Binary variable
be prepacked before transporting or not
Standard Standard price for spare parts Continuous variable
price
Book off The amount of part ordered Continuous variable
Quantity
Order hits The historical order frequency for spare Continuous variable
parts
SEM result Supplier evaluation model score, measure- Continuous variable
ment including company profile, manage-
ment, environment, quality, logistics, af-
termarket, competence, product develop-
ment, finance, productivity, sourcing
Logistics Evaluating suppliers’ performance purely Continuous variable
Material
audit on logistics includinglead time agreement,
Supplier
Results production, planning information and etc.
of all logistics aspects.
Sales level The amount of money from Volvo spends Continuous variable
Spend on supplier
Delivery Supplier historical performance regarding Continuous variable
precision the percentage of parts delivered on time
Regions Theregionwheresuppliersarelocated, in- Categorical variable
cluding ‘EMEA’, ’APAC’, ’Americas’ (3)
Purchase Whether there is an agreement with ma- Binary variable
agreement terial suppliers including confidentiality
agreement, development agreement, price
agreement, warranty charter etc.
PPM Defective parts per million Continuous variable
QPM ApercentagecalculatedfromPPMforpro- Continuous variable
duction quality
Environment Productionstandardmeasurement,includ- Categorical variable
Certificate ing ‘IATF16949’, ’ISO17025’, ’ISO9000’, (5)
’ISO9001’, ’ISO9002’
Quality Quality standard measurement, ’QS9000’, Binary variable
Certificate ’VDA6’
52 |
Chalmers University of Technology | 6
Results: Models and Evaluation
In this chapter, based on the previous result of available features in the company that
are associated with the lead time, two prediction models of lead time deviations are
generated using various machine learning techniques. The results are presented by
classification report which include precision, recall, f1 score. The feature importance
for the most relevant features is also introduced
6.1 Classification Report
The confusion matrix result of deviation prediction model in supplier lead time and
truck arrival time are represented below in table 6.1 and table 6.2 respectively. The
precision score represents the accuracy of prediction. For example, with random
forest model for SLT, it predicts 25,826 (=13626+1824+10376) observation to be
‘Late’ while 10376 of them actually arrived late. The precision will be 0.4 out of 1
(=10376/25826). The recall score represents the missing of capturing the occurrence
of a class. For example, with catboost model for truck arrival time, it correctly pre-
dicts111observationtobe‘early’whilethereis176(=62+111+3)casesinfactbeing
early delivery. Therefore, the recall score is calculated as 0.63 out of 1 (=111/176),
which means 37% of early delivery is not predicted by the model to be ‘early’. The
higher the recall score, the lower the number of missing capture.
Table 6.1: Confusion matrix for SLT models (columns being predicted classes and
rows being actual classes)
catboost Gradient Boosting Random Forest
On Time Early Late On Time Early Late On Time Early Late
On Time 58617 16813 17006 90100 531 1805 67560 11250 13626
Early 2202 6377 1947 8012 2010 504 3012 5690 1824
Late 3333 2904 10994 13002 527 3702 4541 2314 10376
55 |
Chalmers University of Technology | 6. Results: Models and Evaluation
Table 6.2: Confusionmatrixfortruckarrivaltimemodels(columnsbeingpredicted
classes and rows being actual classes)
Catboost Gradient Boosting Random Forest
On Time Early Late On Time Early Late On Time Early Late
On Time 1408 115 135 2391 50 37 1302 182 174
Early 62 111 3 165 105 2 39 130 7
Late 131 13 109 256 4 120 108 20 125
As the classification report shows in table 6.3 and table 6.4 respectively, for the
SLT model, catboost outperforms the other two methods in terms of recall scores
for deviation prediction (0.61 for ‘early’ and 0.64 for ‘late’). The score is slightly
higher compared to random forest. While random forest overpasses catboost in F1
score where both precision and recall are taken into consideration. Focusing on the
deviation class, the F1 scores in early and late are less than 0.5, far from deploy
level which should be better at least above 0.8.
For the truck arrival time model, random forest has the highest score of recall (0.74,
for ‘early’ and 0.49 for ‘late’) while performing not that well in precision compared
to catboost. Similar to the first model, the average scores of each algorithm for the
deviation class are less than 0.5.
Table 6.3: Classification report for SLT models
Catboost Gradient Boosting Random Forest
Precision Recall F1 Precision Recall F1 Precision Recall F1
On Time 0.91 0.63 0.75 0.81 0.97 0.89 0.90 0.73 0.81
Early 0.24 0.61 0.35 0.66 0.19 0.30 0.30 0.54 0.38
Late 0.37 0.64 0.47 0.62 0.21 0.32 0.40 0.60 0.48
Total 0.78 0.63 0.67 0.77 0.8 0.75 0.78 0.70 0.72
Table 6.4: Classification report for truck arrival time models
catboost Gradient Boosting Random Forest
Precision Recall F1 Precision Recall F1 Precision Recall F1
On Time 0.88 0.85 0.86 0.85 0.96 0.90 0.90 0.79 0.84
Early 0.46 0.63 0.53 0.66 0.39 0.49 0.39 0.74 0.51
Late 0.44 0.43 0.44 0.75 0.32 0.45 0.41 0.49 0.45
Total 0.79 0.78 0.78 0.82 0.84 0.81 0.80 0.75 0.76
56 |
Chalmers University of Technology | 7
Discussion
In this chapter, the results of this thesis are interpreted and discussed. Firstly, the
implication of results on literature is discussed by comparing with current literature.
Secondly, the implication of the result on the case company is discussed. Finally,
the underlying assumption and limitation are presented.
7.1 Implication on literature
This thesis project works on two areas in the theory. The first one is predicting
material suppliers’ delivery precision performance. The second one is predicting the
deviation of truck arrival time.
7.1.1 Predicting Material Supplier Delivery Precision
Although a company’s performance is much affected by its suppliers’ performance
includingdeliveryprecision(Krauseetal., 2007), veryfewliteraturehasinvestigated
the evaluation and prediction of supplier performance in operational level during the
periodofcooperation. Thismayduetothecomplexrelationshipbetweenthesuppli-
ers’performanceandseveralcriteriaofsuppliers(Rezaeietal., 2014). However, with
powerful analysis tools such as machine learning and a large number of instances,
this complex relationship could be examined. For example, Khaldi et al. (2017) and
Jiang et al. (2013) implement machine learning algorithms to evaluate and predict
suppliers’ overall performance based on their historical performance data in several
aspects such as delivery, costs, quality.
This thesis project specifically focuses on predicting supplier delivery precision per-
formance using machine learning. The input features that have been used to train
the prediction model are including not only the supplier historical performance and
evaluation information but also the information of parts ordered from that supplier.
Since different parts have different characteristics which could relate to the difficul-
ties of production, and therefore these characteristics further relate to the deviation
in production.
59 |
Chalmers University of Technology | 7. Discussion
When we examine the result of modelling for SLT deviation, however, the prediction
power is not enough with poor precision and recall score of models on ‘Late’ and
‘Early’ classes where the deviations locate. That means with current features of
parts and supplier information, machine learning models still cannot well capture
the relationship between the occurrence of deviation and these features. That may
due to in majority case, the deviation could result from the production disruption
such as machine break down, labour shortage, and waiting time in the production
line. This information currently is only owned by material suppliers themselves and
not able to be utilized by a buyer company.
7.1.2 Predicting Deviation of Truck Arrival Time
The deviation of transportation is most likely to be subjected to the weather and
traffic situation. The successful implementation of machine learning on predicting
the deviation of the airplane could purely based on the weather information at de-
parture and arrival airport (Belcastro et al., 2016). Whereas, predicting the truck
arrival time based on only weather and traffic information does not achieve a good
result(vanderSpoeletal., 2015). Whenconsideringthefullnetworkstateincluding
physical characteristics of the train and train crew information for predicting the
arrivaltimeoffreighttrain, machinelearningbringslargeimprovementinprediction
(Barbour et al., 2015). To interpret these differences, we can consider and compare
the causes of deviation in each transportation mode. For air transportation, one of
the major causes of the deviation comes from the weather (Belcastro et al., 2016),
while the freight train is less prone to be impacted by the weather but more likely
to be affected by the scheduling of train network. Therefore, the common thing in
these two successful implementations is that they manage to make the causes of de-
viation into features for the prediction model. In contrast, for the arrival time of the
truck, it is affected by not only weather and traffic but also could be largely affected
by factors such as the scheduling information from the consignor and consignee of
truck transportation. Only providing weather and traffic information is not going
to make machine learning model capture the pattern of deviation.
This thesis considers that lead time deviation of the truck is related to the schedul-
ing of transportation and prone to the situation from both consignor and consignee,
that is, for example, the deviation of delivery from the consignors could affect the
pickup precision for LSP and further impact the delivery precision to the consignees.
However, the deviation from consignor used as a feature in the model does not get a
high feature importance score as expected, this may party due to that the deviation
from is only structured as a classification feature with three dimension of ‘Late’,
‘Early’ and ‘On time’ instead of the exact number of deviation.
This thesis also innovatively considers the logistics characteristics of the transported
cargo and further the size of shipment including weight, volume, units. It turns out
some of them highly contribute to the performance of the prediction models. The
scheduling method of transportation such as the truckload and delivery method also
contributes to the prediction model.
60 |
Chalmers University of Technology | 7. Discussion
Although we consider some features of organization and cargo into prediction, and
some of them are much likely to be associated with deviation as feature impor-
tance indicates. This proves that the organizational and cargo features are relevant.
However, the model performance is still limited due to the unavailability of some
important features in the company. A successful prediction model of trucks’ ETA
may not stand alone without considering traffic and weather information.
7.2 Implication on the Case Company
Volvo strives to maintain a high delivery precision from its suppliers. In terms of the
KPI delivery precision, the early dispatches and deliveries should not be regarded
as fulfiling the delivery precision since they are also harmful according to our inves-
tigation.
This thesis mainly investigates predicting the deviation precision from their ma-
terial suppliers and LSP. For these two phases, Volvo has more power in inbound
transportation than material supply. Since Volvo has a platform for scheduling and
coordinating all the transportation which could make all the transportation infor-
mation available, while Volvo currently does not have production information of
their material suppliers. Therefore, while in this project the two phases share the
same goal which is to aid the monitoring process by generating deviation alerts,
the results from these two phases will not have the equal implication on the case
company. In this section, the implication of two prediction models on the company’s
monitoring is discussed separately.
7.2.1 Monitoring on Material Suppliers
While in this phase, the limitation on the modelling is related to the feature selec-
tion for production information in material suppliers. Considering improving the
prediction model performance, information sharing with material suppliers regard-
ing their production disruption is necessary in the future.
The models still achieve some prediction power for deviation of SLT. Therefore, the
feature importance generated for the models may deserve an examination for their
close relationship with deviation (Trevor et al., 2009). For example, for selected
two models’ most important features (Catboost and Random Forest), they share all
top seven important features with a slightly different ranking. It is most likely that
the features of evaluation results from suppliers including delivery precision, SEM
result, logistics audit result, sales level spend are negatively related to the deviation
of lead time. That is the lower the performance of these indicators, the more the
deviation there tends to be.
To examine the influence of characteristics of parts on deviation, for example, the
standard price of parts, as the below Figure 7.1 shows, 44% of orders with the spare
61 |
Chalmers University of Technology | 7. Discussion
parts valued more than 10,000 has deviation while this figure is only 21% for spare
parts with price lower than 1,000. This indicates the higher standard price of the
spare parts, the more likely the deviation of SLT could happen. That may due to
the higher the price, the more complicated of the parts to produce which bring in
the risks of deviation. Therefore, the CMP can pay more attention on the spare
part order with expensive price.
Percentage of Deviated Supplier Delivery by Price
50%
44%, 8381
45%
40%
35%
29%, 83829
30%
25% 23%, 400461
21%, 306407
20%
15%
10%
5%
0%
<1000 1000-10000 >10000 Average
Figure 7.1: Percentage of deviated supplier delivery by price
Percentage of Deviated Transportation by Countries
35% 33%, 1395
30%
25%
20%, 10433
20%
15%, 1932
15%
10%, 3391
10%
5%
0%
Average France Germany Sweden
Figure 7.2: Percentage of deviated transportation by countries
7.2.2 Monitoring Process on LSP
Similar to the monitoring on material suppliers, in the short term, LSP could exam-
ine the relationship between the most important features and lead time deviation.
For example, we take a close look at the deviation by countries indicated by the high
feature importance of ‘France’, ‘Sweden’, ‘Germany’. As Figure 7.2 shows, compared
to average deviation case of being 20%, 33% of transportation from ‘France’ are de-
viated. Therefore, LSP can give special attention on transportation from France to
62 |
Chalmers University of Technology | 7. Discussion
CDC. On the contrary, the transportation from Germany and Sweden to CDC are
performing way better, especially for Sweden, only 10% of truck transport has the
deviation which is 10% less than the average.
In the long term, a dynamic deviation alert could be embedded in the system to
assist the TMC to monitor the performance of LSP. This alert can help TMC pre-
ventively reach LSP to figure out whether there will be a deviation according to the
prediction. This alert could require LSP to examine their operation statues. TMC
could also examine the delivery precision statues from their consignors and the ca-
pacity situation at the consignee warehouse. If there could be a deviation, TMC can
help to take actions to prevent the deviation. If the deviation is irresistible, such
as the extreme weather, some corrective actions could be scheduled to alleviate the
influence brought by the deviation.
Transport Coordinator
Transport Collection Transport Performed
Monitor & React on Group Analyze & Log
Transport Deviations Material Supplier
Manager Specialist
Deviation Perform Material
Material
cause? Supplier Root Cause
Supplier
Analysis
Multileg
Transport
type?
Cross Dock Support
Specialist
Performance followed up
Direct flow or delivery to Logistics Service
ultimate consignee Provider
Supplier Manager
Monitor & React on Perform Logistics
Arrival Service Provider
Root Cause Analysis
End Performance followed up
Figure 7.3: Generating a deviation alert in the process of monitoring LSP
In order to construct a reliable deviation alert, some improvement needs to be
achieved. Since the alert is generated by the machine learning prediction mod-
els, the performance of model on ‘Late’ and ‘Early’ classes is important to improve.
The precision score affect the accuracy of the model and therefore affect the relia-
63 |
Chalmers University of Technology | 7. Discussion
bility of the alert generated, while the recall score reflects the degree of miss capture
of these two classes which means the miss capture of the deviation (Sokolova and
Lapalme, 2009). The recall score could be even more important considering Volvo’s
availability performance. In order to improve the performance of the prediction
model, the linkage of the database between the material supplier phase, inbound
transportation phase and internal warehouse phase should be constructed. It means
transportation booking information should contain the ordered part information.
The optimal result of constructing the database is that when we input one trans-
portation booking ID, all the part number in this shipment will present with all
needed feature information regarding the parts linked, such as their standard price
and order hits. Besides, all the features needed in the supply chain including the
material suppliers (consignor), LSP, warehouse (consignee) should also be linked in
one click. Furthermore, in order to generate a dynamic and reliable alert, the open
source weather and trafficforecast informationshould also be addedin the database.
This result of this new dataset considering all the linkage is shown in Figure 7.4.
Shipment Weather & Traffic Warehouse
TB actual weight Weather forecast along the route Capacity forecast
TB actual volume Traffic forecast along the route
Logistic Service Provider
TB actual unit
Evaluation results
Historical delivery performance
Parts
Transport method
Value
Quantity
Database for
Material Supplier
Transportation
Stackable
Model Evaluation results
Hazardous
Historical delivery performance
Delivery precision
Figure 7.4: Recommendation for linkage of prediction model for truck arrival time
deviation
7.3 Underlying Assumptions
The data from the company database we used are assumed to be accurate and re-
liable. Regarding SLT analysis, we use the same source of data with the logistics
analysts in the company. For material suppliers’ information, we extract data from
the supplier management portal as the SRM are using the same portal. Regarding
truck arrival time, the logistics service portal of Volvo is the place where transporta-
tion data are compiled so it is regarded as a reliable source of accurate data. The
false of this assumption is inaccurate data, which could affect the performance of
the prediction model. However, the effect could be subtle since the missing features
could be the main bottleneck to improve model performance.
64 |
Chalmers University of Technology | 7. Discussion
Another assumption is that the sampled data are representative. For SLT deviation
performance, we extract data of two years data from 2017 and 2018. For truck ar-
rival time, the past rolling one year data is extracted. These data are regarded to be
representative for the whole population. It is assumed the lead time data from each
year is homogeneous and the result can be generalized to the future period. That
means the patterns between the selected features and lead time deviation remain
stable for each year. Even though this relationship could be evolved as years pass,
those features with high importance score are most likely to be important, only the
degree of importance could be reexamined each time when building the model.
This thesis also assumes that for SLT, those causal factors of production disruption
which directly result in deviation cannot be assessed by a buying company. Further-
more, we assume the data the company have can indirectly represent those causal
factors to some degree. However, according to the result of the models generated,
the poor prediction performance indicates that this assumption of the above rep-
resentativeness could be false. The causal factors of production could be not well
represented by the substitute information available on the buying company. This
could lead to a conclusion that, when there is limited access to production informa-
tion, the deviation of SLT could not be well predicted with the machine learning
algorithms from the buying company’s perspective.
7.4 Limitation
The first limitation comes from the scope of this thesis, with only focus on one
business area and one geographical region, this could limit the size of data and
the complexity of features. The performance of the prediction model is most likely
to be affected by the amount of data and dimensions of features. Besides, since
the characteristics of outbound logistics could be very different from the inbound
logistics, with only lead times in inbound transportation investigated, this thesis
project cannot be directly generalized to the outbound phase without adjustment.
This thesis also only use the database from the case company, that result in the
traffic and weather information is roughly represented by the origin country. This
information could be accessed from an open source database and considered in the
prediction model for the future.
The feature selection and data preprocessing also have some limitation. There is a
small portion of ‘outliers’ in historical deviation performance such as late or early
up to 4 years. The reason for these strange numbers is not examined. However,
since the proportion of these outliers and the percentage of deviation classes are
both small, and the outputs of these instances are very likely to be late or early.
Therefore, they are kept in the dataset belonging to either ‘Late’ or ‘Early’ classes to
contribute to the machine learning of minority classes. The influence of this way of
handling potential outliers could be small. Furthermore, for handling missing values
in the features, this mainly exist in the evaluation results of material suppliers. Only
65 |
Chalmers University of Technology | 7. Discussion
average values are used to replace them, which may result in a poor estimation of
variances and correlations for the feature (Schafer and Graham, 2002). Considering
the pattern of the missing value is unclear to us for now, therefore advanced impu-
tation methods remain to investigate. However, as we discuss before, the missing of
key features would be the main reason for the poor performance of deviation predic-
tion model. The exact representation of missing value in those evaluation features
would not improve the prediction model to a large extent.
There is limitation existing in the modelling process. This thesis is only implement-
ing classification models rather than regression models. Considering even though
the regression model could generate the prediction of the exact time of deviation,
this exactness also increases the difficulty of regression models to achieve a better
and reliable result compared to classification models at this very first stage. How-
ever, the construction of classification models could also be improved by further
increasing the granularity of the response variables into more classes such as ‘Very
Late’, ‘Very Early’ to further increase the informativeness of classification models.
This improvement is not tested due to the time limit of this data mining project.
Another limitation in modelling is the lack of examination of the feasibility of algo-
rithms. Since the results of modeling with tested algorithms are not good enough
and closed to each other, the performance is believed to highly relate to the fea-
tures. The algorithms used in modelling are regarded as feasible and optimal from
the knowledge of literature review and expertise of data scientists based on the char-
acteristics of input features and output. For future model improvement, however,
algorithms should be re-evaluated for their feasibility and optimisation for new mod-
elling after the improvement of features.
The SLT modelling also does not take time series into account, since the evaluation
information from the buying company towards their material suppliers last for a
long period. For the two years period, even though considering the event time for
each order, the variance of features could be very low for each instance of the same
spare parts. Therefore, the result is not likely to be influenced by the time factors.
66 |
Chalmers University of Technology | 8
Conclusion
In this chapter, the research finding is summarized by answering four research ques-
tions, followed by the practical and theoretical contribution of this thesis project. Fi-
nally, the recommendations and future researches for the case company and academia
are presented in the end.
8.1 Research Finding
The research questions are covered through analysis and discussion in this thesis
work, therefore these answers are stated below as a summary of the research finding.
RQ 1: What are the benefits of predicting lead time deviation for buying com-
panies?
Certain roles in buying companies are responsible for monitoring the delivery pre-
cision of suppliers and taking actions to deal with deviations. For example at Volvo
SML, continental material planners have the responsibility for monitoring the lead
timesfor thesuppliers. Likewise, transportmaterialplanners monitorthe leadtimes
performance of LSP. The prediction of lead time deviation can help create deviation
alerts that assist those monitors for monitoring suppliers’ delivery performance, and
the benefits of the alerts are to reduce the deviation and decrease the impact of
deviation by taking preventive and corrective actions.
RQ 2: What are the factors that could be associated with lead time deviation
perceived by buying companies?
Thefactorsrelatedtothedeviationofleadtimefromthebuyingcompany’sperspec-
tives can be categorized into three levels. The first one is the part level regarding the
characteristics of parts in demand, criticality, value, life cycles and function groups
for material supplier lead time and the part level regarding the logistic character-
istics including volume, weight, hazardous, custom, stackable, demand and value.
The second level regarding the supplier level, it represents the evaluation results and
historicalformaterialsuppliersandLSPforsupplierleadtimeandtruckarrivaltime
respectively. These evaluation performances are covered in multi-criteria aspects of
the suppliers with special focus on delivery performance. This level may also in-
67 |
Chalmers University of Technology | 8. Conclusion
clude the priority of the buying company within the material suppliers. The third
level of factors could be exclusively existing for transportation lead time which is
related to the actors in the supply chain level including their consignors, consignees
and buying company. The deviation of truck arrival time is affected by the delivery
performance of material suppliers which affects pick-up precision of LSP. The route
of shipments scheduled by the buying company could affect lead time. Besides, un-
loading shipment is also subjected to the capacity of warehouses. The country and
environment including weather and traffic are most like to be relevant.
RQ 3: Which data are available to be used as features when building the pre-
diction model of lead time deviation at Volvo SML?
To turn the above factors into features for modelling, there are a few cases occurred
in this thesis project. There are data which can directly represent the factors such
as the demand, value, stackable, hazardous, custom, evaluation results for material
suppliers. There are data representing the factors at an aggregated level, such as TB
weight and volume data for the total weight and volume in one shipment, segmen-
tation data for integrating function groups and life cycles, country for traffic and
weather. Some factors that are not recorded in the data form, such as the prioritiza-
tion, some factors are not available in the buying company due to that information
is owned by material suppliers such as material suppliers or not integrated into the
databases. These factors are tried to be indirectly reflected by other available data,
such as sales spend level data on suppliers for the prioritization, quality and envi-
ronment certificate for the production capacity of suppliers. However, some factors
could not either find suitable data for indirect representation, such as historical de-
livery precision performance and evaluation results of LSP.
RQ 4: How should the prediction model be built using machine learning con-
sidering the practicality of use in the current stage at Volvo SML?
Through the business analysis and data analysis, the goal for a prediction model
at current stage could be generating a deviation alert for monitors in the buying
companies. Aclassification modelwith theoutput ofthreeclasses ‘OnTime’, ‘Late’,
‘Early’ could achieve this goal. However, modelling with currently available features
for both two phases do not deliver deployable results. In order to improve the re-
sults of modelling, more representative features should be added for capturing the
pattern of deviation. For predicting truck arrival time at CDC Ghent, since Volvo
has more power in this phase, most of the key features can be filled in and improved
with Volvo’s efforts. The prediction model for truck arrival could be improved and
put into use in the future when databases are constructed as expected as figure 7.4
shows. However, for the SLT model, the key missing features could be the produc-
tion disruption information. Predicting material supplier lead time with machine
learning from the buying company perspective is, therefore, regarded as not practi-
cal until production information could be shared with the buying company.
68 |
Chalmers University of Technology | 8. Conclusion
8.2 Practical Contribution and Future Work for
the Case Company
Thisthesisforthefirsttimedemonstratesthepossibilityforpredictingthedeviation
ofleadtimeforVolvotrucksparepartsincludingsupplierleadtimeandtruckarrival
time at CDC Ghent with machine learning. The prediction model on the deviation
of truck arrival time shows its potential deployment after the future improvement
of data linkage for input features, while the model of supplier lead time deviation
is believed to be improved only when some key features including the production
information are accessible from the material suppliers.
In the short run, the feature importance generated by machine learning models al-
ready gives insights into the relationships between the deviation and some of the
most relevant features. Through the examination of these relationships, some char-
acteristics of orders demonstrate much more possibility of having deviation. This
could help monitors in Volvo such as CMP and TMC selectively pay more attention
to orders with these characteristics, and preventively react on the deviation and take
precautious actions.
Regarding the future work for the Volvo, the key moves are related to data manage-
ment. Firstly, some important information has not been logged into the database.
For example, logistics evaluation results of LSP is not integrated into the database
and therefore not able to contribute to the prediction model. The second move is
related to linking the data from different phases and construct them into one com-
mon database for the benefits of data preparation for modelling. For example, since
there is no part number information in the transportation booking information, we
have to manually link the order from material suppliers to LSP. The requirement for
the data linkage is demonstrated in Figure 7.4. The mapping of currently available
data in the company for the relevant features is demonstrated in Chapter 5. Ex-
ternal data sources such as weather and traffic information are required to generate
a dynamic and real-time prediction for ETA truck arrival time. Besides, the data
should be stored and managed for a longer period. For the transportation phase,
current platform merely stores data last for at maximum one rolling year, which
could not be enough for machine learning modelling.
When the construction of the database is done and the performance of the pre-
diction model is improved to a deployment level as expected, the prediction model
could be deployed as deviation alert embedded in the program. The informative-
ness of the deviation alert could be improved by increasing the granularity of output
classes. The regression machine learning models could also be tested with all the
desired features available. This also requires the transportation booking records to
generate the information of the length of deviation in days rather than the current
three classes of being early, on time and late, since the regression models require
continuous values as the response variable.
Another improvement that could be made to pave the road for data mining is to
69 |
Chalmers University of Technology | 8. Conclusion
add a detailed description and the responsible role for each item in the database.
We discover during the project that the clarity of data in the company’s current
database is not satisfying. With no further description for each item especially for
those items with similar names, it increases the difficulty for data practitioners to
select the correct data.
Future improvements in predicting supplier lead time should be under the condition
of information sharing between suppliers and buying companies. Volvo could con-
stantly search for opportunities forproductioninformation sharing with its suppliers
or assist them to predict the lead deviation from their perspective and utilize the
prediction result.
8.3 Theoretical Contribution and Future Research
within Academia
The first theoretical contribution of the thesis is that it is the first trial in the theory
forthebuying company topredictmaterialsupplier’sdeliveryprecision performance
by machine learning. It shows that constructing the characteristics of ordered parts
and material’s supplier evaluation results as well as historical delivery performance
into the input features only deliver a weak prediction power. It indicates production
information of material suppliers such as the production disruption of the orders is
necessary to fully capture the deviation of supplier lead time. The second con-
tribution comes from the second prediction model of truck arrival time deviation.
The factors of truck arrival time deviation are investigated and sorted. The devia-
tion could be associated with the logistic characteristics of the cargo, the delivery
performance of consignors and the capacity of consignees, the macro-environment
including countries of consignors, as well as traffic and weather condition. It turns
out logistic characteristics of cargo are important features. The consideration of
organizational factors is much under the constraints of available data currently in
the company.
For the future research of predicting supplier lead time in academia, the idea of
predicting supplier performance with machine learning could be expanded to other
industries where buyers have more information regarding suppliers. The perfor-
mance is also not limited to the delivery precision and inbound logistics phase since
machine learning has the ability to identify the complex relationship between the
performance and multi-criteria of suppliers. For future research of predicting the
truck arrival time, comprehensive factors should be considered as Figure 7.4 shows.
The regression model should be considered when important features are ready since
the outcome of a regression model is more powerful. The output being the exact
number of deviation could help evaluate the default lead time in the system and
reschedule the safety stock.
70 |
Chalmers University of Technology | A. Appendix: Interview Questions
A
Appendix: Interview Questions
Interview Questions Template
Introduction:
● We describe the purpose and scope of this thesis work.
● We describe the purpose/expected output of this interview
Background Question:
● What is your role in the company?
● What is your responsibility and daily tasks?
● Are you responsible for a certain part segments, or part of supply chain or part life-cycle?
Lead time in inbound process:
● What processes are the inbound delivery process including? How long does each process
take (your lead time)?
● How is this lead time generated in your system? Do you or your department set up the
lead time? If yes, how? How often do you do the planning? Which data do you use?
● How much is the deviation between your theoretical lead time and real lead time? How
often do they delay or ahead of time?
● How does these deviation affect your work and the company in your mind? Who will
take your lead time prediction into consideration when they do their job?
● Is there any certain type of spare part (criticality/frequency/price) or any
carriers/forwarder with the largest lead time deviation?
● Regarding the later process about LSP performance, how do you perceive the pick-up
precision of LSP? (For material supply)
● Regarding the previous process about supplier performance, how do you perceive the
actual time of spare parts ready to be shipped? (For transportation)
● What are the factors do you think that are influencing your lead time? Which factors do
you have available data to measure?
- Internal factors
- External factors
II |
Chalmers University of Technology | Abstract
The transport sector is facing a major challenge in meeting future demands for increased energy
efficiency. In the future, most likely new technologies such as hybridization, electrification and
use of alternative fuels will play an increasing role.
The main purpose for carrying out this study has been to investigate how the environmental load
of the road bound trucks transporting iron ore powder from the mine in Kaunisvaara, Pajala, to
Svappavaara would be altered by a possible electrification. This has been done through an
attributional comparative LCA study of three different drivetrain alternatives for heavy duty
trucks, during 16 years of mining operation. The first alternative is a conventional truck with an
internal combustion engine, constituting the reference alternative which the other two are
compared with. The second alternative is a parallel hybrid electric version of the same truck
without external charging and the third is a parallel hybrid electric version of the same truck
receiving electricity from an overhead conductive catenary.
A handful of components in the parallel hybrid and catenary hybrid drivetrain have been
identified, including a lithium-ion battery and an electric motor of different sizes. These
components were studied throughout their life cycle: raw material extraction, production of
components, drivetrain assembly, use phase, recycling and disposal. For the catenary hybrid
alternative also the production of the extra needed infrastructure has been taken into account.
In order to quantitatively assess the environmental impact of these different phases, five different
environmental impact categories have been used: global warming potential, abiotic depletion, and
emissions of hydrocarbon, emissions of particles and emissions of nitrogen oxides.
The results show that both the parallel hybrid and catenary hybrid are better solutions than the
conventional truck in general. However, the catenary hybrid is the more favorable choice in both
impact categories as well as for the studied emissions. The use of electricity instead of diesel
provides enormous savings in environmental impact.
Furthermore, it is shown that the largest contribution to the environmental load for the parallel
hybrid clearly comes from the Li-ion battery. This is due to the amount of advanced materials
included in the battery and that raw material extraction of these materials is very energy
consuming. For the catenary hybrid it is the infrastructure which has the largest environmental
impact. This is due to the large amount of material that is used and, again, the environmental
impact comes mainly from the raw material extraction.
The life cycle phase that has the largest environmental impact is clearly the use phase with its
enormous savings in fuel and thus also in environmental impact. However, even large changes in
energy consumption do not change the final choice of the most favorable solution. Also when
changing the electricity mix to a lot dirtier production the catenary hybrid is still outperforming
both the reference vehicle and parallel hybrid.
This life cycle assessment does not provide values with exact precision for the final results.
However, more important is that the final results and conclusions are very robust.
iii |
Chalmers University of Technology | Sammanfattning
Transportsektorn står inför en stor utmaning att möta framtidens krav på ökad energieffektivitet. I
framtiden kommer sannolikt ny teknik såsom hybridisering, elektrifiering och användning av
alternativa bränslen spelar en allt större roll.
Det huvudsakliga syftet för att genomföra denna studie har varit att undersöka hur miljö-
belastningen för lastbilar som transporterar järnmalm från gruvan i Kaunisvaara, Pajala, till
Svappavaara skulle påverkas av den en eventuell elektrifiering. Detta har gjorts genom en
orsaksinriktad (attributional) och jämförande LCA-studie av tre olika alternativ för drivlinor till
tunga lastbilar, under 16 års gruvdrift. Det första alternativet är en konventionell lastbil med en
förbränningsmotor och den referens som de andra två alternativen jämförs med. Det andra
alternativet är en parallellhybrid, dvs. en elektrifierad version av samma bil men utan möjlighet
till extern laddning och den tredje är en helelektrisk parallellhybrid, en elvägshybrid, av samma
bil, men då elektricitet fås från en luftburen kontaktledning.
En handfull komponenter i parallellhybridens och elvägshybridens drivlinor har identifierats,
däribland ett litium-jon-batteri och en elmotor i olika storlekar. Dessa komponenter har studerats
under hela deras livscykel: utvinning av råmaterial, tillverkning av komponenter, montering av
drivlinan, användningsfas, återvinning och avfallshantering. För elvägshybriden beaktas också
produktionen av den nödvändiga tillkommande infrastrukturen.
För att kvantitativt kunna bedöma miljöpåverkan av dessa olika livscykelfaser har fem olika
miljöpåverkanskategorier använts: potentialer för global uppvärmning och abiotisk
resursanvändning, samt emissioner av kolväten, partiklar och kväveoxider.
Resultaten visar att både parallellhybriden och elvägshybriden är bättre lösningar än den
konventionella lastbilen. Elvägshybriden är dock ett mycket fördelaktigare val sett till både
miljöpåverkanskategorierna och de studerade emissionerna. Användning av el istället för diesel
ger enorma besparingar i miljöpåverkan.
Dessutom visas det att det största bidraget till miljöbelastningen för parallellhybriden kommer
från Li-jon batteriet. Detta beror på att mängden av avancerade material som ingår i batteriet och
dess energikonsumerande råvaruutvinning. För elvägshybriden är det infrastrukturen som har den
största miljöpåverkan. Detta beror på den stora mängd material som används och att
miljöpåverkan även då kommer främst från råvaruutvinning.
Den livscykelfas som har störst påverkan på miljön är användningsfasen med sina enorma
besparingar i bränsle och därmed också i miljöpåverkan. Även stora förändringar i
energiförbrukningen förändrar inte det slutliga valet av den mest gynnsamma lösningen. Även
vid byte av elmix till en smutsigare produktion är elvägshybriden fortfarande den mest
gynnsamma lösningen.
Denna livscykelanalys ger inga slutresultat med hög numerisk noggrannhet. Däremot, och
mycket viktigare, de slutliga resultaten och slutsatserna är mycket robusta.
iv |
Chalmers University of Technology | 1. Introduction
The transport sector is facing a major challenge in meeting future demands for increased energy
efficiency. To achieve a significant reduction in energy consumption development and/or
changes in many areas are required. Today, for example, a lot of work is done to train drivers to
drive more energy efficient, to optimize the existing drivetrains and to vehicle aerodynamics. In
the future, most likely new technologies such as hybridization, electrification and use of
alternative fuels will play an increasing role.
The focus of this project is set on issues related to the drivetrain in a life cycle perspective. The
aim is to increase the knowledge about environmental aspects of plausible future technologies for
energy-efficient transportation with heavy commercial vehicles. Thereby the project also aims to
provide support for decisions and strategies for vehicle electrification and the ability to meet
energy efficiency targets and at the same time decrease the overall environmental load of the
product in a larger systems perspective.
1.1 Background of the Iron Ore Mine Pajala Project
The global demand of iron is increasing, with the consequence that iron ore mines in operation
will increase their activity, and that mines that today are in fallow will be resumed. Also,
exploration of new iron ore deposits can be expected. (Trafikverket, 2012)
As a result, Northland Resources AB is planning to start mine operation in Kaunisvaara in Pajala
in 2013, creating a need for a transportation solution for the iron ore powder. The current plan is
to transport it by truck from the mine in Kaunisvaara to Svappavaara and then by train to the port
of Narvik. Preliminary 1,5 million tons of iron ore powder will be transported from the mine
annually, increasing to around 4,6 million tons from 2015. Because of the new mining
establishment, the Swedish Government gave Trafikverket the mission to investigate the
conditions for upgrading the standard of roads between Kaunisvaara in Pajala and Svappavaara in
Kiruna municipality, and the ability to meet the arisen transport need and its effects.
The increased traffic will become a major burden on the existing road network, which today is
not good enough for transport of this kind of weight, and large improvements of the current
infrastructure must be made. The assignment to Trafikverket also included an analysis of the
possibilities to carry out the transportation with trucks that run on electricity from the road.
(Trafikverket, 2012)
There are three main stakeholders for electrification of heavy trucks for the transportation
solution in Pajala, and their main interests are:
1. The mining company – Expects a significantly reduced cost due to both anticipated
reduced maintenance on the trucks’ engines and significantly reduced energy cost.
2. The automotive industry – Receives a demonstration facility. If it proves to be successful,
it could contribute to development of the electrification system that could become a major
export product for both the automotive industry and other stakeholders.
1 |
Chalmers University of Technology | 2. Life Cycle Assessment Methodology
In this chapter the life cycle assessment (LCA) concept is going to be explained and presented. It
is an overviewing summary based on The Hitch Hiker’s Guide to LCA (Baumann & Tillman,
2004) with additional remarks of the ISO standard 14040 series (ISO, 2006).
2.1 Introduction to LCA
LCA is used to analyze the environmental impact of products and services. The whole life cycle
of a product is followed, from its “cradle to grave”. The process starts with the extraction of raw
material from natural resources, the “cradle”, continues with production and use, and ends when
the product reaches the disposal, the “grave”.
Natural resource use and pollutant emissions are described in quantitative terms as showed in the
Figure 2-1 below.
Figure 2-1 The life cycle model (Baumann and Tillman, 2004)
The LCA methodology contains a whole framework for how such studies are done and
interpreted, as showed in Figure 2-2 below. First, the purpose and the object of study for the LCA
are specified in the goal and scope definition. Second the inventory analysis implies the
construction of the life cycle model and that emissions and resources used are calculated. In the
third phase, the impact assessment, relations between emissions, resources used and various
3 |
Chalmers University of Technology | environmental problems are established through the act of impact classification and
characterization. In the end of this phase the different environmental impacts may be put on the
same scale through weighting, making it easier to compare the results.
Figure 2-2 The framework of an LCA according to the ISO standard 14040 (2006)
The strength of LCA is that it takes the whole product system in consideration. This avoids the
sub-optimization that may occur if only a few processes are focused on. It allows comparisons
between alternative products since the results are related to the function of the product, rather
than the product itself.
The data collected for an LCA can either represent some type of geographical average or be site
specific depending on the type of process that specific data represent as well as the purpose of the
study. For example, in the case of a raw material, which may have many different suppliers, an
average value is often preferable and if it is traded on a global market it is best represented by a
global average. It is rather seldom necessary to go too deep in the level of detail, when an
indication of the environmental load often is preferable. Another reason is that the data collection
is very time consuming and rather complex.
LCA includes normally only environmental aspects. However, in the impact assessment, different
optional weighting procedures may include both economic and social considerations, when the
relative importance of various environmental impacts is given values on one single scale.
LCA is also an important tool when learning about the relationships between the processes or
when exploring the environmental properties of the studied product system.
2.2 Goal and Scope Definition
In the goal and scope definition phase of the LCA study, the purpose and the content of the study
are decided upon. This part is crucial – it is important that the goal and scope definition are stated
4 |
Chalmers University of Technology | correctly and the questions asked are as precise as possible. This is because different purposes
require different methodology and different methodological choices give different results.
2.2.1 Goal of the Study
The goal of an LCA study should clearly state the reasons for doing the study, the intended
application and to whom the results should be presented to.
It is important that the goal is stated and explained in the beginning of the study, but as seen in
Figure 2-2, LCA is an iterative process. This means that the purpose of the study can be redefined
and additional purposes can be added along the way. Preferably, purposes are stated as questions,
and some examples of how they may be phrased questions are given below (Baumann & Tillman,
2004):
Where are the improvement possibilities in the life cycle of this product?
Which are the activities in the life cycle that contributes the most to the environmental
impact associated with this product?
What would be the environmental consequences of changing certain processes in the life
cycle in such and such a way?
What is the environmentally preferable choice of products A, B and C for a designated
application?
2.2.2 Scope of the Study
The scope will define the system boundaries of the LCA study that are needed to answer the
stated goal, according to the ISO standard (2006). Again, as LCA is an iterative process and also
the scope of the study can therefore be redefined along the way.
2.2.2.1 Flowchart
To get a general idea about the system studied, an initial flowchart normally is constructed at this
stage of study. It does not need to be very detailed, since it is intended to provide an
understanding of what is going to be covered in the study.
2.2.2.2 Functional Unit
The functional unit is a reference unit, used to normalize all the inputs and outputs included in the
study. This is done to be able to compare all the aspects included in the study.
A process could have more than one possible function. However, the selected one should be
dependent on the goal and scope and be measurable. Also the function, corresponding to the
functional unit, describes the product or process purpose.
For stand-alone LCA studies, studies of single products or processes, the choice of the functional
unit is seldom critical. In comparative studies, however, it may be a more difficult task since the
functional unit is then used for comparison between the different alternatives. Then the functional
unit must represent the function of the compared alternatives in a rational way, and all the
different alternatives must fulfill the function rather well.
5 |
Chalmers University of Technology | 2.2.2.3 Impact Categories
It is important to consider which environmental impacts to take into account in LCA studies. The
ISO standard (2006) prescribes three main groups of categories which should be covered:
resource use, ecological consequences and human health. These are then often divided into
subcategories, for example global warming, acidification and resource depletion to mention some
of them. However, not all studies take all possible categories into account – it is a part of the
scope definition and depends on the earlier stated purpose and intended audience of the study.
Also, which data to collect follow the choice of impact categories as not all emissions contribute
to all types of environmental impacts and vice versa.
2.2.2.4 Different Types of LCA
There are two main types of LCA studies which are called attributional and consequential. The
first one answers questions like “What environmental impact can this product be held responsible
for?” or “What is the difference in environmental impact between different alternatives?”. It
examines the effects of or difference in effect between products or services when they are in
operation, reflecting the causes of the system. The consequential type answers questions of the
type “What would happen if …?” and it compares the foreseeable environmental consequences of
a decision by modeling the effects of change. The attributional type of LCA has normally a
retrospective approach in time, since it compares the causes of a system and consequentially has
a prospective approach in time, due to the fact that it models the effects of a future change.
When deciding what type of LCA to use it helps to analyze how the outcome of the LCA will be
used. As seen in Table 2-1 below, a summary of which type of methodology goes with which
type of LCA.
Table 2-1 Characteristics of accounting type and change-oriented LCA models (Tillman 2000)
Type of LCA/ Attributional Consequential
Characteristics
System boundaries Additivity Parts of system affected
Completeness
Allocation procedure Reflecting causes of a system Reflecting effects of change
Partitioning System enlargement
Choice of data Average Marginal (at least in part)
System subdivision - Foreground & background
2.2.2.5 System Boundaries
System boundaries are defined during the goal and scope part of the LCA. However, exact details
can often not be decided until enough information has been collected during inventory analysis.
According to Baumann and Tillman (2004) the system boundaries need to be specified in several
dimensions:
6 |
Chalmers University of Technology | Boundaries in relation to natural systems,
Geographical boundaries,
Time boundaries,
Boundaries within the technical system:
o Boundaries related to production capital, personnel, etc.
o Cut-off criteria, means that flows with negligible influence on the main results can
be ignored – they are “cut-off”.
o Boundaries in relation to other products’ life cycles. Requires allocation
procedures.
2.2.2.6 Allocation problems
Several products or functions can share the same process during their life cycles as products are
linked in networks. This can cause an allocation problem, i.e. how should the environmental load
be distributed between the products.
There are three basic cases when allocation problems are encountered:
1. Multi-output. Processes that result in several products. An example is a refinery process.
2. Multi-input. Waste treatment processes, e.g. landfill, that have input consisting of several
products.
3. Open loop recycling. When one product is recycled into a different product. Quality
losses are often a result of open loop recycling. Some examples of this type of process are
recycling of food packaging into other types of packaging and recycling of energy from
waste incineration.
Allocation problems can sometimes be avoided through increasing the level of detail in the
model. However, this is only true for multi-output and multi-input, but not for open loop
recycling.
2.2.2.7 Data Quality Requirements
Data quality requirements should be defined so that the goal and scope of the LCA study are
reached. Depending on which data that is used the LCA will give different results and also
provide different reliability of the results. When talking about data quality Baumann and Tillman
(2006) discusses relevance, reliability and accessibility as main features.
Relevance describes whether the used data actually represent what it should. When the
availability of suitable data is limited and approximate data are borrowed from other LCA
studies, the relevance might become a problem. Furthermore, Baumann and Tillman (2004)
explain that precision is an aspect that is connected to reliability of data. It concerns the
numerical accuracy and uncertainty of data. However, reliability also depends on the consistency
with which it has been collected and documented and also on the competence of the person or
organization that collected the data. Data are more credible if they can be reviewed and that is
7 |
Chalmers University of Technology | only possible if they are documented transparently. Transparent documentation also supports the
reproducibility of the results and the reproducibility is connected to the accessibility of the data.
Major assumptions and limitations of the study should be described and explained in the goal and
scope definition. However, some limitations can be a result of problems that comes up along the
study. One problem could be the project resources, for example if there is not enough time to go
deep enough in the level of detail. Then the limitations of the study will be updated. This is
another example that shows that LCA is an iterative proves.
2.3 Procedural Aspects and Planning of an LCA Study
Before setting up an LCA some important activities need some considerations. These are
described in this subchapter.
2.3.1 Critical Review
Critical review is a method to verify whether an LCA study has met the requirements for
methodology, data and reporting according to the ISO standard (2006), although generally
regarded as optional by the standard. However, if the results of the LCA study should be revealed
to the public, the ISO standard states that it must be conducted.
Different types of critical reviews can be made, from internal or external experts, or by interested
parties.
2.3.2 Actors in an LCA Project
The practitioner is the one conducting the LCA study; it can be a single person or a group of
people. Often there is also a commissioner separate from this group, who requests the LCA study
to be made. The commissioner has the task to state the goal of the LCA, but it often becomes too
vague to use as a basis for the study. For this reason the goal and scope definition should be seen
as an interaction between commissioner and practitioner so that the general goal is rephrased to a
more specific purpose, preferable in the form of a question. Another part that is discussed
between the two is the planning of the study so it will stay within the time and budget constraints.
Since the largest amount of work is the data collection, it is important to plan the data collection
in detail. For example which data sources to use for which parts of the life cycle.
2.4 Inventory Analysis
Life cycle inventory analysis, LCI, involves data collection and calculation procedures to
quantify relevant inputs and outputs of a product system according to the ISO standard (2006).
The inventory analysis includes the construction of a detailed flow model of the technical system.
It is an incomplete mass and energy balance, where only the environmentally important and
relevant flows are taken into consideration. LCI models are simplified to become static and
linear, and do not use time as a variable.
8 |
Chalmers University of Technology | The activities of an LCI include (Baumann and Tillman, 2004):
1. Construction of the flow chart according to the system boundaries decided on in the goal
and scope definition.
2. Data collection for all the activities in the product system followed by documentation of
collected data.
3. Calculation of the environmental loads (resource use and pollutant emissions) of the
system in relation to the functional unit.
As data is collected during this part it is sometimes necessary to revise decisions taken during the
goal and scope definition, and as mentioned earlier this leads to an iterative process between
these two parts.
2.4.1 Construction of a Flow Chart
In the goal and scope definition a general flow chart was constructed. In the inventory analysis
the flow chart is made much more detailed. It shows all modeled activities and the flows between
them. As mentioned before, the inventory analysis is an iterative process which means that when
more information is collected about the processes it will lead to revisions in the flowchart with
more details.
2.4.2 Data Collection
The most time consuming part of the LCA study is the data collection. It can be very difficult to
find data about some activities in the studied process and then assumptions and limitations are
inevitable.
2.4.2.1 Which Data
Numerical data on the inputs and the outputs to all modeled activities need to be collected. These
are inputs of raw material and energy, inputs and outputs of products and emissions to air, water
and land and other environmental aspects. Also descriptive and qualitative data need to be
collected to support allocation.
2.4.2.2 Data Sources
There are many different technical procedures taking place in a life cycle and it is almost
impossible for the practitioner to be an expert on all the fields represented. Therefore, experts
within specific subjects need to be consulted and communication between the expert and the
practitioner is then very important. Of course there are other data sources as well, e.g. different
data bases, other companies and so on.
2.4.2.3 Planning for Data Collection
Before collecting data the practitioner needs to know in advance how the processes work to be
able to ask the right questions and be prepared for dialogue with various experts. Also the
practitioner must decide for which processes average data is preferred and where site-specific
data need is necessary.
9 |
Chalmers University of Technology | Another aspect of importance is to have a strategy for addressing and handling confidentiality
issues with suppliers and decide if the suppliers will be given the opportunity to see how their
data is used and interpreted.
2.4.2.4 Validation of Data
The ISO standard (2006) requires a check of the validity of the collected data. Such checks can
be done by comparisons with other data sources, or by performing mass and energy balances. The
practitioner shall also check if the data quality requirements formulated in the goal and scope are
fulfilled, and if the data sets collected are within the system boundaries.
2.4.3 Calculation Procedures
When all the data is collected and the flow chart is drawn completely an LCI is calculated in the
following steps (Baumann and Tillman, 2004):
1. Normalize data for all the activities where data have been collected.
2. Calculate the flows linking the activities in the flow chart, using the flow representing the
functional unit as a reference.
3. Calculate the flows passing the system boundaries, again related to the flow representing
the functional unit.
4. Sum up the resource use and emissions for the whole system.
5. Document the calculations.
2.4.4 Allocation procedures
In previous subchapter allocation problems are described. The solution is referred to as an
allocation procedure and there are two main methods: allocation through partitioning or by
system expansion. Partitioning means that the up-stream resource consumption and emissions
associated with the multiple processes are divided between them based on for example weight.
System expansion means that the system boundaries are enlarged, for example that an industrial
system is credited with the environmental load from the heat production that is avoided in the
district heating system when it receives surplus heat.
The ISO standard (2006) sets out following procedure for dealing with allocation:
1. Whenever possible, allocation should be avoided by:
a. Increasing the level of detail of the model
b. System expansion
2. Where allocation cannot be avoided, the environmental loads should be partitioned
between the system’s different products or functions in a way which reflects the
underlying physical relationship between them.
10 |
Chalmers University of Technology | 3. Where physical relationship alone cannot be established or used as the basis for
allocation, the inputs should be allocated between the products and functions in a way that
reflects other relationships between them.
Baumann and Tillman (2004) support a different method for allocation procedures based on the
reason for the product’s existents. This puts focus more on the economic aspects than the ISO
standard (2006). One example could be extraction of gold, where the largest part of the extracted
mass is another mineral. Therefore, if the ISO standard rules are followed, the largest percent of
the environmental load should be allocated to the mineral and not to gold. However, the reason
for extracting anything at all from that mine is that it consists of gold, and this can be taken into
account in the study by using the economic values of the mine products as the basis for the
allocation. Baumann and Tillman (2004) also say that the allocation method should depend on
which type of LCA that is carried out. They think that partitioning is applicable to attributional
LCAs and system expansion is more relevant for consequential LCAs. This is also shown in
Table 2-1.
2.5 Impact Assessment
Life cycle impact assessment (LCIA) aims to describe the environmental consequences of the
environmental loads in the LCI. The LCIA is achieved by using the results from the LCI and
translate the loads into impacts, such as acidification, ozone depletion, global warming,
toxicological impacts on human health, effect on biodiversity, etc. These impacts are sub-
categories to three general categories of environmental impacts often considered in LCA studies:
resource use, human health and ecological consequences.
One reason for making this translation is that it is easier to communicate the results from an
LCIA than an LCI. People not familiar with LCA or environmental systems analysis, probably
relate easier to the potential for acidification than to SO .
2
Another reason is to improve the readability of the results; the number of parameters from the
LCI can be up to 200 or more, but it can be reduced significantly to about 15 when using LCIA,
or even down to one by weighing across the impact categories.
In this sub-chapter the methods of doing an LCIA are described and discussed.
The LCIA are divided into mandatory elements and some optional element, as seen in Figure 2-3
the below (ISO 2006).
11 |
Chalmers University of Technology | Figure 2-3 The LCIA are divided into mandatory elements and optional element (Baumann and Tillman, 2004)
The core sub-phases of LCIA are classification, characterization and weighting and they are often
used when carrying out an LCA. When using ready-made LCIA methods many of the sub-phases
are already included inside the method. There are several of ready-made LCIA methods and the
practical advantage of them is that the environmental information for various pollutants and
resources are aggregated to a characterization indicator or even a single number, an index
including weighting. The procedure of the impact assessment is already inside the ready-made
LCIA methods.
2.5.1 Impact Category Definition
In this part the set of impact categories will be described in more detail than in the goal and scope
definition. When deciding which impact categories to include several things should be considered
as completeness, practicality, independence, possibility to integrate in the LCA calculations,
environmental relevance and scientific method (Baumann and Tillman, 2004). Some conflicts
may occur when considering all of these aspects, and then a decision on the best solution must be
made by the practitioner.
12 |
Chalmers University of Technology | 2.5.2 Classification
In this sub-phase of the LCIA the LCI result parameters are sorted and assigned to the various
impact categories. The easiest way to find out which parameter that should be assigned to which
impact category is through published lists. In these there are various substances together with
their equivalency factors listed per impact category.
For example CML-IA is a database that contains characterizations factors for LCIA (Institute of
Environmental Sciences, CML). In some cases certain environmental loads corresponds to more
than one impact category and need to be listed in all affected categories. However, this can only
be done when the impact categories are independent of each other, otherwise it will lead to
double counting.
2.5.3 Characterization
This sub-phase is a quantitative step. When calculating the magnitudes of potential environmental
impacts per category, the equivalency factors, which are defined by expert sources such as IPCC,
are used while modeling the cause-effect chains. For example, if acidification is the impact
category, all the emissions causing acidification (SO , NO , HCl, etc.) in the LCI are added up
x x
based on their equivalency factors. The sum of all these emissions is an indication of the potential
acidification impact.
2.5.3.1 Methods for Characterization
Characterization methods translate environmental load into impact, based on scientific methods,
from chemistry, toxicology, ecology, etc. For pollutants a combination of their physicochemical
properties and their effect in the environment are considered. For resources, land use, noise,
casualties, etc. other modeling principles, based on occurrence or frequencies, are used.
Since the environmental system is rather complex, some impact categories have several
alternative characterization models. There are also categories where characterization factors are
lacking or have incomplete sets of equivalency factors.
The emission-caused impacts such as acidification, eutrophication and global warming, have
more developed characterization methods than those mentioned above. When such methods are
lacking, conditional assessment factors can be developed for the specific study in order to
evaluate the results from the inventory analysis. Otherwise the impact assessment points out
certain environmental impacts and neglects others (Baumann and Tillman, 2004). These
assessment factors could be developed by LCA practitioners with good environmental
knowledge. However, it is more common to separate the results which is lacking on the
characterization method from the others and present them under different heading, for example
flows to other technical systems.
2.5.4 Normalization
In this step the results from the characterization are connected to the actual or predicted
magnitude for each impact category. Normalization is carried out when a comparison between
the results in relation to one specific studied system can give more transparency. The aim of
13 |
Chalmers University of Technology | normalization is to gain better understanding of the size of the environmental impacts that is
caused by the studied system.
2.5.5 Grouping
In this sub-phase the characterization results are sorted into one or more sets. However, this is a
noncompulsory step, and it is occasionally unnecessary, but it can make the results more
transparent. It can also be very useful for the analysis and presentation of the results. For example
the characterization results can be sorted into global, regional, local impacts or impacts with high,
medium and low priority.
2.5.6 Weighting
This sub-phase can be defined as the qualitative or quantitative process where the relative
importance of the different environmental impacts is weighted against each other. Furthermore,
the relative weights are expressed by their weighting factors.
The environmental harm of a pollutant or a resource is indicated relative to other pollutants or
resources in ready-made LCIAs. Characteristic for the weighting methods is that all
environmental problems are measured on a single scale, and therefore it is possible to calculate
the total impact of a system into one number. It is obtained by multiplying all environmental
loads of the system by their corresponding indices and summing them up.
∑
According to the Baumann and Tillman (2004) the methods for generating the weighting factors
are mainly based in the social sciences and on several kinds of principles, for example the
European Network for Strategic Life-Cycle Assessment Research and Development (LCANET)
and European Environmental Agency (EEA). These are:
Monetarisation. With this method our values concerning the environment are described as
the cost of various kinds of environmental damage or as the prices of various
environmental goods. One important aspect is how values are described for goods for
which there is no market (therefore no price). However, a price can be derived from
peoples willingness-to-pay (i.e. one question asked can be how much they are willing to
pay to avoid extinction of a species, for example) or revealed by their behavior (e.g the
difference in price of similar houses close and far away from an airport reveals the cost of
noise coming from the airport).
Authorized targets. This method uses the difference between current levels and target
levels of pollution, and it can be used to derive weighting factors. The target levels can be
formulated by national authorities as well as by companies. This approach could be said
to be based on a so called distance-to-target thinking.
14 |
Chalmers University of Technology | Authoritative panels. In this approach panels are used. They can be made up of, for
example, scientific experts, government representatives, decision makers in a company or
residents in an area. The panels typically describe and rank various types of impacts so
that weighting factors representative for their view can be derived.
Proxies. With a proxy method, one or a few parameters are stated to be indicative for the
total environmental impact. Energy consumption and weight are some examples of proxy
parameters.
Technology abatement. The possibility of reducing environmental loads by using different
technological abatement methods (e.g. filters, etc.) can be used to set weighting factors.
This method can be said to be based on a distance-to-technically-feasible-target thinking.
There will never be a consensus in the weighting element in LCIA, since both ethical and
ideological values are involved. Many engineers therefore have an awkward relationship to
weighting, and the use of weighting factors often lead to discussion about whether they are
“scientifically correct” or not, whether the values are representative or not, etc.
Many LCIA methods are described in principle, but lists with indices for various substances have
been developed for a small number of them. Those methods use different means, i.e. weighting
principles, to obtain the one-dimensional indices. Each method reflects different social values and
preferences, since the determination of the relative harm of different environmental impacts is a
value-laden procedure.
2.5.7 Data Quality Analysis
To better understand the significance, uncertainty and sensitivity of the LCIA results additional
analysis may be needed, for example relevant sensitivity analysis. This is a part of the
interpretation phase of the LCA where findings from the LCI and LCIA are combined together in
order to reach conclusions and recommendations. These findings can also be used for reviewing
and revising the goal and scope of the LCA as it is an iterative procedure. According to Baumann
and Tillman (2004) these techniques are used in order to identify:
the most polluting activities in the life cycle.
the most responsive inventory data, i.e. a sensitivity analysis, where the data describing
the activities in the life cycle for which minor changes in value change the ranking
between compared alternatives.
the most responsive impact assessment data, i.e. a sensitivity analysis.
the significance of alternative methodological choices, e.g. different types of allocation,
also a type of sensitivity analysis, and
the degree of uncertainty in the results (uncertainty analysis). Uncertainty is introduced to
the calculations when input data are estimates, intervals or probabilities.
15 |
Chalmers University of Technology | 3. Drivetrain Technology
In this chapter the three alternative drivetrain solutions are going to be explained more in detail.
The first alternative is a drivetrain from a conventional truck with an internal combustion engine
(ICE) and a transmission system. The second alternative is a parallel hybrid electric vehicle with
the same ICE and transmission system as the conventional truck but with additional electric
components described in this chapter.
The last studied alternative is a truck which uses electricity supplied through conductive transfer
from overhead electric wires. The drivetrain in this alternative consists of the same parts as in the
conventional alternative plus an extra large electric machine and a pantograph. In this last
alternative there is also infrastructure needed to deliver the electricity to the vehicle which must
be included in the LCA.
3.1 Drivetrain of a Conventional Truck
Internal combustion based motor vehicles have been built for more than a century and engines are
an affordable and well-established power source. Below the drivetrain of a conventional truck is
going to be presented.
3.1.1 Drivetrain Architecture
The drivetrain of a truck often have the following torque-transmitting elements (Basshuysen and
Schäfer, 2004):
A internal combustion engine.
A torque converter
A gearbox. Consisting of an initial movement element and the actual speed-reduction
gearing system.
A final drive gearing systems.
Possibly an integrated starter-motor/alternator (ISG).
In the case of four-wheel drive a power divider will also be included.
Other functions are balancing the engine and vehicle traction requirements, reduction of engine
rotation irregularities and distribution power to the wheels.
The distribution of power is achieved by an exchange of torque between the front and rear axles.
The gearbox adapts the movement of the engine to a significantly larger, and required, torque.
(Basshuysen and Schäfer, 2004).
3.1.1.1 Internal Combustion Engine
Internal combustion engines are energy conversion devices which extract stored energy in a fuel
through the combustion process and deliver mechanical power (Husain, 2011).
16 |
Chalmers University of Technology | ICEs for conventional vehicles are designed to operate over a wide range of power and torque,
and compromise is often necessary to deliver acceptable efficiency and performance throughout
its operating regime.
ICEs used in automobiles, trucks and buses are of the reciprocating type, where the reciprocating
motion of a piston is converted to linear motion through a crank mechanism.
There are two types of reciprocating ICEs, and they are the spark-ignition (SI) engine and the
compression-ignition (CI) engine. The two types are more commonly known as gasoline/petrol
engine and the diesel engine depending on the fuel used for combustion. The main difference
between the two is the method of initiating the combustion. (Husain, 2011).
To start with the SI engine, a mixture of air and fuel is drawn in and a spark plug ignites the
intake of the engine, which are called the charge. To help to measure the right amount of fuel in
response to driver demand, electronic control devices are used. The SI engines are relatively
light, lower in cost and used for low power applications, as in conventional cars. (Husain, 2011)
In the CI engine, which is more suitable for heavy duty vehicles, air is drawn in and compressed
to such high pressure and temperature that combustion starts spontaneously when fuel is injected.
These engines are more suitable for applications in the high power range, such as trucks.
Problems with NO emissions in CI engines can be solved through catalytic conversion and
x
emission after-treatment components. (Husain, 2011)
However, since all three drivetrain alternatives will have the same internal combustion engine,
and since it is a comparative LCA study, the effect on the environment for producing the engine
will not be included.
3.2 Topology of Hybrid Electric Vehicles
The definition of a hybrid electric vehicle (HEV) is that it uses two different types of sources to
deliver power to the wheels for propulsion. The most common setup is to combine an ICE with
one or more electric machines. The main reasons of using hybridized vehicles are to improve fuel
economy, reduce fuel consumption and reduce emissions (Husain, 2011).
Compared with conventional drivetrains, HEVs generally provide better fuel economy through
regenerative braking, less engine idling and more efficient engine operation. The drivability gets
better since the electric motor characteristics better match the road load. Also, emissions of
greenhouse gases are decreasing with HEVs and since the fuel economy is better the fossil fuel
consumption is reduced (Mi et al, 2011).
17 |
Chalmers University of Technology | 3.2.1.1 Modes of the Parallel Hybrid
In a parallel hybrid drivetrain, the engine and the electric machine can be used in the following
modes (Mi et al, 2011):
Electric machine-alone mode. When the battery has sufficient energy, and the vehicle
power demand is low, then the engine is turned off, and the vehicle is powered by the
battery only. The electric energy in the battery is converted to kinetic energy by the
electric machine operating as a motor with positive torque.
Combined power mode. At high power demand, both the ICE and the electric machine
supplies mechanical power. The necessary engine torque is reduced in purpose to save
fuel.
Engine-alone mode. During highway cruising and at moderately high power demands the
engine provides all the power and the electric machine is unused.
Power split mode. When the engine is on, but the vehicle power demand is low and the
battery SOC is also low, then some of the engine power is converted to electricity by the
electric motor acting as a generator to charge the battery.
Stationary charging mode. The battery is charged by running the electric motor as a
generator driven by the engine, without the vehicle being driven.
Regenerative braking mode. The electric machine is operated as a generator during
vehicle braking to convert the vehicle’s kinetic energy into electric and store it in the
battery. The electric machine here acts a part of the vehicle brake system to collect energy
during braking.
3.2.2 Important Components of the Parallel Hybrid
The components of the Scania parallel hybrid system, included in the life cycle assessment, are
explained in this sub-chapter.
19 |
Chalmers University of Technology | 3.2.2.1 Electric Machine
An electric machine is an electromechanical device used for electrical to mechanical energy
conversion and also vice versa. It is both a motor and a generator. The electric machine can
process supplied electric energy and deliver torque to the propulsion system, but it also processes
the power flow in the reverse direction during regeneration when the vehicle is braking (Husain,
2011).
Figure 3-3 An example of an electric machine (Arthur’s Engineering Clipart)
Thus, the electric machine is called motor when it converts electrical energy to mechanical, and it
is called generator when the power flow is in the opposite direction. The braking mode in electric
machines is referred to as regenerative braking. Above in Figure 3-3 an example of an electric
machine is seen.
Losses occur in electrical, mechanical and magnetic forms during the conversion process, which
affects the electric machine’s efficiency. However, the efficiency of electric machines is quite
high compared to other types of energy conversion devices.
The major advantage of an electric machine compared to an ICE is that the electric machine can
provide full torque at low speeds and the instantaneous power provided can be two or three times
the rated power of the motor. (Husain, 2011).
There are both DC and AC electric motors. The DC motors are too big and require a lot of
maintenance; therefore electric and hybrid vehicles often use AC motors, in this case a type based
on permanent magnets called permanent magnet synchronous motor (PMSM).
3.2.2.2 Inverter
An inverter is a device that converts direct current, DC, from the battery to alternating current,
AC, to drive and control the electric motor. The inverter also converts AC to DC when it takes
power from the generator to recharge the batteries. The required voltage can be constructed with
20 |
Chalmers University of Technology | the use of appropriate transformers, switching and control circuits. Below in Figure 3-4 an
example of how an inverter is shown.
Figure 3-4 One example off an inverter (Real Work Trucks)
3.2.2.3 DC/DC
A DC/DC converter, as seen in Figure 3-5, changes the system voltage from one level to another.
The input is a filtered DC voltage, although it may be unregulated. The output is a regulated DC
voltage, and multiple outputs can be designed for many applications. There are both isolated and
non-isolated converters. The 12V electronics in the electric and hybrid vehicles are supplied with
a high- to low-voltage DC/DC converter, and this needs to be of the isolated type (Husain, 2011).
Figure 3-5 An example of DCDC converter (Pues EV)
3.2.2.4 Battery
Hybrid and electric vehicles in general need some type a portable supply of electrical energy. The
electrical energy is typically obtained through conversion of chemical energy stored in devices
such as batteries and fuel cells for example (Husain, 2011).
21 |
Chalmers University of Technology | Batteries are the most convenient and developed choice of energy storage available for electric
and hybrid vehicles. High specific power, high specific energy, high charge acceptance rate for
both recharging and regenerative braking, and long calendar and cycle lives are wanted features
of batteries to use for hybrid and electric vehicles. However, although most of these requirements
can be met in some combinations technically, the cost of batteries is often still too high to be
commercially applicable. (Husain, 2011)
There are two types of batteries, primary and secondary batteries. The main difference is
rechargeability – primary batteries are not rechargeable while secondary batteries are. The
batteries used for hybrid and electric vehicles are for obvious reasons of the secondary type.
The main types of batteries used or being considered for hybrid applications are, according to
Husain (2011):
Nickel-metal-hydride (NiMH)
Lithium-ion (Li-ion)
Lithium-polymer (Li-poly)
Sodium-sulfur
The most promising battery among the four are the Li-ion, but there are several of different types
of Li-ion battery technologies being developed. An example of a Li-ion battery is seen in Figure
3-6.
Figure 3-6 A Li-ion battery (DIY Trade)
To obtain as long life time as possible the battery state of charge (SOC) is often limited for
discharging and charging, for example to 25% of the total possible capacity. This means that a
large share of the battery capacity is not used. An alternative to batteries for some hybrid truck
applications are super capacitors which do not rely on chemical reactions and therefore have a
much longer lifetime. They can be operated between a SOC of 0% and 100% without degrading
22 |
Chalmers University of Technology | the lifetime, and at the same time withstand a large number of charge/discharge cycles. (Fuhs,
2008).
3.2.2.5 Hybrid Power Unit Housing
The housing of the hybrid power unit is a box made of aluminium and steel2 that contains the
battery, inverter and the DC/DC unit, but also the cooling and warming unit. However, the two
last parts are not included in this study.
3.3 Drivetrain of a Vehicle with Electricity from the Road
Electrified heavy vehicles are not new. From 1920 to 1960 trolleybuses was quite common in
Swedish cities. However, this was before the diesel engine came into production and after its
introduction the overhead contact wires was dismantled. Nevertheless, trolleybuses are still used
in south of Europe and other parts of the world.
A lot of resent interest for electrifying heavy vehicles has focused on solutions with continuous
electricity supply during driving instead of battery-powered solutions. The reason is that in the
foreseeable future it is not considered feasible to implement battery solutions for heavy transports
driving long distances due to the limited amount of energy and power batteries can supply.
(Trafikverket, 2012)
According to Trafikverket (2012) there are today three main technologies for continuous
electricity supply from infrastructure to the vehicles during driving:
Conductive transfer through overhead catenaries,
Conductive transfer via some form of tracks or leaders in the road,
Inductive transmission via electromagnetic fields from the road structure.
The first alternative, as seen in Figure 3-7, is today the most developed technology and also the
one studied in this LCA study. However, there are some important differences between the earlier
mentioned trolley busses compared to electrified long-haulers, for example the environment in
which the transportation work is made, the average speed and what is being transported. Another
important difference is that in cities, where trolleybuses naturally are implemented, the buildings
could be used to mount the electric wires. (Trafikverket, 2012)
2 Interview with Alexei Tsychkov at Scania
23 |
Chalmers University of Technology | Figure 3-9 Inductive power transfer from the road (Scania)
The requirements on reliability to transport the iron ore from the mine to the customer are very
high. Therefore, considering the conductive tracks and the inductive technology, it is more
reasonable to implement demonstration plants before full-scale implementation is made,
motivated by not least the technology immaturity and operational uncertainty but also the
economic and time aspects for the Pajala mine project. (Trafikverket, 2012)
3.3.1 Important Components of the Catenary Hybrid
The components which will be used in the Scania catenary hybrid system, as included in the life
cycle assessment, are explained in this sub-chapter.
3.3.1.1 Electric Machine
The electric machine in the electrified truck has the same technology as in the hybrid alternative
except that it is larger, so see the previous chapter for an explanation of the technology.
3.3.1.2 Pantograph
The electricity supplied to the vehicle is collected from the overhead wires by a pantograph, as it
can be seen in Figure 3-7 above. Since the transfer is conductive, the pantograph must maintain
good contact under all running conditions of the vehicle in order to provide power. The higher
speed is, the more difficult it is to maintain good contact between the pantograph and the electric
wires.
In today’s technology the contact is maintained between the pantograph and the electric wires by
spring or air pressure. However, compressed air pressure is preferred for higher speed operations.
To maintain the pantograph in raised position, the pantograph is connected to a piston cylinder
which remains the required air pressure. (Trafikverket, 2012)
The pantograph itself consists of a contact strip or similar, which is lagging the electric wires and
transfers the electricity to the vehicle. The contact strips are supported by a mechanical structure
25 |
Chalmers University of Technology | which is attached to the vehicle and is adjustable. At least two contact lines must be used to
transmit the electrical energy, and more degrees of freedom to the pantograph is needed in order
the connection to the electrical wires in momentum is safe (Ranch, 2011).
However, it is important to point out that the installation and system integration of the pantograph
on the truck is still under development.
3.3.1.3 Infrastructure
The added infrastructure consists of a large power system consisting of an auxiliary and catenary
system. The catenary system supplies the trucks with propulsion power and the auxiliary system
supplies the signaling, telecom, switch heating, illumination and other systems. Also an
automatic transformer system is used for distributing the power in the catenary system.
As mentioned, the infrastructure includes a telecom system. The telecom system is used to
facilitate communication along the road. It consists of telecom cables, telephone systems,
transmission facilities, radio equipment, detectors, radio towers and buildings for electrical
equipment. All construction, operation and maintenance are included in the assessment of the
infrastructure.
26 |
Chalmers University of Technology | 4. Definition of Goal and Scope for this Study
In this chapter the goal and scope of this LCA study will be defined and explained.
4.1 Goal of the Study
The main purpose for carrying out this study has been to investigate the effect on the
environmental load of a possible electrification of the road bound trucks transporting iron ore
powder from the mine in Kaunisvaara, Pajala, to Svappavaara. This has been done through an
attributional comparative LCA study of three different drivetrain alternatives for heavy duty
trucks, during 16 years of mining operation – this equals the whole lifetime of the mine minus the
estimated construction of necessary infrastructure. In essence, the vehicle cycle, or cradle to
grave, impact of a set of additional components have been compared to the effect of the reduced
total energy use of fuel and electricity in the well to wheel phase when a conventional truck has
been solely hybridized or hybridized with external power supply from the road.
Three questions have been evaluated:
1. What is the difference in environmental impact of the three alternatives for transporting
iron ore powder?
2. Which added component in the drivetrain has the largest environmental impact?
3. Which phase in the life cycle has the largest environmental impact?
4. Assuming the same extraction rate per year, how long time must the mine be operated
before break-even is reached for the two alternatives, with the reference and with each
other?
This study was commissioned by Hybrid Systems Development department at Scania CV AB and
they have also been the intended recipients. Their motive for initiating this study has been to
increase their knowledge about the potential environmental impacts of electrifying the drivetrain
of heavy duty trucks.
4.2 Scope of the Study
Three different alternatives for the drivetrain have been compared. All other parts of the trucks
have been assumed to be exactly the same for all three alternatives and for this reason been left
out of the study.
4.2.1 Options
The three alternatives are:
1. A conventional truck with an internal combustion engine
2. A parallel hybrid electric version of the same truck without external charging
27 |
Chalmers University of Technology | 3. A parallel hybrid electric version of the same truck receiving electricity from an overhead
conductive catenary
Figure 4-1 The three different alternatives studied
In this report the first alternative will be referred to as the reference alternative, the second
alternative will be mentioned as the parallel hybrid alternative and the third as the catenary hybrid
alternative.
The reference alternative, as seen in
Figure 4-1, was the conventional truck which constituted the baseline for the comparison – all
components in this truck were also included the other two alternatives. They have the same
internal combustion engine and from an LCA perspective equivalent transmission systems. This
conventional truck is able to load 65 tons of iron ore powder and the total weight of the truck is
90 tons, which is also the maximum allowed weight for all three alternatives.
The second alternative was the parallel hybrid truck which has been electrified to make the
drivetrain more efficient and use less fuel, but without being externally charged. As the internal
combustion engine and the transmission has been kept the same as in the conventional truck, only
a set of additional main components have been identified by Scania for the study. The complete
cradle-to-grave life cycle of these components have been included from raw material extraction
and production, component manufacturing and assembly in the vehicle, their effect on the fuel
consumption in the use phase, to the waste handling and recycling of materials when they are
scrapped.
The following components were selected: the power electronic controller in the form of an
inverter, the electric machine, the DCDC converter, the lithium ion battery and the hybrid power
unit housing. Furthermore, as the addition of these extra components to the drivetrain increased
the unloaded vehicle weight, it also implied that the weight of the iron ore powder possible to
load had to be reduced in order to keep the maximum weight of the loaded vehicle. This lead to
28 |
Chalmers University of Technology | an increase in the number of trucks needed to transport the same amount of load, however only
five extra trucks in continuous operation over the 16 year period.
In the third alternative, the catenary hybrid truck, a larger electric machine has been installed
compared to the second alternative, as seen in
Figure 4-1. It has been assumed to be factor of three larger regarding material, size and effect
calculations compared to the electric motor in the hybrid alternative. However, for the installation
of the electric machine on the assembly of the driveline the amount of energy used is assumed to
be the same3. Also, an additional component, the pantograph, has been added to the truck and the
lithium-ion battery has been removed.
Furthermore, the infrastructure of the conductive transfer of electricity through overhead
catenaries has been included in the study. As well in the catenary alternative, as in the parallel
hybrid alternative, an increased number of five trucks were needed due to the extra weight of the
added components.
4.2.2 Driving Pattern
Only transportation of iron ore powder by truck from the mine in Kaunisvaara to Svappavaara
has been considered in the study, and the distance has been assumed to be 140 km after the new
road has been constructed. In addition, in Svappavaara the iron ore powder has been assumed to
be transshipped over to train and further transported on railway to the port of Narvik.
4.2.3 Flowcharts
In Figure 4-2, a simple flowchart is representing the life cycle phases of the parallel hybrid
alternative.
In Figure 4-3 the flowchart of the infrastructure for the catenary hybrid alternative is seen
(Nielsen, 1999) and
Figure 4-4 shows the complete flowchart for the catenary hybrid including the infrastructure.
3 Interview with Håkan Gustavsson at Scania
29 |
Chalmers University of Technology | 4.2.5 System Boundaries
As stated earlier the study covered the whole life cycle of the drivetrain of the different
alternatives from the cradle to grave. However, in the well-to-wheel phase, seen in Figure 4-2, the
data for the fuel consumption and energy use refers to the complete vehicle.
The study has focused on the technology available today and it has been assumed that no new
development or drivetrain changes were implemented during the life time of the mine. As stated
earlier the internal combustion engine has been the same in all three alternatives. Note that all
other parts of the vehicle, except the drivetrain, has been regarded as identical and were not
included in the study.
Environmental impacts from the manufacturing of capital goods, such as machines used when
producing the different components for the drivetrain, were not considered, nor the impacts from
activities of the employees.
As mentioned earlier in Chapter 3, the following components were decided together with Scania
as relevant to include in the study for the parallel hybrid alternative:
DCDC converter
Electric Machine (EM)
Li-ion Battery
Inverter
Hybrid power unit housing (HPUH)
For the catenary hybrid alternative the Li-ion battery has been taken away and these were the
components included:
DCDC converter
A larger Electric Machine (EM)
Inverter
Hybrid power unit housing (HPUH)
Pantograph
Infrastructure for electric power transfer
Two important assumptions have been made in discussion with Scania which should be pointed
out. First, no battery change was needed for the trucks during their 2 years of operation at the
Pajala mine4, and, second, the vehicle has been used up to 70 % of its distance life during this
operation5. The rest of its life time it assumed to be reused for another purpose.
4 Interview with Johan Lindström at Scania
5 Interview with Håkan Gustavsson at Scania
33 |
Chalmers University of Technology | 4.2.6 Geographical and Time Boundaries
In consultation with Scania the operation time at the Pajala mine of one vehicle was decided to be
2 years, corresponding to 500 000 km of driving distance for the two studied alternative and the
reference vehicle assuming the driving pattern described in the previous chapter. The life cycle of
the iron ore mine was set to 16 years with a starting point in 2015. This year the construction of
the infrastructure is assumed to be completely finished and from this year and onwards 4,6
million tons of iron ore powder are expected to be produced and transported annually.
Raw material production has been taken place globally or in Europe, and in the modeling of
material production, data from Ecoinvent has been used for global averages data or European
regional data depending on the availability (Ecoinvent, 2007).
Manufacturing of studied components has been assumed to take place in the countries listed
below in Table 4-1 for the parallel hybrid alternative and in Table 4-2 for the catenary hybrid
alternative.
Table 4-1 The components and manufacturing location for the parallel hybrid alternative
Component Manufacturing location
DCDC Sweden
Battery China
Inverter Germany
Electric Machine Germany
Hybrid Power Unit Housing Sweden
Table 4-2 The components and manufacturing location for the catenary hybrid alternative
Component Manufacturing location
DCDC Sweden
Inverter Germany
Electric Machine Germany
Hybrid Power Unit Housing Sweden
Pantograph Germany
Infrastructure Sweden
The assembly of the driveline has been done in Södertälje, Sweden.
The electricity used for the catenary hybrid vehicle has been modeled with the projected average
Swedish electricity mix in 2020. This was chosen as the operational phase will take place the
north of Sweden and because 2020 is the midpoint year of the assumed 15 year mine life length
in full operation.
The entire use phase was taken place in the Pajala area. The vehicle end-of-life has been assumed
to take place in Sweden.
34 |
Chalmers University of Technology | The transport road between the transshipping and the mine has been assumed to be the same in all
three alternatives.
4.2.7 Functional Unit
The functional unit of the LCA study was chosen to be transportation of one ton iron ore powder,
i.e. this was the chosen unit for the use phase, identified as the reference process in the study. In
other words, the results of the study will be presented per ton of transported iron ore powder,
which is the function that the trucks provide.
4.2.8 Limitations
Transportation of components between manufacturing sites and Scania have been excluded in the
study based on the assumption that it would have very small or negligible additional impact. Also
the transport of the trucks from Scania and up to Pajala has been excluded in the study.
Also the information about the different materials has also been simplified to keep within the
time boundaries of the project. For example only one type of steel has been used in the study to
approximate the environmental impact for different steel types.
4.2.9 Types of Impacts Being Considered
The impact categories were decided upon in cooperation with Scania. They are listed below and
explained more in Chapter 6 in this report.
Global warming potential
The emissions: NOx, HC, PM
Abiotic resource depletion potential
The global warming has been chosen since it is a current problem in the world and widely
discussed how to solve the emissions.
The NOx, HC and particle emissions are of interest since these are the emissions from the
operational phase using diesel as fuel for the reference case and the parallel hybrid case.
Abiotic resource depletion was considered as an impact category since the reference vehicle and
the parallel hybrid are consuming large volumes of fossil fuels. However, the catenary hybrid
uses no fossil fuels but far more of others materials due to the demanding infrastructure. It will
therefore be an interesting trade-off between the use of fossil fuels in the parallel hybrid and use
of material in the catenary hybrid for this impact category.
35 |
Chalmers University of Technology | 4.2.10 Allocations
Allocations in material production, production of components and end of life treatment have been
mainly based on weight, since this has been the best available information. Since this is a
comparative study attributional LCA, partitioning has been used in accordance with Baumann
and Tillman (2004).
Allocations with regards to the infrastructure have been made based on time, more specifically
the lifetime of the mine. This is due to the fact the mine in Pajala has a lifetime of 16 years and
the data used to represent the infrastructure has an assumed lifetime of 60 years (Bothniabanan,
2010).
4.2.11 Data Collection
Data for this LCA study has been collected from different sources. The data for raw material
extraction has been taken from Ecoinvent. Ecoinvent was created by the Ecoinvent Centre in
Switzerland and contains life cycle inventory data for various services and products.
To analyze the production of the components invested, data has been collected from other LCA
studies on similar products and from suppliers of the different components. Also, some
assumptions and estimations have been made by technology experts at Scania.
The data for the use phase has been collected from Scanias own Environmental Product
Declaration (Scania, 2012), from the report of Trafikverket (2012) about the mine in Pajala and in
discussion with Scania.
The data for end of life phase has been assumed in discussion with Chalmers and is also collected
through different studies on recycling of materials.
4.2.12 Critical Review
The LCA study was critically reviewed by the supervisors, Anders Nordelöf and Ann-Marie
Tillman from Chalmers University of Technology and Håkan Gustavsson from the department of
Hybrid Systems Development at Scania CV AB.
36 |
Chalmers University of Technology | 5. Inventory analysis
This chapter describes the different processes in the life cycle and the procedures of data
collection, modeling and calculation of life cycle inventory (LCI).
5.1.1 Life Cycle Phases of the Drivetrains
The life cycle of the drivetrain was divided into four different phases:
Production
Use
Maintenance and repair
End-of-life
The processes in the life cycle are taking place in various locations around the world. In all major
processes their corresponding locations are seen below in Table 5-1. These geographical
locations have been used when choosing the electricity mix for manufacturing of the different
components.
Table 5-1 Overview of the processes, their locations and time horizon
Proces Representative location Time
Raw material extraction Global average or European average is used, in 2012
special cases specific region are used.
Manufacturing of Electric Machine Sweden 2012
Manufacturing of Inverter USA 2012
Manufacturing of DCDC USA 2012
Manufacturing of HPHU Sweden 2012
Manufacturing of Battery China 2012
Manufacturing of Infrastructure Sweden 2012
Assembly of Drivetrain Sweden 2012
Use phase Sweden Midpoint
2020
End of Life Sweden 2031
To be in line with the geographical and time boundaries projected future state and country
specific electricity mixes have been used. In Table 5-3 the projected average electricity mix for
Sweden 2020 and 2031 are presented. Data for 2031 has been approximated by using data for
2030. In line with the goal and scope formulation, average data has been used for all electricity
mixes.
37 |
Chalmers University of Technology | Table 5-2 Projected average Swedish electricity grid mixed in 2020 (Gustavsson, Särnholm et al. 2011)
Primary energy Hydro Wind Solar Nuclear Electricity production Electricity production
demand, shares at district heating in industry (assumed
by fuel [%] plants oil)
Sweden 2020 38.5 11.7 2.3 40.8 3.4 3.3
Sweden 2031 40.4 18.3 4.9 30.6 3.1 2.7
The electricity mixes for 2012 for the United States and China has been assumed to correspond to
the data from 2009 (IEA, 2012) and is seen in Table 5-3 below.
Table 5-3 Electricity mix for the United States and China (IEA, 2012)
Primary energy Coal Gas Oil Hydro Wind Solar Nuclear Biofuels Waste
demand, shares by fuel
[%]
United States 45.4 22.8 1.2 7.2 1.8 0.1 19.9 1.2 0.5
China 78.8 1.4 0.4 16.7 0.7 0 1.9 0.1 0
5.2 Production
In the production phase there are raw material extraction, production of components and vehicle
assembly. The data collected for these steps are explained more in detail below.
5.2.1 Raw Material Production
The drivetrain consists of many different materials and data from the production phase for these
were gathered from Ecoinvent. The data used for extraction of raw material has been either
global averages (GLO) or European averages data (RER). This assumption has been made since
the raw material is extracted on different locations around the world and the global and European
average data was found the most representative.
5.2.1.1 Aluminium
European average data has been used for production of aluminium. Including processes has been
cast aluminium ingot production, transports of materials to the plant and the disposal of the waste
(Classen et al, 2009).
5.2.1.2 Copper
In the extraction of copper global average data is used. It includes the pre-treatment of the ore,
the reduction and the refining; the product is used as pure metal or as alloying element in various
technical applications (Classen et al, 2009).
5.2.1.3 Steel
In the production of steel, carbon steel has been assumed with a process that used average global
and European production mix (Classen et al, 2009).
38 |
Chalmers University of Technology | 5.2.1.4 Neodymium
The data of extraction of neodymium has been assumed to take place China (Classen et al, 2009).
Since this was representative due to the fact that China produces 97 % of the world’s neodymium
supply (Milmo, C. 2010).
5.2.1.5 Cast Iron
For the cast iron a composition of 35% scrap and 65% virgin iron has been used. Also, the
process represents a mix of global average and European production mix (Classen et al, 2009).
5.2.1.6 Nylon
In the production of nylon, average European production has been assumed (Hischler, 2007).
5.2.1.7 Brass
Brass was assumed to contain 70% copper and 30% zinc, and the melting and casting of brass
ingots are included. The only data available was from production in Switzerland and it was
therefore used as the best available representation of the global average (Classen et al, 2009).
5.2.1.8 Nickel
Global average data were used for the production of nickel. Included processes were mining,
necessary infrastructure and disposal of overburden and tailings. It also includes the metallurgy
step with the disposal of slag, the infrastructure and the separation of the co-product copper and
production, application and emissions of most agents and additives used in beneficiation and
metallurgy (Classen et al, 2009).
5.2.1.9 Printed Wiring Board
Global average data has been assumed in the production of printed wiring boards. The data
represent a mix of two mounting technologies, surface mount and through-hole mount. Assuming
a mix of 50:50 mix between the two technologies. It includes processes of components mounting
using lead and lead free solder technology (Hischler et al, 2007).
5.2.1.10 Tin
European data has been assumed, including the cradle to gate inventory of world-wide primary
tin production. Transports are also included from the major producers in Europe (Classen et al,
2009).
5.2.1.11 Synthetic Rubber
This type of rubber is used in technical products and was chosen since it was assumed to be the
most accurate rubber available. European data has been assumed and the included processes are
production of the rubber and also the transport of raw material to the production plant (Hischler,
2007).
5.2.1.12 Polymethyl methacrylate, PMMA
PMMA are a thermoplasts which is used in the industry as insulators and it has been assumed to
correspond to the plastic parts in the different components, since the lack of information on what
kind of plastic material which has been used. European data has been assumed and the included
processes are all processes from raw material extraction until delivery at plant (Hischler et al,
2007).
39 |
Chalmers University of Technology | 5.2.1.13 Carbon
Global average data has been used and the included processes are raw material extraction and
production of chemicals used for production of carbon. The transport of material to
manufacturing plant is included. (Hischler et al, 2007)
5.2.1.14 Graphite
This data includes the production of an anode for a lithium-ion battery where the graphite acts as
a Li+-ion accumulator. The data are based on patents and the transportation are based on
Ecoinvent standard estimates. A European dataset are used for global processes. (Hischler et al,
2007)
5.2.1.15 Lithium
Global average data were used for the production of lithium. The lithium was produced from
Lithium chloride electrolysis which delivers the co-products lithium and chloride. The allocation
in this case has been based on stoichiometric calculations according to Ecoinvent. (Classen et al,
2009)
5.2.1.16 Oxygen
European average data has been used. The technology used is cryogenic air separation, and the
products are liquid oxygen, liquid nitrogen and liquid crude argon and no gaseous products are
considered. Allocation factors were calculated from the heat of vaporization and the specific heat
capacity multiplied with the temperature difference from 20 oC to the boiling point. Included
processes are the electricity for the process, cooling water and waste heat. (Hischler et al, 2007)
5.2.1.17 Phosphorus
Phosphorus has been produced with the technology oxidation of phosphorous trichloride. Raw
materials are modeled with a stoichiometric calculation, and energy consumption and transports
are estimated. The transports have been calculated with standard values. The data used has been
based on European averages. (Hischler et al, 2007)
5.2.1.18 Tube insulation
The data for tube insulation used in the study was chosen for its representativeness of technical
applications. The included processes are raw material extraction and the production stages to a
finished product. The dataset used comes from Germany and were assumed to be representative
for the European average data. (Hischler et al, 2007)
5.2.1.19 Electrolyte
The production of the electrolyte for Li-ion batteries covers all processes from raw material
extraction to finished electrolyte. Global average data has been used and the process data.
(Hischler et al, 2007).
40 |
Chalmers University of Technology | 5.2.2 Production of Components
For this part of the life cycle data have been collected for material compositions of the different
components along with the energy, electricity and heat, used to manufacture them from these raw
materials. A waste percentage of 1% in the production for the different components has been
assumed in discussion with Scania.
The electricity mix used to calculate the emissions for the production of the different components
have been taken from the manufacturing country showed in Table 5-1.
5.2.2.1 Electric Machine
The material data for the electric machine in the parallel hybrid alternative has been collected
from Scania. It has a maximum power output of 150 kW and total weight of 83 kg.
The total weight of the permanent magnets in the electric machine was 1.9 kg and the material
composition were 0.5 kg neodymium, 0.25 kg dysprosium, 0.02 kg boron and the rest of the
material was assumed to be iron6. However, data on dysprosium and boron was not available in
Ecoinvent and was therefore assumed to have the same environmental effects as neodymium.
For the assembly energy of the electric machine for the parallel hybrid data was taken from an
EPD of an electric machine made by ABB. The selected reference motor is a flameproof 400 V
AC motor with 22 kWh rated power output and with a total weight of 279.2 kg (ABB, 2002).
After rescaling based on weight and assuming that the manufacturing energy in the EPD
approximately corresponded to the assembly energy of the Scania electric machine, the result was
calculated to be 19 kWh of electricity and 16 kWh heat.
The electric machine for the catenary hybrid does not exist today, and the material data for the
catenary hybrid has been assumed to have a power output of 250 kW in nominal rating. A
rescaling on the same electric machine from ABB, as in the parallel hybrid, was made and the
material data has been made based on the power output and with the same material configuration.
The total weight of the electric motor was calculated to be 278 kg. The assembly energy was
calculated as in the case of the parallel hybrid and was found to be 31 kWh of electricity and 26
kWh of heat.
The linear rescaling of the machines weight and material content based on the power output only
is consistent with reality if it is assumed the electric machine has been prolonged only7. However,
if the larger power output instead would reach with larger radius, the rescaling is instead the
squared value of the material to get the same power output.
5.2.2.2 Inverter
The inverter was assumed to be manufactured in USA. The material data for the inverter for the
parallel hybrid was collected through disassembly and weighting at Scania. The inverter has a
6 Interview with Jörgen Engström at Scania
7 Interview with Jörgen Engström at Scania
41 |
Chalmers University of Technology | maximum alternating current of 300 Ampere which matches the electric machine of 150 kW and
it has a total weight of 18 kg8.
The inverter for the catenary hybrid was assumed to be identical with the one in the parallel
hybrid, even with higher power output of the larger electric machine, based on information from
Scania.9
For the assembly of the inverter, an approximation from an EPD made by ABB was studied and a
converter with the most similar material content was chosen, the ACS 100/140 frequency
converter (ABB, 2002). The manufacturing energy consumption was assumed to correspond to
the assembly energy for the inverter. After rescaling according to weight the assembly energy
was calculated to be 70 kWh of electricity and 39 kWh heat.
The inverter included in the study is based on old technology and has therefore been larger than
necessary. Today smaller inverters with more up to date technology are available.10
5.2.2.3 DCDC
The DCDC converter was assumed to be manufactured in Sweden. The maximum power output
is 7.5kW and total weight is 21.5 kg. The material data has been provided by the supplier.
For the production and assembly of the DCDC converter, an EPD from ABB (2002) has been
used to approximate the used energy. The DCS 400 has been rescaled based to weight, and the
manufacture energy is assumed to correspond to the assembly energy for the DCDC converter.
After rescaling the assembly energy was calculated to be 12 kWh of electricity and no heat is
used.
The DCDC converter used in this study are over dimensioned and based on old technology, but it
was the only available data. Today there is better performing technology is available which also
will be used in the future at Scania11.
5.2.2.4 Hybrid Power Housing Unit
The HPHU has been assumed to be manufactured in Sweden with a total weight of 297 kg. The
value for energy used in the production has been estimated by Scania12. The total length of the
laser cutting has been assumed to be 58 m, and the speed of the laser cutting is around 5 m/min
since the thickness of the steel is average 3 mm thick (LTU, 2012).
5.2.2.5 Battery
The battery was assumed to be manufactured in China and the material data was collected from
the article Life-Cycle Analysis of Production and Recycling of Lithium Ion Batteries (Gaines,
8 Interview with Jörgen Engström at Scania.
9 Interview with Jörgen Engström at Scania.
10 Interview with Jörgen Engström at Scania
11 Interview with Jörgen Engström at Scania
12 Interview with Alexei Tsychkov at Scania
42 |
Chalmers University of Technology | 5.4 Well to Wheel
In the well to wheel phase the vehicle operation have been considered together with the
production of diesel and electricity to calculate the savings in fuel consumption and emissions for
the two hybrid alternatives compared to the reference vehicle. An average speed of 65 km per
hour has been assumed for all vehicles. The reference was found to have an average fuel
consumption of 0.50 liters per km based on the stated fuel consumption per hour given by Scania.
The parallel hybrid alternative is expected to save 1.50 liter diesel per hour according to Scanias
calculations, which gives an average fuel consumption of 0.48 liters per km. The catenary hybrid
values have been based on the Swedish projected average electricity mix for the 2020. The
average electricity consumption has been calculated to 2.7 kWh per km.
The extra added components in the two different alternatives will decrease the load capacity, and
in turn instead increase the number of trucks used to be able to transport the same total amount of
iron ore powder per year. In Table 5-4 the total numbers of trucks used during the lifetime of the
iron ore mine are presented.
Table 5-4 The total number of trucks used in the different alternatives
Alternative Number of trucks
Reference Vehicle 646
Parallel Hybrid 651
Catenary Hybrid 651
5.5 Maintenance and repair
It has been assumed that the difference in maintenance and repair are negligible among the
different alternatives. This was decided together with the Hybrid Systems Development
department at Scania.
No battery change has been assumed in the parallel hybrid alternative, due to the short lifetime of
2 years for the trucks and therefore a change will not be needed.13.
5.6 End of Life
After the truck’s lifetime for transporting iron ore powder from the mine in Pajala has expired, it
has been assumed that both hybrid alternatives will be used for other purposes with less heavy
operation as in the mine. In the case of the parallel hybrid this means that it may still be used as a
hybrid vehicle whereas the catenary hybrid then only can be operated in the engine-alone mode,
as a conventional mode. Therefore, an assumption was made together with Scania that the trucks
could be in operation 30 % longer than the 2 years in Pajala, and in the case of the parallel hybrid
the effect of this reuse was examined for the hybrid powertrain components.
13 Interview with Johan Lindström at Scania
44 |
Chalmers University of Technology | The end-of-life stage has been divided into a recycling phase and waste management phase.
Swedish waste incineration data has been used in the waste management phase since the use
phase has been located in Sweden. The heat produced when incinerating the materials has been
assumed to replace other heat production and was therefore credited for the corresponding global
warming emissions, since this is the major emissions when producing heat (Environmental
Energy Agency, 2011). Landfill has been assumed to be the most accurate waste management
process when recycling was not possible.
The actual recycling process for can be expected to be of an open loop type. Open loop recycling
is when a product is recycled into a different product and quality losses are quite common. As the
activities in the open loop system are shared between two products an allocation problem often
On the other hand, closed loop recycling method assumes that no significant losses in quality
occur when recycling the material. Each product is hereby responsible for its environmental load
associated with virgin material production, recycling and the final waste treatment. Therefore, an
average impact allocated equally among the products depending on the number of times recycling
occurs (Nicholson, A.L. et al. 2009).
Today virgin raw materials are extracted and stockpiled in society in the form of products. Even
in cases when recycled materials would satisfy the quality requirements, virgin material is often
used to be able to produce products with high quality and because the infrastructure for extraction
makes it easily available. Nevertheless, the assumption in this study has been that all recycled
material comes into use in another new product as in the open-loop setup but at the same time it
is still replacing virgin material extraction by the recycled amount, as if modeled by a closed
loop. The idea is that there is a large share of the societal raw material demand which use virgin
resources but which could be met using a mix of virgin and recycled material with lower but
acceptable quality. Different materials have then also been assigned different recycling rates due
to dispersion in various processes and in some cases too large quality losses. These assumptions
are explained further down. Accordingly, with this calculation method the recycling process
resembles more a closed loop recycling method than the supposed open loop, and the positive
effects of the recycling process has been credited the different studied alternatives.
Electronics have been assumed to be dismounted and disassembled by hand to separate larger
metal and plastic casings. Small metal fractions and alloys have been assumed to be fragmented
and not recycled, and then placed in a landfill. The plastic materials have been assumed to be
incinerated with energy recovery. Today there is no effective process in recycle the printed
wiring boards, so they have been assumed not to be recycled and also go to landfill.
The larger metal scrap has been assumed to be recycled according to the following rates:
aluminium and copper with 100 % (Leifsson, 2009) and steel and iron with 70 % due to waste
and quality losses. The remaining 30 % has been assumed to go to landfill.
Lithium from the battery has not been recycled at all. Since the battery recycling market is largely
price driven and with current technology recycled lithium costs as much as five times more than
newly produced lithium. Though, lithium could be recycled to 100% and it is expected to be the
future main source of lithium supply (Waste management world, 2011).
45 |
Chalmers University of Technology | 6.2.1 Results for Global Warming Potential
The first graph, Figure 6-1, illustrates how big the global warming emissions are in the well to
wheel cycle for the different studied alternatives. As seen a large saving has been done in the
catenary hybrid compared to the reference vehicle, but also compared to the parallel hybrid. This
is due to the change of the energy source from diesel to Swedish projected average electricity mix
in 2020.
Figure 6-1 Global Warming Potential of the Well to Wheel cycle for the different alternatives
in turns of g CO-eq./ton iron ore
2
The second graph,
Figure 6-2, shows the global warming emissions for the entire life cycle of the drivetrain, the
cradle to grave cycle. The values in this graph are delta values compared to the reference vehicle,
which is always at the zero line. If the emissions have positive values, the studied alternatives
emit more than the reference vehicle and when the alternatives have negative values the
alternatives provide savings in emissions. Thus, a negative total bar indicates that the alternative
is a better solution than the reference vehicle.
As seen in Figure 6-2, the savings in the well to wheel phase for the catenary hybrid is very large
compared to the saving in the parallel hybrid. However, both alternatives are favorable solutions
compared to the reference vehicle. The catenary hybrid has larger global warming emissions
compared to the reference vehicle in the extraction of raw material phase. However, these added
emissions are clearly counterbalanced and outweighed by the saving in the well to wheel phase.
47 |
Chalmers University of Technology | 6.2.4 Summary of the Results
The results show that both the studied alternatives are favorable solutions compared to the
reference vehicle. The life cycle phase that has the largest environmental impact is without doubt
the well to wheel phase, i.e. when the vehicles are in operation.
The alternative powertrains are a lot better than the reference in almost all studied impact
categories, except for the emissions of particles. In this category the emissions are slightly worse.
Comparing the alternatives with each other, the catenary hybrid is the more favorable choice for
transporting iron ore powder in Pajala.
6.3 Results Divided Into Life Cycle Stages
In order to investigate which of the added components that has the largest environmental impact
pie charts have been produced showing each component’s share of contribution to the different
impact categories and emissions. The included life cycle phases were raw material extraction,
production of component and end of life. The results for the parallel hybrid are shown in Figure
6-9.
Figure 6-9 Relationships between the components in the Parallel Hybrid alternative
As seen in the Figure 6-9 above, the battery has the largest environmental impact in both impact
categories, but also in the emissions of nitrogen oxides and particles. An interesting result is that
the Hybrid power unit housing had such big percentage of the hydrocarbon and particle
emissions. This is due to the large amount of material used in the HPHU.
52 |
Chalmers University of Technology | In the catenary hybrid the constituent with the largest environmental impact is more obvious. As
seen in Figure 6-10, is the environmental impact of the infrastructure an order of magnitude
larger than the components added in the vehicle. The blue part of the first bar in the parallel
hybrid stack is the environmental impact from the battery and the orange part represents the
remaining added components. The green bar is the global warming potential of the infrastructure
and the purple bar is the components in the catenary hybrid without the infrastructure.
Figure 6-10 Environmental impact of the parallel hybrid components, the infrastructure and the catenary hybrid
components excl. the infrastructure in turns of gCO -eq/ton iron ore
2
6.4 Break-even Analysis
In the reference case the trucks have been leaving the mine fully loaded every 7:th minute.
Assuming the same extraction rate of 4.6 million tons iron ore powder per year, it has been
examined how long time the mine must be operated before break-even is reached for the two
alternatives, with the reference and with each other for the two impact categories, global warming
potential and abiotic depletion potential.
As seen in Figure 6-11 the break-even point for the catenary alternative is at roughly four months
with both the hybrid and the reference for the global warming emissions. In order words, if the
mine is going to be used for more than four months the catenary hybrid becomes the most
favorable choice to transport the iron ore powder. The values are still delta values compared to
the reference vehicle, meaning that negative values are savings in environmental impact.
53 |
Chalmers University of Technology | 7. Uncertainty and Sensitivity Analysis
An investigation of the robustness of the final results was conducted for some identified key
assumptions and limitations of the project. The result of this investigation is presented here.
7.1 Mark-up for the Pantograph
The pantograph was never included in the LCA calculations for the catenary hybrid alternative as
explained above. Together with Scania it was decided that a mark-up calculation was the most
accurate method to evaluate the uncertainty caused by excluding the pantograph from the study.
Three different mark-up calculations have been conducted: 25 %, 50 % and 100% of the raw
material extraction and the production of components phases of the catenary hybrid, excluding
the infrastructure, was used.
The results for the mark-ups for the pantograph for the Global Warming Potential are presented
below in Figure 7-1. As seen, even if the pantograph is included in the drivetrain it does not
contribute to any major difference in the final results.
Figure 7-1 Mark-up for the Pantograph in Global Warming Potential in turns of gCO -eq/ton iron ore
2
7.2 Changes in Energy Consumption
Since the well to wheel phase of the catenary hybrid has undoubtedly the largest environmental
impact, changes in fuel consumption data used was made to see if it made any changes of the
results.
55 |
Chalmers University of Technology | A 10 % change in fuel consumption of diesel was tested, both positive and negative changes.
Also, a 10 % variation in electricity consumption of the catenary hybrid was made. In both cases
the changes in the final results were very small. Below, in Figure 7-2 the results from a 10%
lower fuel consumption of diesel combined with a 10 % higher electricity consumption was
assumed and are presented as the “worst case”. To clarify, no other life cycle phase was changed.
Figure 7-2 Changes in the Energy Consumption was made with a 10% lower fuel consumption of diesel and a 10 % higher
electricity consumption in turns of gCO -eq/ton iron ore
2
7.3 Importance of Recycling and Reuse
The recycling rates for aluminium, steel, iron and copper are set to be quite high in the study and
it helps to reduce the environmental impact of both the studied alternatives. It has also been
assumed that the vehicles are reused after the mine operation. A worst case scenario, for the end
of life phase, would be if no reuse or recycling was made at all and all the material was assumed
to go to landfill.
First the importance of reuse is presented in Figure 7-3 below. Here the parallel hybrid trucks
have no reuse phase after the operation in the iron ore mine in Pajala.
56 |
Chalmers University of Technology | 8. Interpretation
In this chapter of the report the results are discussed and final conclusions are drawn. Also some
thoughts and recommendations about the made LCA are presented.
8.1 Discussion
Changing the energy supply from diesel to electricity is the most important explanation for
overall final results. As a consequence, the methodological choice of the electricity mix plays an
important role, but the sensitivity analysis shows that a change of the electricity mix will not alter
which of the alternatives that becomes the most favorable final solution. This is an interesting
conclusion, since a lot of debate today concerns that the European electricity mix possibly will
make the catenary hybrid solution an unfavorable solution compared to reference truck.
This life cycle assessment has, in some cases, used rough approximations and simplified, but
reasonable, assumptions and cannot therefore provide values with exact precision for the final
results. This is partly a consequence of the selected calculation method (spreadsheet software) as
well as the data availability for collection. However, as indicated both by the results themselves
and the uncertainty and sensitivity analysis, the results are both clear and very robust.
Studying the life cycle of the different drivetrain, for all impact categories, production of
components and drivetrain assembly has been shown to have a minor impact compared to other
phases. On the contrary, the well to wheel phase has without doubt the largest effect on the
savings for both the alternatives, but mostly for the catenary hybrid. The construction of
infrastructure has also quite large effect, but with recycling rates that are quite high, the
contribution to the environmental impact in the end turned out to be quite small.
Both alternatives are a better solution than the reference vehicle in all impact categories except
the emissions of particles. This might be due to the poor level of detail when collecting data to
the use phase from Scania’s EPD. Therefore, this rather small increase of emission must be
regarded as uncertain.
The component with the largest contribution to environmental impact for the parallel hybrid is
clearly the battery when studying the raw material extraction, production of components and end
of life. And this is the case although that only part of the energy in the production of the battery
has been included due to lack of data. However, this data lack can be assumed to be small and
have no effect on the overall results of the parallel hybrid, with an uncertainty similar to the
mark-up of the pantograph section 7.1, at the same time as would only strengthen the role of the
battery in the parallel hybrid drivetrain. The main reason for the large impact of the battery is use
of advanced materials in quite large quantities, such as in the active electrode materials, which
are both energy and resource intensive when produced.
Another interesting result was that the hybrid power unit housing had such a large percentage of
the emissions of particles and hydrocarbons.
Another comment about the environmental impact of the different components is that the
difference between the inverter and DCDC is relatively large. They are consisting of
59 |
Chalmers University of Technology | approximately the same materials and ought to be more similar. However, the collection of data
for the inverter was conducted through disassembly and weighting of the different materials, i.e.
with less precision than other parts which may be an explanation.
For the catenary hybrid the infrastructure has, as mentioned earlier, the largest contribution. Also,
if it is compared with the most contributing part of the parallel hybrid, the battery, then it
becomes obvious that there is an order of magnitude difference. However, the savings are then
also much larger in the WTW phase for the catenary hybrid.
When assuming no recycling and reuse at all the catenary hybrid still remains the better option,
despite the large use of materials in the infrastructure. It was expected that recycling would have
a greater impact than it showed – the explanation is again that the savings in the WTW phase are
a lot bigger than the savings in the end of life phase. However, with another, more fossil intensive
electricity mix the end of life phase would have a larger influence on the final results.
As seen in the mark-up for the pantograph of the catenary hybrid alternative, it does not
contribute to any larger difference in the final results. Therefore, the presented results could be
rather representative even without the pantograph.
A final observation is the result for the break-even point. If the iron ore mine only has a lifetime
longer than five months the catenary hybrid is the more favorable solution in an environmental
perspective.
60 |
Chalmers University of Technology | 8.2 Conclusions
Based on the results and discussion the answers to the questions asked in Goal and Scope are the
following:
1. What is the difference in environmental impact of the three alternatives for transporting
iron ore powder?
The Catenary Hybrid is the more favorable choice for both impact categories and for the
selected emissions. This is because the use electricity instead of diesel provides enormous
savings in environmental impact.
2. Which added component in the drivetrain has the largest environmental impact?
For the parallel hybrid it is clearly the Li-ion battery that gives the largest additional
environmental load. This is due to the amount of advanced materials included in the
battery. For the catenary hybrid it is the infrastructure which has the largest environmental
impact. This is due to the large amount of material that is used. In both cases the
environmental impact comes derives from the raw material extraction.
3. Which phase in the life cycle has the largest environmental impact?
For all studied alternatives, including the reference, it is the use phase and the WTW cycle
that that is most important in terms of environmental impact. Both hybrid vehicles have
their largest contribution in this phase and it the amount of savings made here that
determine the overall final results.
4. Assuming the same extraction rate per year, how long time must the mine be operated
before break-even is reached for the two alternatives, with the reference and with each
other?
For the global warming potential the break-even point was at four months and for the
abiotic depletion the break-even point was at four and a half month. This means that if the
mine is in operation longer than four and a half month the catenary hybrid is the better
choice.
61 |
Chalmers University of Technology | 8.3 Recommendations
Based on this LCA study the following recommendations regarding choice of drivetrain for
transportation of iron ore powder from the mine in Pajala, and future LCA work are given.
The catenary hybrid shows larger environmental savings than the parallel hybrid. Therefore the
catenary hybrid drivetrain, including infrastructure, is the recommended drivetrain solution.
For future LCA studies it is also recommended to collect material and process data from the
suppliers continuously to an internal LCI database. However, in the case when no data is
available, literature data or assumptions regarding processes can be used to estimate energy
consumption, material transformation, etc. from similar studies.
8.4 Future work
Some recommendations for future work to complement this study:
The role of other electricity mixes could be more explored.
Explore the role of new or different materials in the components.
Investigate if downsizing of the ICE is possible and if it could lead to more load capacity.
Investigate the accuracy for using Bothniabanan as an approximation to the Pajala road
infrastructure.
Continuous data collection for future LCA studies.
Look at other technologies of continuous electricity supply.
Investigate the particle emissions more closely.
Explore how the results change if other fuels than diesel are used.
62 |
Chalmers University of Technology | It can be estimated that around 10 % of the stated total amount of vehicles do not complete the
assembly line as customer vehicles.15 This value has been subtracted r in the calculation of the
following table:
Alternative Percentages of Consumption of
electricity / heat electricity / heat [kWh]
Reference vehicle 3 % 19,5 / 18,3
Parallel Hybrid 4 % 26,0 / 24,4
Catenary Hybird 5 % 32,5 / 30,5
A.5. Results from the LCIA
The table below shows the exact delta values of the GWP100.
Table A-1: The exact values of the GWP100 in g CO2-eq/f.u.
Alternative/ Raw Manufacturing Construction WtW End of Life TOTAL
Phase material of of
extraction Components Infrastructure
and Drivetrain
Assembly
Parallel 28 8 x -254 -12 -230
Hybrid
Catenary 580 0, 15 28 -6681 -438 -6510
Hybrid
In Table A-2, the real delta values of both the alternatives and the different life cycle phases are
presented for the abiotic resource depletion potential.
Table A-2: The exact values of the Abiotic resource depletion in g Sb-eq/f.u.
Alternative/ Raw Manufacturing Construction WtW End of Life TOTAL
Phase material of of
extraction Components Infrastructure
and Drivetrain
Assembly
Parallel 0,1 0,05 x -2 -0,08 -1,5
Hybrid
Catenary 4 0,001 0,2 -42 -3 -40
Hybrid
15 Interview with Håkan Gustavsson at Scania.
A-4 |
Chalmers University of Technology | Abstract
Keywords: auditory advisory, traffic information, behavior modeling, process min-
ing, data mining.
Driver safety continues to be hugely important to car manufacturers, governments
and drivers. Statistics from the CARE European Road Accident Database (2012)
reveal that there were 1,190,448 accidents and 34,817 fatalities in Europe during
2009.
Recently, researchers have pointed out that Advanced Driver Assistance Systems
(ADAS) should focus more on design for situational awareness to provide the driver
with attention supports. The studies carried out by M. Wang have proved that pro-
viding continuous visual traffic information increased drivers’ safety during highway
scenarios. In addition, drivers perceived the information as non-obtrusive. Studies
found that early warnings or normal driving information presentation reduced the
number and severity of crashes. Recent studies on auditory modality in vehicle use
has also shown great potential that auditory information may be more effective than
visual modality. Results of studies have proven that auditory information improve
safety in driving, shorten response time, enhanced accuracy and increase drivers’
situation awareness.
Toverifytheeffectsoftheauditoryadvisorysystem, datawithrespecttothedrivers’
behavior collected from experiments need to be analyzed and evaluated. Process
mining, i.e., extracting valuable, process-related information from the event logs,
complements existing approaches to Business Process Management (BPM). BPM
primarily focuses on analysis of process management and the organization of the
workfromprocessautomationandprocessanalysis, andaimstoimproveoperational
business processes, possibly without the use of new technologies. On the other hand,
BPM is often associated with software to manage, control and support operational
processes. As BPM heavily relies on process models, process mining plays a very
important role in raw data analysing. Process mining which focuses on processes
but uses the real data is bridging the gap between classical process model analysis
and date oriented analyses like data mining and machine learning.
In this case, related to the driving behaviour, data like steering wheel angles, ve-
locities, accelerations, reaction times, longitudinal/lateral position and etc. will
contribute to build model to illustrate how drivers will behave with and without the
auditory advisory system in different scenarios.
The objective of this thesis work is to develop and verify a 3D auditory advisory
traffic information system (3DATIS) based on design requirements generated from
previous studies and investigate the conceptual design for its safety value and pos-
sible positive/negative adaptive behavior of the drivers to 3D sound information
presentation in a car simulator.
This report presents how we design the 3DATIS system and illustrates how it influ-
ences drivers’ behavior.
vi |
Chalmers University of Technology | 1
Introduction
In this chapter, the background of the thesis goal and the related works will be
introduced.
1.1 Background
Road transportation by use of vehicles enables nations as well as individuals to reap
the benefits of the movement of goods and people. The benefits could for example
be the access to better jobs, markets, health care and education. However, with an
increasing number of automobiles on the road, the traffic situation is becoming more
and more challenging for the drivers. According to the World Health Organization
(WHO), road traffic accidents caused approximately 1.25 million worldwide death
in 2015[1]. That is to say, in every minute, 3 people were killed in traffic accidents.
Predicted by WHO, in 2030, the road traffic injures will go up to rank 5 in the
list of leading causes of death from 2.2% to 3.6% which will exceed lung cancers,
tuberculosis, HIV, and diarrhoeal diseases[2].
Figure 1.1: Worldwide number of road traffic deaths
Asillustratedinfigure??, theworldwidenumberofroadtrafficdeathswasincreased
in the past decade. However, during the same period the world population increased
by 5% [1]. Furthermore, from 2010 to 2013 the number of registered automobiles
increased by 16% worldwide[1]. This demonstrates that in the past years, the road
safety efforts have saved lives. Hence, driver safety continues to be hugely important
to car manufacturers, governments and drivers.
Active and passive safety systems are two primary safety systems which are used in
vehicle industry nowadays for reducing the effects of collisions. For instance, seat
1 |
Chalmers University of Technology | 1. Introduction
belts and airbags are two well known examples of passive safety systems which pro-
tect passengers to avoid injures after the collision happens. Active safety systems
on the other hand are used help drivers to understand the state of the automobile
to both avoid and minimise the effects of a crash.
Active systems today fall under the general term of Advanced Driver Assistance
Systems (ADAS). Increasing demand for ADAS is seeing new generations of cars
equipped with numerous sensor technologies powering the aforementioned systems.
In the past decades, the early ADAS designs only focused on the warning zone which
means the warning voice would give to the drivers less than 2.5 seconds. Such a
time will give the drivers shortest time to react to the emergency events in order to
avoid the collision happen, according to T.J. Triggs et. al [3].
However, unnecessary information presented to the driver can lead to high visual at-
tention and mental workload. Furthermore, a lot of the traffic information presented
from ADAS are warning signals that are usually activated in potentially dangerous
near-crash and pre-crash situations. The playing of a warning voice only make sense
when a crash is going to happen immediately. Also, the frequency of warning should
be rare as well. Otherwise, it could cause ‘cry wolf’ effect which makes the drivers
neglectthewarningfortherealpotentialriskandfailtofollowthedesignedreaction.
Hence, in the next step, the ADAS system shifts to a ‘higher’ level which pro-
vides advisory information rather than warnings. Those particular forms of ADAS
are called Advisory Traffic Information Systems (ATIS), and they support decision
makingonalongertimescale, i.e. onthetacticalandstrategicallylevels, asopposed
to on the operational level.
1.2 Advisory Traffic Information System (ATIS)
According to Summala[5], the aim of drivers is to drive in their comfort zone. While
the border of the comfort zone is surpassed and thus the safety margin violated,
the drivers will feel uncomfortable and try to adapt their behaviors to corrective
actions. To maintain comfortable driving, drivers require more information from
the surrounding traffic environment which makes it possible for them to handle the
potential risks. How can the surrounding information be sent to the drivers effec-
tively to support them in some ways that they feel comfortable?
The studies carried out by[8] have shown that providing continuous visual traffic
information increased drivers’ safety during highway scenarios. In addition, drivers
perceived the information as non-obtrusive. Studies found that early warnings or
normal driving information presentation reduced the number and severity of crashes
[6].
M.Wang et al. [6] pointed out that the visual traffic advisory system can help
make the behaviors of drivers ‘safer’ in comparison to a reference group in some
specific traffic scenarios. The number of collisions can also be reduced significantly.
2 |
Chalmers University of Technology | 1. Introduction
1.3 Hypothesis
Recent studies on auditory modality in vehicle applications has shown great po-
tential and illustrated that auditory information may be more effective than visual
modality. Additionally, the results of other studies have shown that auditory in-
formation improves safety in driving, shortens response time, increased accuracy
and improve the situational awareness of drivers[3][4][8]. Hence, in this thesis we
would like to find a way how to express and research the auditory information in
order to improve the behavior of the driver of the vehicle with respect to safety and
reliability.
1.4 The main research questions
The objective of this thesis work is to design and verify the 3D auditory advisory
traffic information system (3DATIS).The performance of the system will be inves-
tigated based on the safety value and the possible positive or negative adaptive
behaviors of the drivers when subjected to the 3D sound information. The driving
tests will be carried out in a car simulator.
The master thesis shall answer the following questions:
• How to design the 3DATIS?
1. How to play the voice?
2. What kind of melody should be used?
• How do 3DATIS influence driver’s behaviors?
1. What kind of data need to be collected?
2. How does the data have to be processed to be able to build from it a model of
the driving process?
3. How then to from that data build process models of different scenarios?
• In which stages will 3DATIS have a noticeable effect, either positive or nega-
tive?
1. What kind effect took place (Was the reaction faster? Did the car deviate from
the expected path?)
2. Why did the effect happen?
3. What is the best that can happen?
1.5 Process mining
To verify the effects of the auditory advisory system, data representing the behavior
of the drivers during the experiments need to be analyzed and evaluated. Process
mining, i.e extracting valuable, process-related information from the event logs,
complements existing approaches to Business Process Management (BPM)[8]. BPM
3 |
Chalmers University of Technology | 2
Methods
2.1 Overview
The whole process of the thesis work is illustrated in the Figure 2.1
Figure 2.1: The process of the thesis work
During the experimental design all the details of the experiments and the processes
need to be verified. Additionally, all desired data, i.e. the data useful for verifying
the performance of the 3DATIS, need to be specified. What type of information,
how to display the information and how to eliminate the bias also need to be taken
into consideration. In terms of data analysis, methods need to be found in order to
process the data and conclude results relating to the main research questions defined
in the Section 1.4.
2.2 Experiment design
2.2.1 Time to Collision (TTC)
In ADAS, time to collision is a very essential element in vehicles’ human machine
interface (HMI) design. Time to Collision (TTC) has been verified to be a very
effective way to eliminate the severity of the vehicle collision and recognize critical
or normal behavior. As illustrated [9] in Figure 2.2.
5 |
Chalmers University of Technology | 2. Methods
Level TTC Distance to Collision
Informative(white) >6s & <9s <4.5m
Advisory(orange) 3s to 6s <3.5m
Critical(red) <3s <2.5m
Table 2.1: The explanation of M.Wang’s ATIS interface design
rear right, and one to the back. The white color signifies that the TTC for all the
road users is between 6 and 9 seconds. To the right in Figure 2.3 is shown the user
interface when there are two pedestrians, in close vicinity to the ego vehicle, one in
front, and one to the rear right in the blind spot. In addition, a road user is very
close to the ego vehicle on the immediate right, the red color signifying a TTC of
less than 3 seconds.
2.2.2 Design principle
Unlike the visual advisory system, the drivers who use auditory advisory system
will get the information passively, rather than observing the interface at regular
intervals. That is to say, the drivers themselves cannot control the flow of the in-
formation. This behavior of the warning system might leave drivers irritated and
annoyed.
Hence, based on the pilot experiments, the informative information feels a bit re-
dundant. Figure 2.4 illustrates how the system will play the alarm to the drivers
with respect to TTC
Figure 2.4: The design of the 3DATIS system
As can be seen in Figure 2.4, the threshold of the advisory is shortened between 2
seconds to 4 seconds, according to the decision from the scenario that held before
the experiment started, to reduce the mandatory input to the drivers which will
cause a "annoy feeling". On the other hand, the warning zone of TTC is set less
7 |
Chalmers University of Technology | 2. Methods
than 2 second. That is to say, when the event happened, the system will play two
different tones with respect to the two situations.
2.2.3 Design of Advisory Traffic Information System (3DATIS)
The simulator room is in the SAFER, Vehicle and Traffic Safety Centre at Chalmers
which is located in Lindholmen Science Park. The hardware equipment is shown in
Figure 2.5
Figure 2.5: The hardware setting
In this experiment, one PC running the STISIM Drive® software which which ex-
ports the real-time geographic (coordinates of road users and etc.) and physical
(speed, steering wheel angle, acceleration and etc.) data was used driver simulator.
A Logitech G25 Racing Wheel, including pedals and a gearbox was installed on the
frame. Another PC running MATLAB was used to recieve and process the data ob-
tained from the simulation environment. A 5.1-channel surround-sound system was
used to emit the auditory information in the 3DATIS prototype. The arrangement
of the loudspeakers and the seating positions of the participants were calibrated
according to Dolby 5.1 home theatre speaker guidelines. The speaker system used
was a Logitech model Z-5500. Sound-absorbing curtains were installed on three
sides of the test area to ensure a good surround effect. A HD projector was used to
project the simulated drive scenarios on the front wall. Two webcams were installed
to record what the drivers saw on the road, as well as and their reactions to the
incidents, e.g. steering, braking. This video data was synchronized with simulation
data to better allocate the starting point of the driver reactions with respect to any
incident reactions to the incidents.
STISIM
8 |
Chalmers University of Technology | 2. Methods
STISIM Drive® is a programmable driving simulator developed by STISIM Drive
company in the United States. As the experiment required the driver to operate
the vehicle in different traffic situations, an open, programmable, and expandable
virtual reality driving simulation software engine is essential. STISIM Drive® pro-
vides such a platform and enables researchers to edit the traffic scenarios to fit the
requirements.
Pure Data
Pure Data (Pd) is a visual programming language developed by Miller Puckette in
the 1990s for creating interactive computer music and multimedia works[7]. In this
experiment, the program will play the corresponding pitch and tone. Figure 2.6
illustrates a user interface of PureData. The activation of the different information
levels is based on two parameters: safety margin (SM), and TTC.
Figure 2.6: The user interface of Pure Data
In this study, the prototype of 3DATIS have been developed with the goal to provide
auditoryadvisoryregardingthesurroundingenvironment. Theinformationcontains
directional risk levels in relation to the participant’s own vehicle. The objective is to
develop a system capable of letting the driver know where any surrounding object is,
and how urgent corrective action is. From the experience during building the system
in MATLAB, the expected TTC threshold was too small to handle the situations.
Hence, in order to give the participants enough time to react to an emergency situ-
ation, the threshold of the TTC was raised to 3 to 6 seconds, and 0 to 3 seconds for
the advisory and critical zone respectively. Table 2.1 illustrates the said properties.
9 |
Chalmers University of Technology | 2. Methods
Information Sound effects Time to Colli- Distance to
level sion Collision
Advisory Original sound 3s to 6s <3m
sample
Critical Increased pitch <3s <2m
and frequency of
looping
Table 2.2: Thresholds for warnings in terms of time to collision (seconds) or dis-
tance to collision (meters).
As can be seen in the Table 2.2, distance to collision as another criteria is added.
That because if two cars were driving in parallel, the relative speed will be zero and
thus TTC can be infinite. In this situation, TTC could not capture the decrease of
the distance in lateral position which still might cause the collisions.
The sound effect is another essential part in the 3DATIS design. Hence, choosing
a satisfying and a high reliability sound is a critical work. From the beginning, the
Swedish automotive OEM’s sound sample databases provided several sound samples
which had been tested with their global potential customers who gave high accep-
tances. However, those sounds were initially designed without critical level which
is to say, they did not have multi-frequency sound. Hence, a professional acoustic
designer from the automotive OEM, designed a hitting-bamboo-like sound which
consist of a sharp transient and a short tail. The short transient has many over-
tones which contributes to it make a multi-frequency sound. The acoustic tail is to
decrease the annoyance and give the comfort into the signal similar to the plucking
of a string. Hence, this sound has two characteristics which help the drivers in some
particular situations: pleasantness and directivity. This sound sample had been
tested by several sound experts and research groups to guarantee it was appropriate
for the excepted advisory use instead of just alerting.
In this experiment, the 3DATIS prototype was developed in a combination of Pure
Data and MATLAB. The Pure Data patch was designed to express the sound effects
which related to the auditory information levels as shown in the Table 2.2.
2.2.4 The traffic incident scenarios design
Toexpresstherealtrafficsituationinagoodwayinthesimulator, theresearchteam
of M.Wang had observed over 100 naturalistic driving videos. Thus, the scenarios
used in this experiment covered common critical situations from those videos.
10 |
Chalmers University of Technology | 2. Methods
Cutin
Figure 2.7(b) shows car b initially parked at the road shoulder suddenly start mov-
ing and cut out to participant’s lane and then cut back to the road shoulder. As car
c limits participant’s driveable region, changing lane for the participants may not
be a good choice. The considered safe reaction is to brake or steer slightly to the left.
Pedestrian
In this event, the road is narrowed to include only two lanes. A pedestrian will
suddenly walk out from the front of a car parking on the right road shoulder. The
pedestrian will not show up until the participants drive really close to the pedes-
trian, thus, the reaction time will be very short in this scenario. The considered safe
reaction in this scenario is to turn the steering wheel quickly as fast as possible.
Intersection
When participants restart the car after stopping before the traffic light, two pedes-
trians standing oppositely will start to cross the street (they will run the red light)
from one side to the other. Furthermore, the pedestrian on the right hand side will
be hidden behind the A pillar of the vehicle. In this scenario, the considered safe
approach is to brake and wait for the pedestrians to pass by.
Overtaking The car driving in front of the participant will be driving with signifi-
cantly lower speed, forcing an overtake to occur. While the driver has the attention
on the front vehicle, another vehicle will suddenly appear from behind in a very
high speed and overtake the participant with a relatively close distance as shown in
Figure 2.7(e). The considered safe approach is to turn back to the right lane after
overtaking the car b as quick as possible.
2.3 The overall procedure
Before the formal test, the participants need to fill in personal information like age,
gender, occupation, the time having driving licence etc. Then, the simulator will be
introduced to participants including the steering wheel, the transmission mode, the
screen and the pedals. After that, the description of the goal and procedure will be
given to the participants, as well as how the 3DATIS works, and the participants
are suggested to keep the velocity at 60 kilometers per hour, follow the traffic signs,
try to keep themselves in the right lane if there is no need to change the lane and so
on. Then, the participants will have a training section on the simulator to be able
to adapt the simulator like feeling the feedback of the steering wheel, getting used
to the positions of the pedals, and more importantly, to make sure the participants
have a correct understanding of the 3DATIS system. When the participants feel
properly prepared, the experiment is taken to the next step. In the formal testing,
the participants tested twice: once with the advisory from 3DATIS system and
once without. To eliminate the bias, both the order of the two tests and all the
scenarios which were introduced in Section 2.2.3 are randomized. Furthermore,
each test would freeze twice with stopping the program and black screen. When the
systemwasfrozen,theparticipantswererequiredtofillouttwosituationalawareness
12 |
Chalmers University of Technology | 2. Methods
questionnaires based on situation awareness global assessment techniques (SAGAT)
and situation awareness rating techniques[18]. Also, the freeze is randomized.
2.4 Participants’ performance measurement
To research how the participants will react to the emergence, it is needed to extract
the data from stage to stage from the raw data set. As the researchers would like to
know how the 3DATIS influences the drivers’ action, there are five important time
points needed to be found in the data set, as shown in Figure 2.8 (the reaction point
and emergency event start can change the order as their position depends on their
time), so that driving behaviors hided in two groups of data for 30 participants (one
with advisory from the 3DATIS and one without) can be researched comparatively.
More specifically, for each participant, the behaviors data from advisory voice on to
reaction point and reaction point to over take the objective car need to be extracted
from the raw data.
Figure 2.8: Timeline of key points.
Figure 2.9 shows an example of the data set collected from one participant.
Figure 2.9: An example of raw data.
In Figure 2.9, each column has it own meaning. For example, column A holds the
absolute time starting from switching on the simulator to the end and the samples
taken approximately eight times per seconds. Some important variables that will
13 |
Chalmers University of Technology | 2. Methods
be used in the data analysis, are shown in Table 2.3: The Column V holds the
Column A Relative time
Column C Lateral acceleration
Column E Velocity
Column F Absolute distance
Column H Steering angle
Column I Gas pedal
Column J Deceleration
Column K Times of collisions
Column O Advisory indicator (0=off; 1=on)
Column V Longitudinal distance to the road user
Column W Lateral position to the road user
Column X Forward speed of the road user
Table 2.3: Thresholds for warnings in Terms of time to collision (seconds) or
distance to collision (meters).
value 999 and Column X keeps an certain initial value until the road user appear
in the drivers’ sight and that point can be the an index of the point the objective
car show up, and Column W holds the a certain initial value until the road user
starts behaving which forces the participants to have to do something to avoid the
collisions. And the changing point of Column W is also an index that we divide the
process into two basic periods, before the emergency and after the emergency.
Findingthereactionpointisthemostessentialandcomplexpartinfindingthesefive
points. The drivers’ behaviors are essentially given by changes in three quantities:
Steering wheel angle, acceleration (gas pedal) and deceleration (brake pedal). For
example, participant 2’s Cutin scenario behavior is illustrated in Figure 2.9. The red
curve is the acceleration (increase means hit the gas pedal). The blue curve shows
the steering wheel angle (decrease means turning left and increase means turning
right). The yellow curve is the brake pedal (negative value when braking). Finally,
the purple line as the lateral distance of the car parked in the road shoulder, which
starts to move at the time indicated by the green vertical line. We define the valley
value of each curve is the action-ending point then move backward to find the peak
value as the action-start point. After that, by comparing the three action-start
points, we can find the earliest one to be the reaction point. As can be seen in
Figure 2.9, since there is no braking (Column J holds only 0’s), only the steering
wheel and gas pedal need to be compared. Obviously, the action of releasing gas
pedal is earlier than the turning the steering wheel. Hence, the reaction point of
this scenario is on the point where the red vertical line is, in Figure 2.10.
14 |
Chalmers University of Technology | 2. Methods
Figure 2.10: An example of the behaviors curves from the collected data.
The precise measurements for each stage of each scenario are as follows:
1.Number of collisions: A direct indicator giving the number of collisions during
the driving scenario. Clearly, the objective of the 3DATIS system is to reduce this
number.
2. Response time: the TTC at the reaction point
3. Steering performance (degree):
The Mean Steering wheel angel is the absolute mean value of turning the steering
wheel to deal with the event, between the action-start and action-end points. Max-
imum Steering wheel angle (SA): The relative value between the maximum steering
wheel angle and minimum one which shows the range of the steering wheel angle.
4. Gas pedal performance (feet/second2)
Mean acceleration is the absolute mean value of acceleration due to the throttle
during the stage.
Maximum acceleration is the difference between maximum and minimum accelera-
tion during the stage. It gives the amplitude of the acceleration.
5. Brake pedal performance (feet/second2)
Mean deceleration is the absolute mean value of deceleration due to the brake pedal.
Maximum deceleration is the difference between the maximum and minimum decel-
eration during the stage and gives the amplitude of the deceleration.
6. Video recording: The behavior of the participants during the experiment were
recorded by two cameras which is used to research drivers’ behaviors. The video
can be matched with the raw data set, used to have a double check with drivers’
15 |
Chalmers University of Technology | 3. Results
AscanbeseeninTable3.2,boththemaximumvaluesandthemeanvaluesdecreased
in Cutin. On the other hand, in Pedestrian, the three values illustrated in Table
3.2 were increased. Considering that the number of collisions decreased for both
scenarios, the decrease for Cutin and the increase for Pedestrian were both the right
behavior. Because in Cutin scenario, the participants would have a preparation
when they saw the car parking on the road shoulder (the object), and they would
finely adjust the car before the object moves. That means the participants reaction
became more accurate and stable. However, in Pedestrian they know they need
bigger reaction to avoid collision under the help of 3DATIS as the situation was
more urgent.
3.2.3 TTC at the reaction point
TTC at the reaction point will decide when and how much time the drivers have
left to handle the emergency driving situation. The results are shown in Table 3.3.
Scenario BASE Mean BASE STD UI Mean UI STD
Cutin 2.27 0.74 3.32 1.75
Redcab 3.43 1.02 6.93 4.74
Intersection 4.01 1.73 6.67 3.29
Pedestrian 3.06 1.41 5.48 2.80
Overtake 0.57 0.56 2.80 1.14
Table 3.3: TTC at the reaction point.
Table 3.3 illustrated that with the help of 3DATIS, the TTC at reaction point of
all scenarios got improved approximately two to three second. That is to say, the
drivers brought their reaction forward so that they could have two or three seconds
more to deal with the emergency.
3.2.4 The result of P1 (from advisory voice on to reachtion
point)
As presented in Section 2.4, the measurements illustrated by scenarios as followed
and the significant difference only shown in P1 (from advisory voice on to reaction
point). The abbreviation:
STD: standard deviation
M: mean
Acc: acceleration
Longi:longitudinal
Pos: position
Ang: angle
In Cutin, all the three measurements decreased which fits the conclusion from Sec-
tion 3.2.2, participants’ behaviour became more accurate and stable with the help
of 3DATIS. However, the standard deviation of steering wheel angle increased. This
20 |
Chalmers University of Technology | 3. Results
Cutin Measurements BASE Mean BASE STD UI Mean UI STD
Lateral_Acc_STD 0.22 0.19 0.08 0.04
Longi_Acc_STD 0.10 0.12 0.05 0.05
Steering_Ang_STD 1.17 0.32 0.96 0.37
Table 3.4: P1 of Cutin.
means before the emergency event start, participants had already tried to adjust
their lateral positions by using steering wheel.
Redcab Measurements BASE Mean BASE STD UI Mean UI STD
Longi_Pos_M 7.39 0.39 7.08 0.45
Longi_Pos_STD 0.17 0.10 0.11 0.07
Table 3.5: P1 of Redcab.
Intersection Measurements BASE Mean BASE STD UI Mean UI STD
Lateral_Acc_M 0.01 0.02 0.03 0.05
Lateral_Acc_STD 0.05 0.04 0.10 0.08
Longi_Acc_M 1.51 1.53 0.72 1.53
Steering_Ang_STD 0.60 0.23 1.23 0.75
Table 3.6: P1 of Intersection
As can be seen in Table 3.6, the mean and standard deviation of lateral acceleration
increased as well as the standard deviation of steering wheel angle. This means the
participants were given the awareness when using 3DATIS and they controlled their
steering wheel in order to avoid the collisions. On the other hand, the decrease of
longitudinal acceleration also shows that they knew they should drive safer in this
situation.
3.3 Drivers’ subjective feedback to 3DATIS
To measure the overall acceptance and the two sub-measures, usefulness and satisfy-
ing, the mean values of nine items on the two acceptance subscales of all participants
were calculated. Considering the scale starts at -2, the usefulness ratings are high,
especially on item 1,7 and 9. The questionnaire is shown in the Appendix, Ques-
tionnaire Part3
Relatively, most scores are neutral or negative which indicate that the auditory
information might be irritating during the drive. All rates regarding the satisfaction
score are lower than the rating for usefulness. We performed a paired t-test to eval-
uate the participants’ perception on usefulness and satisfaction. The results showed
that the usefulness rate was significantly higher than satisfaction.
21 |
Chalmers University of Technology | 3. Results
As shown in Table 3.7, usefulness score was calculated from item 1,3,5,7 and 9,
meanwhile satisfying score was calculated from item 2,4,6 and 8.
Item Number Mean STD χ2
1. Useful 30 0.97 0.78 24.00
2. Pleasant 30 0.10 0.92 15.00
3. Good 30 0.37 0.93 18.00
4. Nice 30 -0.40 1.04 10.66
5. Effective 30 0.43 0.94 25.66
6. Likable 30 -0.40 0.89 20.00
7. Assisting 30 0.87 0.94 25.00
8. Desirable 30 0.23 1.04 12.33
9. Raising alertness 30 1.13 0.73 33.33
Usefulness 30 0.77 0.77 29.67
Satisfying 30 -0.18 0.91 20.67
Table 3.7: P1 of Intersection
3.4 Discussion
The results of SAGAT questionnaire shows that the 3DATIS system helps the
driversgainawarenessofthesurroundingenvironment. EspeciallyforthePedestrian
and Overtake test cases, the environmental understanding is significantly increased.
These two scenarios have a commonality: both of them have a disadvantage to the
participant that the objects are all hidden in blind spots so that the participants
will not see them until in close range. So the fact is that the systems give the par-
ticipants understanding of the surrounding environment even though they cannot
see the objects. On the other hand, participants’ understanding to the Redcab de-
creased. The reason behind this was explained in the Section 3.1.1.
In Cutin, with the help of 3DATIS, participants’ mean and maximum accelera-
tion shows a decrease tendency which means the action became more accurate and
stable. It shows the participants had prepared for the event.
It is interesting to see that in Redcab, the participants’ performance was oppo-
site to other scenarios. The maximum and mean acceleration at the reaction point
mentioned in Section 3.2.2 are significantly increased, compared to the BASE. It
means the participants were not prepared and performed in a panic. The mecha-
nism behind this is probably that there are two objects in the scenario, and one
of them is the disturbance and used to interfere with participants’ judgment. It
resulted in that when the voice was triggered, the participants could not distinguish
whichobjectisthevoicereferredtoandthuscausingthedropindriverperformance.
22 |
Chalmers University of Technology | 3. Results
In Intersection, besides the TTC at the reaction point, there is no obvious differ-
ence can be detected from the measurements. That is to say, except that participant
would behave earlier, the reactions including turning the steering wheel, releasing
the gas pedal and brake performance were more or less the same. It is also in-
teresting to notice that although the object of Cutin and Intersection behave in a
similar way, i.e. both of them are suddenly moving from right side to the left side,
participants reacted differently for the two cases. Drivers were more prone to decel-
erate in the Intersection, whereas in the Cutin the participants more often bypassed
the obstacles. As Swedish drivers are used to stopping for pedestrians crossing the
road, the result might be different for drivers from other parts of the world, where
stopping for pedestrians is not so common.
In Pedestrian, the road narrowed down to a single lane and the object (a pedes-
trian) walks unexpectedly from behind a car parked at the right side of the road.
Hence, the participants cannot see the object beforehand, unlike the other scenarios.
Therefore, in this particular case, participants need to quickly release the gas pedal
and turn the steering wheel to avoid collision, which means that if the standard
deviations of these quantities is large, the indicates a "good" reaction. When double
checking the video recorded from the live experiments, it is also noticed that without
the aid of the 3DATIS, a significant proportion of the participants fully applies the
brakes when seeing the pedestrian[18], indicating that the drivers were uninformed
of the presence of the obstacle.
In Overtake, the obstacle approaches from the rear at a relatively high speed of
100 km/h[18]. Hence, the advisory sound would directly trigger at warning level (as
it is decided by TTC) and therefore there would be only approximately two seconds
left for the participant to take action to avoid the collision. From the observation of
the videos, we noticed that when the advisory voice was triggered, most participants
started to regulate the steering wheel or the gas pedal immediately.
23 |
Chalmers University of Technology | 4
Process mining
When talking about the data, one can always ask four questions in any situations:
1. What happened? e.g. The ego vehicle deviate from the designed path.
2. Why did it happen?e.g. Why people deviate from the designed path?
3. What will happen? i.e. What can we learn from historic information to make
predictions about what is happening at this point.
4. What is the best that can happen? i.e. We want to use analytics to recom-
mend certain things that’ll improve the situation.
In Chapter 3, we explained "What happened?" and "Why did it happen?", but still,
we cannot directly answer the third and fourth question. Hence, we need to find a
special way to analyse those data to get answers. Thinking that all the behaviors
to the emergency, including turning the steering wheel, hitting the brake pedal, re-
leasing the gas pedal etc. are actually a continuous process, if we can find a way to
build the process model, maybe we can answer the last two questions.
Wil M.P. van der Aalst [8] introduced an algorithm to discover the process model
in Petri Net. Verwer et al. [10] introduced a timed syntactic pattern recognition
to solve the limitation that Real-time automata has never actually been applied to
real data. Salehi et al. [11] created a machine-learning algorithm to identify a car
and its driver from detailed driving data.
From the reference above, one can make a prototype in the mind, a discrete event
model related to drivers’ behaviors and time may be built by mining the raw driv-
ing data. Hence, we will try to use the process mining methods to analyse driving
behaviors.
To verify the effects of the auditory advisory system, data with respect to the driver
behaviors collected from the experiments need to be analyzed and evaluated. Pro-
cess mining, i.e, extracting valuable, process-related information from the event
logs, complements existing approaches to Business Process Management (BPM) [8].
BPM primarily focuses on the analysis of process management and the organization
of the work from process automation and process analysis, and aims to improve op-
erational business processes, possibly without the use of new technologies. On the
other hand, BPM is often associated with software to manage, control and support
operational processes. As BPM heavily relies on process models, process mining
plays a very important role in raw data analyzing. Process mining focuses on the
25 |
Chalmers University of Technology | 4. Process mining
process but utilizes the actual data, are bridges the gap between classical process
model analysis and date oriented analysis like data mining and machine learning.
In this case, related to the driving behavior, data like steering wheel angles, ve-
locities, accelerations, reaction times, longitudinal/lateral position, will contribute
to build a model to illustrate how drivers behave with and without the auditory
advisory system in different scenarios.
Process mining is a relatively new research discipline that sits between machine
learning and data mining on the one hand and process modeling and analysis on
the other hand [11]. The idea of process mining is to discover, monitor and improve
real processes (i.e not assumed processes) by extracting knowledge from event logs
readily available in today’s systems [11]. Hence, we can consider the process mining
as a "super glue" between data and process as well as the fusion between business
decision makers and IT people.
Figure 4.1 gives an overview in a sense. It illustrates that process mining connects
BPM and classical data in this pattern.
Figure 4.1: The relationship connecting with data mining and BPM.
Figure 4.1 illustrates that process mining establishes links between the actual pro-
cesses and their data on the one hand and process models on the other hand. As
shown in Figure 4.2, the "World" represent the different types of raw data. Most
information systems store information in unstructured form. For example, raw data
is sometimes scattered over many tables. In such cases, event data exist but some
efforts are needed to extract them.Thus by using a software system (in a company, it
can be provided by the IT department according to the requirements from the data
26 |
Chalmers University of Technology | 4. Process mining
analysts), the event log which records the necessary information in some specific for-
matcanbegotten. Then, byusingsomealgorithm(herewewilluseAlpha-algorithm
[8]), the process model can be discovered to analyse the "World", i.e to answer the
questions illustrated at the beginning of this chapter.
Figure 4.2: The steps and relationships in process mining. [8]
4.1 From raw data to event logs
After finishing the experiment, the raw data set includes all information from the
simulator including the time, velocity, steering angle, brake pedal, gas pedal, ac-
celeration, distances to road users, TTC to the road users, road users’ speed and
etc.. Those data exist in the columns of the raw data set. How can we extract the
eventlogsfromtherawdatatofittherequirementsofdiscoveringtheprocessmodel?
For the row data, we have three main columns of data need to be extracted i.e.the
data from the steering angle, gas pedal and brake pedal. We need to compare the
order of each reaction in the data stream, as there can be not just one action for each
stream. However, to transform the row data to the event log will become sophisti-
cated because there are much noises in the columns. There are many reasons that
can cause those shakes or tingles, like muscle control and driving habits. Hence, we
can consider these “useless behaviors” as noise superimposed on the “right behav-
iors”. The task is thus to separate the noise from the actual driver behavior.
The spectral subtraction method [24] is a simple and effective method for noise
reduction. As the noisy is uncorrelated, and there are no connection with any oc-
curring incident, the spectral subtraction will be a good choice to remove the noise
27 |
Chalmers University of Technology | 4. Process mining
4.2 Event logs
Figure 4.5 shows a standard event log sheet used in process mining. Furthermore,
theeventlogsshowninFigure4.5aregeneratedusingdatacollectedfromthedriving
experiments. This figure shows that there are three fundamental variables, as can
be seen by the respective columns.
4.2.1 An example of the event log
Figure 4.5: An example of the event log of Cutin.
In Figure 4.5, a single row does not mean a complete process instance, just an event.
Since a data set used in process mining consists several events, these data are often
referred to as an event log. In an event log:
• Each event reflects to a behavior that was implemented in the whole process.
• Multiple events are linked together in a process instance or case.
• Logically, each case forms a sequence of events—ordered by their timestamp[8].
As seen in the Figure 4.5, one case ID is constructed by a complete series of han-
dling behaviors during the scenario Cutin. Thus, the number of the drivers can be
the case ID. There are several activities occurring for each case and each of them
corresponds to a driving behavior.
• Gas off: releasing the gas pedal.
• Accelerate: hitting the gas pedal.
• Brake: hitting the brake pedal.
• Turn left: turning the steering wheel left.
• Turn right: turning the steering angle right.
30 |
Chalmers University of Technology | 4. Process mining
• Event on: The road users move the car/start to cross the road suddenly.
The activities above are assumed reasonable for the process mining. Furthermore,
the time-stamps columns show the handling time of each action clearly. The times-
tamps give additional information when investigating the driver behavior patterns.
For example, by comparing the mean value of each action’s time and frequency, the
statistically most likely driving behavior, based on the choices of all participants,
can be determined for each scenario.
From the data sample in Figure 4.5, we can see why even doing simple process
related analyses, for example count the frequency of each behaviors, or the time
between activities, is no way to use standard tools such as Excel. Process instances
are scattered over multiple rows in a spreadsheet and can only be linked by adopting
a process-oriented meta model. For example, if we look at the rows 6-11 in Figure
4.5, you can see one process(Person 2) that starts with the status Registered on 1st
February 2016, 00:00 (a relative value), moves on to 00:00:03 where we can see the
complete action last for three seconds.
There are three types of process mining that can use event logs as illustrated in
the Figure 4.2, discovery, conformance, enhancement. The first one is discovery.
The discovery technique which we will use in this thesis takes an event log and
thus creates the process model without using any other types of data. An example
is the Alpha-algorithm[13] which will be introduced in the coming sections. This
algorithm uses the event log to create a Petri net to describe the behavior. For
example, by using the Alpha-algorithm, the Petri net can be directly built without
using redundant information.
4.2.2 Case ID
A case ID is a defined instance of an action/process. The precise meaning of a case
depends on in which stage in the process the case is found. For example:
• In a hospital, booking a doctor is one case.
• In a purchasing process, the case could be the writing of a purchase order.
• In a police station, recording reporters’ information would be a case
For every event/activity, one need to identify which case it belongs to so that the
process mining algorithm can distinguish between the different executions.
The case ID can determine the scope of the process and where the process starts
and ends respectively. In fact, the case ID can be constructed in more than one way.
An example could be the following:
If professors from Division of Systems and Control would like to buy some stuff
in their offices, like printers, computers or some new chairs and etc., then the pur-
chasing process can be set up in two approaches:
31 |
Chalmers University of Technology | 4. Process mining
1. One can regard the processes of a specific lead through the purchasing funnel as
the process you would like to analyze. Thus the product lead number will be the
case ID. E.g. computer 1, chair 2 and etc..
2. On the other hand, one can consider the whole purchasing process for a professor
as the process scope and thus, professors’ names can be the case ID.
The two alternatives are all logical and reasonable based on the goal of purchas-
ing appliances and what kind of reault that people want to get from the data.
Rule #1: The case ID determines the scope of the process[17].
4.2.3 Activity
An activity constructs one step in the process. For example, if a professor wants to
buy a new computer for his office, then the process may contain the following activi-
ties: Writingreport,Sendingtothesecretary,Approvedbythefinancialdepartment,
Request rework, Ordering from the shop, Refuse etc.. Some of the activities might
be executed more than once. For instance in the example above, the action “Writing
report” might be executed every time when “Request rework” occurred.
All the activities in the process or the different procedures should be named. If
one has only one activity for each case, then the event log is not specific enough.
The events can also include outlier actions not only just the information that at-
tracts people. Of course, it is also necessary to “clean” those outliers before the
analysis to access a relatively better conclusion.
Rule #2: The activity name determines the level of detail for the process steps[17].
4.2.4 Timestamp
The third mandatory data is the timestamp which displays the time when each ac-
tivity occurs. It is essential for both building the sequence for the activities and
investigating the timing characters of each case. It is all right to just have a start
time, however, if an event log also includes a complete time column, like illustrates
in Figure 4.5, the time of each activity period can be found. This is also called
execution time or activity handling time[17]. The activity period helps one to di-
vide the inactive waiting time from the active one and this is one of the advantages,
compared to having only a start time column.
Rule #3: If you don’t have a sequentialized log file, you need timestamps to
determine the order of the activities in your process[17].
32 |
Chalmers University of Technology | 4. Process mining
4.2.5 Other variables
Additional variables can also be added in the event log to analyze other character-
istics or properties of the process. The suitable additional variables depend on the
domain of the process. As this thesis will not use other variables to analyse the
behavior of the drivers, additional attributes will not be introduced. Please see [17]
for more detail on this.
4.3 Petri net
As Petri nets are foundational process notation and oldest and best investigated
process modeling language, this thesis will use the Petri net to represent the process
model. Even if the graphical notation is simple, there are still many analysis tech-
niques that can be used to analyze the Petri nets[?][15][16]. A Petri net describes
an "action flow", and consists of places and transitions. The network structure is
constant, however, supervised by the firing rule, tokens can flow through the “pro-
cess stream”. The state of a Petri net is determined by the distribution of tokens
over the places.
Definition (Petri net): A Petri net N = (P,T,F) where P is a finite set of places,
T is a finite set of transitions such that P∩T = 0, and F ⊆ (P×T)∪(T×P)is a set
of directed arcs, called the flow relation. A marked Petri net is a pair (N,M), where
N = (P,T,F) is a Petri net and where M ∈ B(P) is a multi-set over P denoting the
marking of the net. The set of all marked Petri nets is denoted N.
Figure 4.6: An example of Petri net, traffic light.
33 |
Chalmers University of Technology | 4. Process mining
Here we will use an elementary example to illustrate the Petri nets. A traffic light
is a very simple process consisting of three states which is obviously the red, yellow
and green light, as shown in Figure 4.6.
The process model for the traffic light has the states red, green and yellow as well
as the transitions to shift from one state to another. Here we can see the Petri net
is static which means that the process is fully described and the net will not be
subject to change. However in the Petri nets there are so called tokens, of which
there can be several. The tokens can move from one place to another place. In this
case, when the transition R2G fires, the token will be moved from the state red to
the state green. Then, the transition G2Y can fire moving the token from green to
yellow. Finally, the token will be moved from the yellow to the red as a result of
firing Y2R.
34 |
Chalmers University of Technology | 4. Process mining
4.4 Process discovery: Alpha-algorithm
Process discovery is the main job of the process mining task. From an event log,
a process model can be built , capturing the behaviors found in the event logs.
This section will describe the method mentioned in the previous section, i.e. the
alpha-algorithm.
Figure 4.7: A sketch map explaining Play-in, Play-out and Replay. Picture from
[8].
One of the most essential elements of process mining is to build a strong connec-
tion between a process model and extracting the “reality” from an event log [8],
according to Van der Aalst et al. [8], to Play-in, Play-out, and Replay. The process
discovery corresponds to the Play-in. For Play-in, opposite to the Play-out, the
actions behind the event logs will be regarded as inputs and the aim of which is to
build a model. The alpha-algorithm is such an example of Play-in technique.
Definition (Process discovery):A process discovery algorithm is a function that
maps an event log onto a process model such that the model is “representative” for
the behavior seen in the event log[8]. The key step is to find such a model.
35 |
Chalmers University of Technology | 4. Process mining
4.4.1 From an example to access to Alpha-algorithm
As a function, we know that there should be input and output of the Alpha-
algorithm. We know that the output of the alpha-algorithm is the process model(it
can be a Petri net, BPMN model etc.), so what kind of input does the Alpha-
algorithm use? When we apply the Alpha-algorithm and we just focus on control
flow like the Activity column in the Figure 5.4, we ignore the resources and other
data elements, as well as the actual timestamps of the events taking place. The
ordering is the only thing that need to be taken into consideration. Hence, we
can convert such an event log to a multiset of traces (because the same trace can
appear more than one time) and for each trace there is a sequence of the names
corresponding to the activities. Here is an abstract example:
L = [< a,b,c,d >3,< a,c,b,d >2,< a,e,d >]
1
Here, we can see the log L contains six traces (the index on the right corner mean
1
the number of each order) and they can be considered as six different cases. They
are modeled as a sequence of activity names. So the sequence a, b, c, d were exe-
cuted three times (there are three traces of that type).
The goal of the alpha-algorithm is to capture L to a process model automatically,
1
no matter what kind of way to represent it e.g. BPMN, Petri Net etc.. Here, we
will start the Alpha-algorithm with the ordering relations without considering the
frequency or some other attributes. In this case, we are only interested in finding in
the log.
As we can see in L , the sequence < a,b,c,d >3 represents that the event/be-
1
havior a is followed by b, b is followed by c and c is followed by d. This relationship
is called direct succession. The relations between the event are listed as followed:
• Direct succession: x > y, iff for some case x is directly followed by y, thus for
L , the following relationships can be given:
1
a > b
a > c
a > e
b > c
b > d
c > b
c > d
e > d
• Causality: x → y, iff x > y and not y > x. Hence we can get the following
relations:
a → b
36 |
Chalmers University of Technology | 4. Process mining
a → c
a → e
b → d
c → d
e → d
• Parallel: x||y, iff x > y and y > x. In this case, the direct succession relation
holds in both directions:
b||c
c||b
• Choice: x#y, iff not x > y and not y > x, which means x is never followed by y
and vice versa:
b#e
e#b
c#e
a#d
The relations mentioned above are used to learn patterns in the process. For exam-
ple, if we see a is followed by b but b is never followed by a, Figure 4.8 can be used
to model this behavior.
Figure 4.8: Causality patten: a → b.
If we find that a is sometimes followed by b, but never the other way around and
at the same time, a is sometimes followed by c, but c is never followed by a, and b
and c never followed one another, we can learn the XOR-split pattern as followed.
Figure 4.9: XOR-joint pattern: a → b, a → c, and b#c.
37 |
Chalmers University of Technology | 4. Process mining
Hence, based on these patterns, a Petri net can be automatically constructed from
Table 4.1 as in Figure 4.13:
a > b
a > c a → b
a > e a → c b#e
b > c a → e b||c e#b
b > d b → d c||b c#e
c > b c → d a#d
c > d e → d
e > d
Table 4.1: The relationships extracted from L .
1
Figure 4.13: The result produced by the Alpha-algorithm from L .
1
This example shows the basic idea of how can we transfer the event log to the pro-
cess model. However, in this case, it is actually already a bit more involved and
furthermore, the Alpha-algorithm can handle more situations.
Here, let us look at L in more detail. When we take an event log, we can talk
1
about so-called footprints and Table 4.2 illustrates them. Table 4.2 illustrates the
relation between each sequence displayed by Footprint.
4.4.2 Algorithm
Afterthebasicideahasshown,theAlpha-algorithmcanbedescribedasfollowing[23]:
Definition (Alpha-algorithm):
1. T = {t ∈ T | ∃ ∈ σ}
L σ∈L
39 |
Chalmers University of Technology | 4. Process mining
a b c d e
a #L → L → L #L #L
1 1 1 1 1
b ← L #L ||L → L #L
1 1 1 1 1
c ← L ||L #L → L #L
1 1 1 1 1
d #L ← L ← L #L ← L
1 1 1 1 1
e ← L #L #L → L #L
1 1 1 1 1
Table 4.2: Footprint of L .
1
2. T = {t ∈ T | ∃ = first(σ)}
I σ∈L
3. T = {t ∈ T | ∃ = last(σ)}
O σ∈L
4. X = {(A,B) ∈ X |A ⊆ T ∧ A 6= ∅ ∧ B ⊆ T ∧ A 6= ∅ ∧ ∀ ∀ a →
L L L L a∈A b∈B
B ∧ ∀ a #L ∧ ∀ b #L }
a1,a2∈A 1 a2 b1,b2∈B 1 b2
5. Y = {(A,B) ∈ X | ∀ A ⊆ A0 ∧ B ⊆ B0 =⇒ (A,B) = (A0,B0)}
L L (A0,B0)∈XL
6. P = {p | (A,B) ∈ Y }∪{i ,o }
L (A,B) L L L
7. F = {a,p | (A,B) ∈ Y ∧ a ∈ A}∪p | (A,B) ∈ Y ∧ b ∈ B}∪
L (A,B) L (A,B),b L
{(i ,t) | t ∈ T }∪ {(t,o ) | t ∈ T }
L 1 L O
8. α(L) = (P ,T ,F )
L L L
L is an event log over some set T of activities.The first step that we take is that
we scan the event log (T ) to see what are the activities or what are the transitions
L
that are appear. We just look at the symbols that occur in the event log. These will
be the activities in the process model, and each corresponding to a transition.Then,
step 2 check which is occurred (T ) as the first activities in some traces and T as the
I O
last one (step 3). The fourth step to the sixth step are the key of Alpha-algorithm.
If we think about the process discovery in terms of Petri nets, step 4, 5 and 6
are all about discovering places. As can be seen in the Figure 5.14, we want to
discover places by identifying sets of transitions, A and B where A are the input of
the place and B are the output transitions for the place[8].
Figure 4.14: Place p(A,B) connects the transitions in set A to the transitions in
set B. Resource from[8].
40 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.