University
stringclasses 19
values | Text
stringlengths 458
20.7k
|
|---|---|
Colorado School of Mines
|
Table A.6. Continued
Bor Calci Potas Lithi Magne Sodi Stron Chlo Bro Nitr Sulf Sample Time Taken
Sample
on um sium um sium um tium ride mide ate ate number during Test
Sample 4 stage 2 936 549. 4625
402.6 4996 980.8 5102 4 0 5.9 157
reject .6 1 2
957 1848
Sample 4 Perm 7 46.1 107.4 1673 38.8 1507 0.3 0 11.9 38.2
.1 7
930 2413
Sample 4 Perm 8 72.3 135.2 2185 53 1967 0.4 0 11.7 37.3
.5 9
973 102. 3005
Sample 4 Perm 9 178.5 2773 76.8 2535 0.6 0 11 36.7
.2 3 3
Sample 4 stage 3 101 688. 5514
524.1 6157 1298 7394 5.2 0 5.3 195
reject 4 8 8
917 1931
Sample 4 T200 60.9 126.7 1928 49.3 1752 0.4 0 0 37.6
.6 5
1084 8517
First brine 696 1723 798.3 8060 3552 13 0 130 422
4 9
130 949. 1036 7618 148.
Final brine 659 7426 1720 7 0 247
6 5 3 2 1
27.
First RO Perm -0.1 0.1 0.7 0 0.6 0 0.4 0 0 0
9
50.
Second RO Perm -0.1 0 1.1 0 0.9 0 1.8 0 0.1 0.1
1
Third RO Perm 82 -0.1 0 2.8 0 2.3 0 8.9 0 0 0.1
167
Fourth RO Perm -0.1 0.2 3.5 0 2.8 0 15.2 0 0 0.1
.1
Permian Basin Produced Water
Cycle 1, Step 1,
First RO Perm 0.3 0 0.6 0 0 43.3 0 5.8 0 0 0 1
63 mins in
Cycle 1, Step 2,
2nd RO Perm 2.2 0 1 0.2 0 83.1 0 80.4 0.4 0.3 0 2
60 mins in
Cycle 2, Step 1,
3rd RO Perm 0.5 0 0.4 0.1 0 35.9 0 4.7 0 0 0 3
15 mins in
Cycle 3, Step 2,
4th RO Perm 3.4 0 0.4 0.1 0 51.9 0 26 0.1 0.3 0 4
13 mins in
120
|
Colorado School of Mines
|
Table A.6. Continued
Bor Calci Potas Lithi Magne Sodi Stron Chlo Bro Nitr Sulf Sample Time Taken
Sample
on um sium um sium um tium ride mide ate ate number during Test
60. 1477 2440
Sample 4 T150 66.2 165.1 20.6 14.4 14.6 75.6 0 28
4 6 2
61. 150. 3343 5125
Initial Feed Cycle 2 367.9 39.2 24.4 37.2 171.5 12 25
7 7 0 0
62. 1292 2088
Initial Feed Cycle 3 36.7 154.9 21.1 5.6 6 92.8 101 23.4
2 1 3
83. 197. 4692 6769
Sample 4 T100B 504 59.5 42.1 50.9 287.3 16.3 29.8
6 1 1 3
77. 1342 2318
Sample 4 T200 33.2 136.7 19.6 6.8 5.2 76.3 30.3 25.4
8 3 3
74. 5168 7709 262.
Initial Feed Cycle 1 1232 701 22.4 625.3 464.5 345.2 0
3 4 3 2
Initial Feed Cycle 1 72. 5395 8178 258.
1136 688.8 23.2 610.7 446.3 329.9 5.5
Sample 2 3 1 8 3
Sample 4 1st stage 71. 2007 3349
95 238.9 29.8 20.5 22.2 149 55.3 203
reject 4 0 0
Sample 4 2nd stage 77. 119. 2478 5347
292.1 36.8 23.7 27.7 238 40.4 390
reject 8 2 9 6
Sample 4 3rd stage 80. 164. 4261 6560
393.6 47.4 33.9 39.9 292 32.1 509
reject 7 4 6 4
Sample 3 stage 3 65. 4204 6729
204 477.3 49.9 35.6 51.9
reject 6 8 8
96. 7528 1E+0 226
Cycle 1 Brine 2102 1001 78.7 1197 815.4 2370
6 1 5 4
557. 131. 7062 9269
Cycle 2 Brine 87 927.7 147.4 160.4 3470 742
8 6 6 7
140 264. 5867 7945 475.
Cycle 3 Brine 758.8 88.1 62.6 78.6 2578 218
.1 4 6 4 8
122
|
Colorado School of Mines
|
ABSTRACT
The MineSENTRY (Mine Safety and Rescue through Sensing Networks and
Robotics Technology) research project has developed a robotic system designed to
extend the range of wireless communications in mining environments. The robotic
system was designed to assist disaster response teams by allowing the team to control
teleoperated robots while remaining safely outside. The proof-of-concept system con-
sists of a teleoperated vehicle and an Autonomous Mobile Radio (AMR). The AMR is
tasked with maintaining equal radio signal strength between the teleoperated vehicle
and the operator.
System testing was conducted at the Colorado School of Mines’ Edgar mine. Sev-
eral simple controllers were developed to guide the AMR during these tests with
mixed results. Although the guidance controllers did not always perform well, the
results and data gathered proved useful for further development. Despite these per-
formance difficulties, the results confirmed the capability of the MineSENTRY design
to maintain communication within a mining environment.
To address problems encountered with the tested guidance controllers, a unique
path-trackingcontrollerisproposedwhichtakesadvantageoftheinherentstructureof
mining environments. The proposed controller departs from contemporary methods
by combining the environment and path into a single description. Further devel-
opment is necessary before the controller can be implemented on the AMR robot;
however, the controller has been successfully tested in simulation with encouraging
results. It was found that the controller may alleviate the need for accurate and
expensive localization typical of current navigation methods.
iii
|
Colorado School of Mines
|
CHAPTER 1
INTRODUCTION
Robotics have gradually dispersed from the factory floor to commercial, private,
and military applications. Within the last few years there has been a growing interest
in applications for humanitarian missions such as search and rescue (SAR). Augment-
ing SAR missions with robotics can potentially increase response time, efficiency, and
safety. Rescue operations benefit from mobility and advanced sensor technologies
allowing robots to reach or perceive environments that human rescuers cannot.
1.1 Motivation
Due to the harsh environments in which SAR robots are intended to operate,
development of SAR robotic systems has proven difficult and a number of key prob-
lems have yet to be solved. Reliable communication between robots and operators is
one of these key problems. Mining environments are particularly difficult as neither
tethered nor wireless systems guarantee reliable communication even during normal
mining conditions. Wire tethers can experience a number of physical conditions
which result in either immobilization or communication loss. High frequency, high
bandwidth radio signals suffer greatly from propagation losses, resulting in severely
reduced range. Therefore, improving communication is a necessary component to
successfully utilize robots in mine rescue operations.
Improved communications would allow SAR robots to travel farther into mines
while providing invaluable sensor data to the rescue workers outside. In particular,
air quality must be known before rescue workers are permitted to enter the mine.
If the mine’s air quality measuring systems fail during a disaster, borehole drilling
is used to lower sensors from the surface into the mine—a time consuming process.
Robots could be deployed quickly and provide extensive air quality information and
1
|
Colorado School of Mines
|
Figure 1.1: MineSENTRY Project Concept [1]
allow rescue workers to make better informed decisions. Furthermore, robots could be
teleoperated from outside the mine and used to search for victims long before human
operators could safely enter.
1.2 MineSENTRY Project
TheMineSENTRY(MineSafetyandRescuethroughSensingNetworksandRobotics
Technology)projectwasaimedatdevelopinganddemonstratingamulti-agentrobotic
platform for mine search and rescue operations as a solution to the inherent commu-
nication problems. The research project was carried out at the Colorado School of
Mines (CSM) and was funded by NIOSH Grant #1R01OH009612-01 through CSM’s
Center for Automation, Robotics, and Distributed Intelligence (CARDI). The intent
was to create an ad hoc wireless network utilizing multiple Autonomous Mobile Ra-
dios (AMRs) which can dynamically adjust to maintain communication between the
operator(s) and the lead teleoperated vehicle. To demonstrate the viability of this
system, a robotic system was developed at CSM and tested at Edgar mine in Idaho
Springs, Colorado—CSM’s experimental mine.
Conceptually, the MineSENTRY robotic platform maintains communication be-
tween a teleoperated leader vehicle, a modified Bobcat front-end loader, and the
operator (Figure 1.1). The operator remains near the mine entrance while operat-
ing the front-end loader located deeper within the mine, where it would otherwise
2
|
Colorado School of Mines
|
be outside the communication range of a direct wireless link. Several AMRs follow
the front-end loader into the mine and position themselves to create a wireless mesh
network. The AMRs act as nodes within the network, each relaying information to
the next node until the information reaches its destination. The AMRs attempt to
maintain the highest radio signal strength (RSS) possible by automatically readjust-
ing their positions. By using autonomous, mobile relays, the MineSENTRY platform
retains the freedom of wireless systems while drastically extending the robots reach
into the mining environment.
1.3 Thesis Objectives
To create the MineSENTRY robotic system, the MineSENTRY team had to de-
sign, obtain, or modify the teleoperated leader vehicle, AMR, and radio systems. In
particular,theAMRwasdevelopedwiththejointeffortofseveralMineSENTRYteam
members. The author’s primary responsibilities involved the design and construction
of the AMR, including actuation, hardware, power distribution systems, and controls,
although the author assisted in other aspects as well. Originally, software develop-
ment was undertaken by other team members; however, due to their departure prior
to project completion, the author assumed responsibility for completing the control
software.
Theobjectiveofthis thesisistodetailthedevelopmentoftheAutonomousMobile
Radio, with a focus on the actuation, power, software, and control systems. Specifi-
cally,theguidancecontrollersusedduringproof-of-concepttestingaredescribedalong
with observations regarding their performance. Motivated by these observations, a
unique control approach is proposed for use with the AMR system; also, simulation
results are presented and analyzed. Documentation is provided to allow others to
continue, improve, or expand on this research.
3
|
Colorado School of Mines
|
1.4 Chapter Overviews
To provide the aforementioned information, the thesis will cover each topic in the
following chapters:
Chapter 2 - Literature Review
In this chapter, a brief description is given regarding the use of mobile robots in
mining environments and SAR applications. Furthermore, an overview of the current
state of mobile robot navigation research provides the reader basic background infor-
mation needed to understand later topics.
Chapter 3 - Development
Details of the AMR development are provided in this chapter. The topics covered
are AMR actuation, hardware, electrical, and software development. This chapter
is not intended to be a complete description of the AMR systems; rather, it com-
plements the theses of other MineSENTRY team members, [1, 2], to provide the
documentation necessary to allow others to reconstruct the MineSENTRY systems.
Chapter 4 - Proof-of-Concept Testing
The results of several MineSENTRY tests, conducted at the Edgar mine, are pre-
sented in this chapter. The results are focused on the guidance controller(s) used
during the tests to guide the AMR through the Edgar Mine. While many difficulties
were encountered that adversely affected performance, several successful tests were
conducted resulting in useful experimental data. This data motivated the develop-
ment of an alternate guidance control, described in Chapter 5.
Chapter 5 - Controller
The unique guidance controller proposed by the author for use with the Mine-
4
|
Colorado School of Mines
|
CHAPTER 2
LITERATURE REVIEW
The MineSENTRY project seeks to develop a robotic system for use in mine
search and rescue (SAR) operations. To gain a better understanding of the difficulties
involvedandthereasonsforundertakingsuchaproject, afamiliaritywiththecurrent,
germane research is necessary. Specifically, the current research on mobile robots,
their application in mining environments, and their application in search and rescue
operations must be evaluated.
2.1 Rescue Robots and Mines
In recent years there has been a great deal of interest in using mobile robots in
search and rescue (SAR) applications [3, 4]. The advantages of mobile robots and
SAR applications are diverse. For example, robots can provide communication and
sensor information while traversing terrain and environments which prove hazardous
for human rescue workers. Mining environments are an excellent example of one such
environment. In the event of a disaster, mines can quickly become inhospitable and
dangerous for both the miners and search and rescue teams. Mine SAR operations
would therefore benefit greatly from the capabilities of mobile robotic tools; however,
mines themselves present a number of challenges for mobile robots.
On January 2, 2006 an explosion occurred at the Wolf Run Mining Company’s
Sago Mine located near Sago, West Virginia, hereafter referred to as the Sago mine
disaster [5]. Approximately 20 hours after the explosion, the Mine Safety and Health
Administration’s rescue robot arrived from a repair facility in Knoxville, Tennessee
and was deployed in conjunction with a rescue team just a few hours later. The
robot, nicknamed V2 (Figure 2.1), was sent in to measure carbon monoxide levels
which would allow the rescue team to advance more quickly into areas where CO
7
|
Colorado School of Mines
|
Figure 2.1: MSHA’s mine rescue robot, a Remotec ANDROS Wolverine robot [6].
levels were safe. Unfortunately, after a few hours of use, an operator error resulted in
damage to the robot’s wheels, disabling it for the remainder of the rescue operations.
The Sago mine disaster is an example of how robots may be used during actual
SARoperations. However, thisdisasteralsorevealsthechallengesinvolvedwithusing
robotsinsuchaninhospitableenvironment. Someofthesechallengesincludemobility
inroughterrain, control, powerrequirements, andcommunication[7]. MSHA’srescue
robot utilizes a 5000 foot fiber-optic cable for communication which is susceptible to
damage and limits the robot’s range. Other researchers exploring robotic solutions
for mining environments have encountered similar difficulties with communication.
Abandoned mines are another safety concern for which robotic solutions are being
researched. In 2003, a robot nicknamed ‘Groundhog’ (Figure 2.2) was used to explore
and map the main corridor of an abandoned mine in Mathies, Pennsylvania [8]. The
robot was designed to operate autonomously outside of wireless communication, but
experienced problems during testing. Intervention over the wireless link proved un-
successful due to the weak signal strength. Similarly, Pilania and Chakravarty are
focusing on communication in the development of their mine SAR robot [9].
8
|
Colorado School of Mines
|
Figure 2.2: Groundhog, the abandoned mine mapping robot [10].
The goal of the MineSENTRY project was to develop a solution to the communi-
cation problem encountered in mining environments. The solution is a wireless mesh
network comprised of autonomous robots that dynamically adjust to maintain com-
munication. Each node (robot) within the network relays received signals to adjacent
nodes. In the MineSENTRY system, the operator transmits to the nearest AMR,
and the signal is relayed to the teleoperated vehicle. This simultaneously solves both
communication range and reliability issues, provided enough AMRs are available to
create a solid mesh network. However, with or without communication all robots (i.e.
the AMRs) are expected to have a certain degree of autonomy and operate indepen-
dent from human control. Thus there is a need for autonomous navigation, which has
its own set of problems that must be addressed for the MineSENTRY project to be
successful.
2.2 Navigation
The literature on mobile robot navigation is extensive and diverse. Although
the research is extensive, no universal solution exists because each situation often
9
|
Colorado School of Mines
|
poses its own unique challenges. In general, navigation methods can be divided
into landmark, beacon, dead reckoning, and map-based methods [11]. Map-based
methods are of particular relevance to the MineSENTRY project and are frequently
researched. To implement a map-based navigation system three challenges must be
addressed: mapping, localization, and path planning [12]. Additionally, the process
of following the planned path, referred to as path tracking, is a commonly researched
topic.
2.2.1 Mapping and Localization
Mapping is the process of exploring an environment and using sensor data to
produce a map which is useful (e.g. for localization) [13]. The guidance controller
proposedinChapter5isdesignedtoutilizesurveymapsthatarecommonlyassociated
with active mines. Because a survey map for the mine being used in testing is readily
available, nospecificmappingtechniquesareimplementedandthusnotechniqueswill
be discussed. However, for the survey map to be useful, the map must be described
in a unique fashion which will be discussed in Chapter 5.
Localization is the process of determining the robot’s position and orientation
(collectively known as pose) from map and sensor data [13]. When implementing
or developing a localization algorithm, the absence of information known before and
during run-time greatly increases the complexity—and therefore difficulty—of the
localization process. For example, global localization is the process of localizing with
no initial information; in other words, the robot begins lost and must use the map to
determine its location. Additional complexity is encountered in the ‘kidnapped robot
problem’, which is global localization that further requires the robot to recognize if
it has been picked up and moved. Conversely, if the robot only needs to track its
deviation from a known initial position, the localization process is known as path
tracking [14]. Path-tracking localization is particularly important to the guidance
10
|
Colorado School of Mines
|
controller proposed in Chapter 5, which will be elucidated in Section 2.2.3.
Originally mapping and localization processes were executed sequentially; how-
ever, recent research focuses on simultaneous localization and mapping (SLAM) al-
gorithms that perform mapping and localization processes simultaneously. The work
of Antoni Burguera is of particular relevance to the AMR. His extensive research ex-
tendsSLAMalgorithms(whichtypicallyrequirehighaccuracyanddatarichscanning
laser sensors) to the relatively inaccurate and sparse data associated with ultrasonic
range-finding (sonar) sensors—the sensor of choice for the AMR [15].
2.2.2 Path Planning
Path planning, also frequently referred to as motion planning, is the process of
findingacollisionfreepaththrough2-Dor3-Dspace. Thisprocessisroughlydivisible
into global and local path planning. Global planning involves finding paths over long
distances to a goal location, typically with some sort of map; whereas local planning
involves finding the immediate motions the robot must make. Early research into
path planning generally focused on local methods, such as the potential field method
[16]. However, these methods had the disadvantage of being easily trapped in local
minima [17]. This disadvantage prompted research into other planning methods, such
as the Vector Field Histogram (VFH) [18] or Generalized Voroni Graph (GVG) [19]
methods.
Other approaches to path planning compute paths by finding the configuration
space (C-space) of the robot; then searching the C-space for acceptable solutions
to reach the goal state [11]. Configuration-based methods have the disadvantage of
being computationally intensive, but advances in computer technology have reduced
the cost of such methods. For example, the Groundhog robot applied the A* search
algorithm in C-space to plan its paths [8]. An advantage to configuration space
methods is its applicability to both global and local path planning.
11
|
Colorado School of Mines
|
2.2.3 Path-Tracking Control
Onceacollisionfreepathisfoundviaanyofthenumerouspath-planningmethods,
a path-tracking controller is required to guide the robot along the desired path. Path-
tracking control is a very diverse research topic. Path-tracking controllers use a
large variety of techniques including fuzzy logic [20], neural networks [21], spatial
methods [22], and many others. Often vehicle controllers focus on incorporating the
kinematics (or dynamics) of the vehicle being controlled [23]. However, if accurate
localization data is available, and the operating conditions allow the vehicle dynamics
to be ignored, simple path-tracking controllers may be implemented. For instance,
the Groundhog robot [8] utilized a simple Proportional-Derivative (PD) controller to
track and follow the paths generated by its path planner.
The guidance controller proposed in Chapter 5 is categorized as a path-tracking
controller. However, the proposed controller adopts a unique approach to the path-
tracking problem in an effort to reduce the need for accurate localization. To ac-
complish this, the algorithm requires a unique way of describing the path. This path
description is a key contribution of the thesis.
2.3 Summary
There are many advantages to using mobile robots in mining environments; how-
ever, there are also technical challenges that impede their implementation. MSHA’s
mine rescue robot, V2, and the abandoned mine mapping robot, Groundhog, are ex-
amples that exemplify both the advantages and challenges. Recognizing these chal-
lenges, the MineSENTRY project seeks to address the problem of communication
within mining environments. To accomplish this goal, the MineSENTRY research
team sought to develop the Autonomous Mobile Relay (AMR) robot; which, as its
namesake implies, requires a certain degree of autonomy. Current research into mo-
bile robot navigation reveals many options that could have been used to implement
12
|
Colorado School of Mines
|
CHAPTER 3
DEVELOPMENT
BeforediscussingthecontrolalgorithmsimplementedontheAMR,itisimportant
to understand the platform upon which it was developed. As previously mentioned,
the MineSENTRY research team built the AMR robot using a golf cart as the vehicle
platform. In order to convert a golf cart into a robot, a number of systems had to
be developed including actuation, hardware, electrical, and software. Each of these
topics are discussed separately in the remainder of this chapter.
3.1 Actuation
To transform a golf cart into the AMR robot, the control electronics must be able
to control the steering, brake, and throttle. Thus, actuation was installed that allows
electronic commands to be realized as mechanical motion. The steering and brake,
both purely mechanical on the stock golf cart, required the addition of electrome-
chanical actuators; however, the golf cart’s throttle was electromechanical initially,
and only required additional circuitry to allow computer control. Before adding any
actuation, several requirements were established to govern the design process. These
requirements will be discussed in the following sections; however, a universal require-
ment was to keep modification to the factory golf cart minimal. By reducing the
amount of modification necessary, the design becomes nearly modular. Provided the
necessary components as a kit, a small team of technicians could easily mount an
identical system onto another golf cart with minimal time and effort.
3.1.1 Steering Overview
The requirements for the steering actuation included the necessity to return the
vehicle to fully manual steering. This capability is extremely important both during
15
|
Colorado School of Mines
|
softwareandhardwaredevelopment. Duringtesting, itoccasionallybecamenecessary
to manually drive the vehicle back to the lab to fix software problems that would have
otherwise rendered the vehicle immobile.
Allowing for manual steering requires the steering shaft to be mechanically de-
tached from the actuator. The mechanical disconnect is necessary because back-
driving the servomotor’s gear box results in unneeded mechanical wear and could
damage the gears; furthermore, back-driving the servomotor adds unneeded resis-
tance to the steering during manual operation. Damage to the servomotor gear box
became a primary concern because of the relatively low output and input torque limit
specifications. The steering can be easily back driven by applying force at the wheels,
creating concern that a solid impact at the wheel could cause a transient torque at
the servomotor in excess of the gear box limit. The DC servomotor was also capable
of exceeding the input as well as the output torque limit of the gear box.
Several solutions for protecting the gear box were considered including commer-
cially available friction clutches, friction-based mechanisms, and shear pins. Most
commercial friction clutches had insufficient torque ratings given the design require-
ments. Friction-based systems, such as a cone shaped friction clutch, could provide
both gear box protection and a means of disengaging the steering. However, to pro-
vide the torques required the clutch would either need to be prohibitively large or
require high axial forces while engaged. Thus, the author decided to use a shear pin
for the gear box protection due to its simplicity and reliability. In order to disengage
the steering actuator, a positive locking assembly was designed to lock and unlock the
driving pulley from the steering shaft. A full description of the final steering assembly
as well as the actuator description and selection process is described below.
16
|
Colorado School of Mines
|
Figure 3.1: Modified Steering Column
3.1.2 Steering Assembly
Before going into the details of the steering assembly, it’s necessary to describe
the stock steering assembly of the golf cart. The stock golf cart steering assembly is
very simple; consisting of a hollow steering column, a steering shaft, steering wheel,
universal joint, and a rack and pinion drive. The steering column is a steel tube
that bolts to the firewall and houses the steering shaft. The steering shaft is directly
attached to the steering wheel and to a universal joint behind the firewall. Finally
the universal joint is connected to a rack and pinion steering assembly (Figure 3.1).
To minimize themodificationsrequiredtoaddactuatedsteering, theentire assem-
bly is attached to the steering column near the steering wheel. The only modifications
needed to attach the steering assembly are to shorten the steering column and add a
Woodruff key seat to the steering shaft. The actuator assembly clamps to the end of
the steering column and replaces the stock bushing used to stabilize the end of the
steering shaft.
AsshowninFigure3.2,thereareseveralcomponentsusedinthesteeringassembly.
The servomotor which directly drives the drive pulley. The drive pulley is not directly
17
|
Colorado School of Mines
|
Figure 3.2: Steering Actuation Assembly
fixed to the servomotor output shaft; rather, a steel sleeve is pinned to the servomotor
output shaft and the drive pulley can freely rotate around the sleeve. The end of
the sleeve is tapped and keyed. An end cap is notched to match the sleeve’s key
and is fastened to the sleeve with a cap screw. The end cap rotates the drive pulley
throughashearpin. Theshearpinprotectsthesteeringactuatorgearboxbybreaking
if the torque on the output shaft exceeds a certain limit. The drive pulley drives
the steering pulley through a timing belt. The steering pulley is mounted on the
steering shaft but may freely rotate on the shaft. To drive the steering shaft, a
spline assembly is keyed to the steering shaft. The spline assembly sits next to the
drive pulley, and by virtue of the splines can rotate with the steering shaft and
have some axial motion. The axial motion of the spline assembly is used to fix the
steering shaft to the drive pulley via a steel pin. The geometry of the purchased
steering pulley required the addition of an inset brass disk to provide the mating
holes for the steel pin. By adjusting the spline assembly to engage or disengage the
steering pulley, the servomotor may be mechanically disconnected from the steering
column. This allows the golf cart to be driven manually without back-driving the
18
|
Colorado School of Mines
|
Figure 3.3: Brake Actuator Assembly
requirement can be easily achieved.
3.1.5 Brake Assembly
The selected brake actuator is a common linear actuator with potentiometer po-
sition feedback. The stroke length, force limit, and linear speed were selected based
on the previously mentioned specifications. To allow manual operation of the brake
at all times, the brake actuator is not rigidly connected to the brake; rather, a sleeve
is connected to the actuator piston and a push rod is mounted to the brake pedal
lever (Figure 3.3). This configuration allows the actuator to push on the brake lever
but retract independently of the brake. Note that if the brake actuator is left in
an extended position while the golf cart control systems are turned off, the brake
cannot be released manually. To resolve this, the golf cart control systems are set to
automatically release the brake when placed in manual mode.
The manual braking requirement created some small problems with the remote
or automatic control modes. The stock golf cart has no mechanism to prevent the
21
|
Colorado School of Mines
|
Figure 3.4: Unmodified Golf Cart Throttle Sensor
allows manual operation at any time.
Unfortunately mimicking the sensor was not an easy task. The sensor itself es-
sentially acts as a variable resistor which changes its resistance based on the position
of the plunger. The difficulty arises from the fact that the sensor is not grounded
but rather floats at some voltage above ground. Without access to the stock golf
cart controller’s circuitry, or a schematic thereof, this behavior could not be reliably
reproduced. Consequently to mimic the sensor, the throttle control circuit had to be
electrically isolated from ground through an isolating transformer. Figure 3.5 shows
the throttle control circuit, which is mounted underneath the seat near the golf cart
controller. For a more detailed description of the throttle control circuit, please refer
to [2].
3.2 Hardware
This section discusses the hardware required to mount or enclose the electrical
components of the AMR, particularly sensors. In total, 24 sensors are mounted to the
golf cart: fourteen ultrasonic range finders (sonars), four infrared range finders (IRs),
four Hall effect encoders, one string potentiometer, and one brake switch. Detailed
23
|
Colorado School of Mines
|
Figure 3.8: Brake Switch
enclosure that houses the Vehicle Controller, power supplies and distribution boards
was carefully selected, modified, and mounted to the golf cart. The actuator control
board, which provides the high-power drive circuitry and low-level feedback control
for both actuators, was mounted in the front of the vehicle in a pre-existing storage
compartment.
The primary enclosure is a Bud Industries steel NEMA enclosure and was orig-
inally selected to contain the Vehicle Controller (microcontroller board), actuator
power supply, power distribution unit, and the Autonomous Controller (single board
computer). Although the space was allocated, the single board computer was aban-
donedinfavorofanetbookastheAutonomousControllerduetosoftwareissues. The
enclosure is mounted using the golf bag rack on the golf cart (Figure 3.9). The bag
rack has many pre-existing holes which may be tapped using a 1/4-20 tap. Simple
mounting brackets were built to attach the primary enclosure and vibration dampers
were used to reduce the shock loadings felt by the components mounted inside. A
hole was cut on the bottom of the enclosure to allow cable routing and has an inset
nylon brush grommet to help protect the enclosure from dust and debris. A hole was
cut in the back of the enclosure to mount a cooling fan. The cooling requirements
27
|
Colorado School of Mines
|
The control electronics are organized such that lower-level functionality (e.g. com-
municating with sensors) is handled by the Vehicle Controller which is a digital signal
processing microcontroller from Microchip. The Vehicle Controller reads the sensor
outputs and converts the raw electrical signals to integer values with appropriate
units. For example, the sonars output an analog signal that the microcontroller con-
verts to distance measurement in millimeters. The Vehicle Controller also handles
communication with the actuator control board which drives the actuators.
The actuator control board is a Roboteq AX3500 driver board that is capable
of driving/controlling two actuators (Figure 3.10). The AX3500 implements a PID
control loop and can accept a variety of position feedback signals. Both the brake
actuator and the steering servomotor are controlled by this board. The servomotor
uses an optical encoder for position feedback and the brake actuator uses an analog
potentiometer for position feedback. The position setpoints are sent by the Au-
tonomous Controller to the actuator control board via an RS-232 connection to the
Vehicle Controller. Other than some low-level control functions, such as releasing the
brake or calibrating steering, the Vehicle Controller simply relays brake and steering
commands from the Autonomous Controller to the actuator controller.
The Autonomous Controller is a netbook laptop running control software. The
Autonomous Controller handles higher-level functions such as interpreting sensor sig-
nals and guiding the robot through the environment. Additionally, the Autonomous
Controller communicates to the other AMRs and the base station through the Rajant
radiomeshnetwork. TheAutonomousControllercanreceiveandinterpretcommands
from the base station that determine how the AMR should move to maintain radio
signal strength. These operations are implemented using the high-level programming
language Python.
33
|
Colorado School of Mines
|
3.4 Software
Operation of the AMR requires a variety of software. There is software imple-
mented on the base station which oversees and controls the mission objectives. The
Autonomous Controller software communicates with the base station and determines
the specific AMR control actions. Finally, there is firmware implemented on the Ve-
hicle Controller which handles all the lower-level control functions of the AMR. The
author did not contribute to the development of either the base station software or
the microcontroller firmware, so neither will be discussed. The details of the base sta-
tion software and microcontroller firmware may be found in [1] and [2], respectively.
Initial development of the Autonomous Controller software was performed by Chris-
ter Karlsson and Ken Anderson, but was completed by the author with contributions
from Chris Meehan.
The autonomous control software is implemented in the high-level programming
languagePythonandmayberunfromaLinuxorWindowscomputer. Theuserinter-
acts with the software through a custom graphical user interface (GUI), a screenshot
of which is shown in Figure 3.13. Under Global Settings, the user must first select
the communication port and perform a communication test to initiate a link with
the Vehicle Controller. Currently, the communication is accomplished with a USB
to serial adapter cable. If the Vehicle Controller is on and properly connected to
the computer (Autonomous Controller), the communication test will connect to the
Vehicle Controller and start passing data. Once the communication test is performed
successfully the other GUI functionality becomes available.
Most of the GUI’s screen real estate is used for monitoring sensor feedback and
displaying status information. On the bottom edge of the window, a series of indica-
tors display the AMR control mode, drive direction, and brake status. In the center
of the window there are two graphing areas which can each plot data for one sensor.
The graph source can be selected via the drop-down boxes under Global Settings,
34
|
Colorado School of Mines
|
communication method.
Besides queues, flags are used to communicate the status of the threads. The
flags are Boolean values and are primarily used to assist thread communication. For
example, when the RC-Comm thread receives new data it checks the flags of the other
threads to see if they are ready to receive more data. If the flag is cleared, RC-Comm
puts data into the corresponding queue and sets the flag. When the receiving thread
retrieves the data from the queue, it clears the flag to indicate it is ready for more
data. In the case of the Rajant server, the control thread uses flags to indicate when
it has completed the most recent command and is awaiting more commands.
3.5 Summary
To summarize, the actuator assemblies allow manual, remote, and automatic con-
troloperationofthevehicle. TheVehicleControllerhandlesalllow-leveltaskssuchas
gathering sensor data, monitoring E-stops, and relaying information to/from the Au-
tonomous Controller. The Autonomous Controller handles higher-level tasks such as
communicating with the mesh network, guidance control, and user interface. Finally,
the software defines the operation of both the Vehicle Controller and Autonomous
Controller. With a basic understanding of the hardware, electronics, and software
implemented on the AMR, we may now begin to discuss the guidance controller de-
veloped on the AMR platform.
39
|
Colorado School of Mines
|
CHAPTER 4
PROOF-OF-CONCEPT TESTING
The MineSENTRY team performed proof-of-concept testing at the Edgar mine
in Idaho Springs, CO. For these tests, three simple guidance controllers were de-
signed and implemented on the AMR’s Autonomous Controller with the intent of
autonomously navigating a portion of the Edgar mine’s Army tunnel. While a fully
functional guidance controller would be expected to control the speed, brake, and
steering of the vehicle, these test controllers focused on controlling the AMR steering
only. During the testing described in this chapter, the speed was set to a constant
value (maintained by the Vehicle Controller); the MineSENTRY base station auto-
matically commanded the AMR when and how far to move. The test controllers
were:
• A fuzzy logic controller designed to recognize key situations and react appro-
priately.
• Two basic wall-following controllers:
– a center-drift wall-following controller designed to ignore side-passages
(designated as Controller 1).
– awall-followingcontrollerdesignedtoswitchbetweenleft,right,andcenter
following (designated as Controller 2).
Each guidance controller was tested within the hallways of CSM’s engineering
building, Brown Hall, as well as the Edgar mine in Idaho Springs, CO. Detailed
descriptions of the proof-of-concept testing, specifically regarding the radio signal
strength results, may be found in [1]. The development, testing, and performance of
the guidance controllers is detailed in the following text.
41
|
Colorado School of Mines
|
Figure 4.1: Brown Hall Test Path
hallways. Because the infrared sensors are mounted in different locations this also
required that the controller be trained with a new set of situations specifically for the
IR range finders. Unfortunately, sensor data is unavailable for these tests because the
Python control software was incomplete at this time. Despite the issues encountered,
some successful results with this controller were obtained.
While testing in the Brown building, the desired path for the robot was a straight
line past a T-junction (Figure 4.1). After a few dry runs to tune the controller
and adjust the threshold values for detecting passages, the controller successfully
guided the robot past the T-junction to the opposite end of the hallway several times.
However,itwasnotedthattheperformancearoundtheT-junctionwasnotcompletely
reliable. Conversely, the performance in the straight sections of the hallway was
excellent. Several test runs were conducted with the robot set initially to the right
or left of the hallway or turned at angles from straight. In each instance, the robot
successfully managed to navigate to the center of the hallway and straighten out.
A significant amount of the controller tuning and adjustment was focused on
obtaining the desired response at the T-junction. If the controller did not correctly
identify the T-junction as a passage on the left, the robot would begin to turn left
thendriveintotheoppositecorner. Even if thepassagewasdetected, therobotwould
tend to veer to the left side and over-correct at the reappearance of the wall on the
44
|
Colorado School of Mines
|
Figure 4.2: Typical failures encountered while tuning fuzzy controller.
left, or in some cases not correct at all (Figure 4.2).
Once the controller was properly tuned, the performance near the T-junction
became more reliable as long as the robot was nearly centered and driving straight.
The tuned controller made several successful test runs, driving both East to West
and vice versa, where the T-junction had almost no noticeable effect on the path of
the robot. At this point, we determined that the fuzzy logic controller was ready for
testing at the mine.
4.1.3 Experimental Results - Edgar mine
OnMarch11, 2010theAMRwastransportedtoEdgarminetoperformtheproof-
of-concept testing previously described. The fuzzy logic controller’s performance was
mixed however and passage detection performed poorly. Even in the straight sections
of the mine the controller would tend to veer to one side and the robot had to be
stopped before impact. After debugging it was discovered that a few sonar sensors,
including two of the 45◦ sonars important for passage detection, had been damaged in
transport and were reporting fixed data readings (Figure 4.3). After discovering the
damage to the sensors, the fuzzy-logic guidance controller was disabled. The proof-
of-concept testing, of the base station and radio systems, continued with a human
operator to steer; while the Autonomous Controller regulated the AMR’s speed as
commanded by the base station.
The MineSENTRY team returned to Edgar mine on April 2, 2010 for further test-
45
|
Colorado School of Mines
|
Sonar 2 Sonar 3
6 6
4 4
2 2
0 0
0 20 40 60 80 100120140160180 0 20 40 60 80 100120140160180
Time (sec) Time (sec)
Sonar 5 Sonar 6
6 6
4 4
2 2
0 0
0 20 40 60 80 100120140160180 0 20 40 60 80 100120140160180
Time (sec) Time (sec)
Figure 4.3: Plots showing sonars 3, 5, and 6 which were damaged during transport.
Sonar 2 was working and is shown for comparison. The AMR was moving between
20-45 seconds and was stopped otherwise. The spikes seen in sonar 3’s data were due
to communication errors that were resolved in later versions of the code.
ing of the radio system. The failed sensors had been replaced and extra precautions
were taken during transport to protect the AMR. Unfortunately, even with properly
functioning sensors the controller was unable to navigate past side passages. Thus,
theteststookplaceinastraightsectionoftheArmytunnelasindicatedinFigure4.4.
The controller performance in the straight section of the mine mirrored the per-
formance observed in Brown Hall. The AMR successfully followed the leader (a
MineSENTRY team member with a Rajant radio) using radio signal strength (RSS)
information relayed by the base station. The AMR repositioned five times during
the test. While the AMR was stopped, the leader would reposition and the RSS
signal strength information logged. The approximate path taken by the AMR as well
as the sensor data are shown in Figure 4.5. The sensor data is split between the
45◦ sensors and the side facing sensors to compare the performance. Interestingly,
the performance of the 45◦ sensors appear to be better than the side facing sensors
in these tests, but the plot for the side facing sensors shows three sensors on either
46
)m(
egnaR
)m(
egnaR
)m(
egnaR
)m(
egnaR
|
Colorado School of Mines
|
Figure 4.4: Testing area during the April 2, 2010 tests. The AMR traveled from right
to left and stopped at the turn.
side. The variation observed in the side facing sensors is due to calibration inaccu-
racy and signal filtering, resulting in slightly different reported distances on the three
side facing sensors. In later experiments the sonars were re-calibrated and electronic
filtering added which improved the reliability of the sensors. Table 4.1 summarizes
the observed performance of the fuzzy logic controller.
The issues with the sonar sensors create difficulty assessing the performance of
the fuzzy logic controller. The controller may perform better if tested with properly
calibrated, functioning sensors. However, it was decided the results of the April 2
testing was clear enough to pursue other controller solutions.
Table 4.1: Guidance Controllers Performance Summary
Controller Summary
Location Controller Performance Summary
Brown Hall Fuzzy Logic Successfully navigated T-junction, but is sensitive to
tuningparameters. Excellentperformanceinthehall-
way.
Edgar Mine Testing on Mar. 11, 2010, controller failed due to
Fuzzy Logic
broken sonar sensors.
TestingonApr. 2,2010,controllercouldnotnavigate
side passages, but performed well in straight sections.
47
|
Colorado School of Mines
|
fixed number of the previous sensor values, the buffer is updated based on the odome-
ter data. The buffered values represent average sensor readings taken in increments
of 0.2 meters, and a history is maintained for the last four meters. The buffer also
stores values indicating if the robot was left, right, or center following at the time
the value was buffered. When a passage is encountered, the buffer is searched to
determine what the setpoint distances were while the robot was still center following.
Furthermore, the distances are computed differently from Controller 1. Controller
2 uses the six sensors which point directly to each side of the robot. The distance to
the left or right wall is computed by taking a separate weighted average of the sensors
for each side. The weighting is determined by the magnitude of the sensor values;
the smaller sensor readings are weighted heavier than the larger sensor readings.
Originally, the sensor values were averaged, but this was modified during testing in
BrownHall. Theweightingwasimplementedtoensurethatsensorreadingsindicating
close proximity to the wall were not ignored when averaged with a larger reading (e.g.
when the robot is angled toward the wall).
4.2.2 Experimental Results - Brown Building
As with the fuzzy logic controller, both of these controllers were tested in the first-
floor hallway of Brown Hall. The desired path and desired response were also the
same: drivepasttheT-junctionwhilemaintainingthestraightestpathpossible. Both
controllers were tuned separately to achieve this goal and both performed slightly
differently within the hallway. Because these tests were used to design and tune the
controllers in preparation for testing at the Edgar mine, no sensor data was logged
for later analysis. However, Figure 4.6 shows the difference between the 45◦ and side-
facing sensors utilized by Controller 1 and Controller 2, respectively. The sensor data
was logged at a later date while driving the AMR through the hallway manually.
Controller 1 tended to veer towards the left when passing the T-junction. After
50
|
Colorado School of Mines
|
Figure 4.6: Comparison of the 45◦ and side-facing sonar sensors’ readings while
driving the AMR in Brown Hall.
passing the T-junction, it often settled into a low amplitude steady-state oscillation
while following the center of the hallway. The simple passage detection performed
very well, but due to the aforementioned problems with the 45◦ sonars detecting the
smooth walls the control was jittery (i.e. the steering jerked frequently).
Conversely, Controller 2 tended to follow the right wall more closely when the wall
on the left disappeared at the T-junction. After passing the T-junction, it tended
to over-correct but settled into following the center of the hallway steadily. Due to
the averaging of multiple sonars, as well as the more consistent performance of the
side-facing sonars, the control response was smoother than Controller 1.
4.2.3 Experimental Results - Edgar mine
Both of these controllers were tested on November 9, 2010 before the the main
proof-of-concept testing with the intention of using the better performer in the fully
autonomous tests. The proportional control of both wall-following controllers needed
to be tuned for the mine. In each case, the proportional gain had to be reduced.
However, the performance characteristics of these controllers seemed to switch in the
51
|
Colorado School of Mines
|
(a) Edgar mine Map
(b) Controller 1 (c) Controller 2
Figure 4.7: Plots showing approximate AMR path and sensor readings while navi-
gating passed the cross-cut. Only the sensor readings from the six side facing sensors
are shown for clarity.
mine. Controller 1, which showed steady-state oscillation in the Brown Hall hallway
showed little to no oscillation in the mine tunnels. On the other hand, Controller
2 showed a fair amount of steady-state oscillation in the straighter portions of the
mine tunnels. The performance of the controllers at the Y just after the cross-cut,
the locations of which are shown in Figure 4.7(a), changed.
The cross-cut was believed to be the most difficult location for the controllers
to navigate. However, both controllers managed to successfully navigate past this
location in testing (Figure 4.7). The Y further down actually proved to be more
difficult for both controllers. Controller 1 never successfully passed the Y in the
mine. Upon entering the widened area, the trajectory of the robot carried too far
to the right where it encountered the corner of the Y. Controller 2 also tended to
do the same thing, but after a number of adjustments managed to bear left enough
52
|
Colorado School of Mines
|
Figure 4.8: AMR navigating using Controller 1 in the tunnel after the Y (moving
from right to left). Only the sensor readings from the six side-facing sensors are
shown.
to pick up on the right wall again and continued through the tunnel. Unfortunately
after passing the cross-cut, Controller 2 would often begin oscillating at some distance
down the tunnel and would have to be stopped before hitting the wall (Figure 4.8).
It is important to note that during the proof-of-concept tests the base station
would command the AMR to stop around the Y. This could have been detrimental to
Controller 1, which used buffered sensor values to account for side passages. Because
thebufferingwasonlybasedontime, thebufferwouldfillwithreadingsfromthesame
location. This problem should not have affected Controller 2 which buffered previous
sensor values based on the distance traveled as opposed to time. Controller 2 had
one successful navigation past both the cross-cut and the Y as shown in Figure 4.9.
Unfortunately the radio signal strength information received from the base station
was incorrectly calibrated for this test run. This caused the AMR to move in small
increments of approximately 10 ft; thus, this particular test run was stopped short so
adjustments could be made at the base station.
53
|
Colorado School of Mines
|
Figure 4.9: AMR successfully navigating past the cross-cut and the Y in the Army
Tunnel.
Despite the difficulties encountered with these controllers, a number of successful
proof-of-concept tests were completed. Due to the difficulty of resetting the Bobcat
leader for each test and the controller’s unreliable performance near the Y, the proof-
of-concept test was divided into two parts. In the first part, the AMR started near
the base station while the Bobcat was driven further down the Army tunnel. The
AMR was unmanned and autonomously adjusted to maintain radio signal strength
between the Bobcat and the base station. Once the AMR reached the Y, it was
manually repositioned past the Y and the experiment continued. If the controller had
proved capable of autonomously navigating past the Y consistently then no human
intervention would be necessary for the entire duration of the test. Table 4.2 provides
a summary of the performance for each of the tested guidance controllers.
4.3 Sonars in the Mine
The performance of the ultrasonic range-finding sensors (sonars) within a mining
environment is very important to consider while developing a controller for the AMR.
54
|
Colorado School of Mines
|
Table 4.2: Tested Guidance Controllers Performance Summary
Controller Summary
Location Controller Performance Summary
Brown Hall Fuzzy Logic Successfully navigated T-junction, but is sensitive to
tuningparameters. Excellentperformanceinthehall-
way.
Controller 1 Successfully navigated T-junction, but had steady
state oscillation in the hallway.
Controller 2 SuccessfullynavigatedT-junction,nosteadystateos-
cillation in the hallway.
Edgar Mine Testing on Mar. 11, 2010, controller failed due to
Fuzzy Logic
broken sonar sensors.
TestingonApr. 2,2010,controllercouldnotnavigate
side passages, but performed well in straight sections.
Controller 1 No issues with cross-cut, but could not navigate
passed Y afterwards. No steady state oscillation ob-
served in straight tunnel sections.
Controller 2 No issues with cross-cut, successfully navigated
passed Y once. Steady state oscillations in straight
sections required human intervention to prevent
crashing.
The data gathered during the tests on April 2 and November 9 of 2010 reveal some
potential issues that should be addressed in future development of the MineSENTRY
robotic system.
Consider Figure 4.10 which shows the approximate path and sensor data for one
of the November 9th tests. The sonars’ results have been separated for clarity and
comparison. The plots for the 45◦ sonars indicate that, similarly to the results ob-
tained in Brown Hall, the incident angle may cause a large variation in the reported
distance. This result was unexpected because the walls within Edgar mine are far
from smooth. The rock walls of the Edgar mine are ragged and broken and deviations
from average wall position easily exceed 1 ft (≈ 300 mm). However, these errors only
appear in the forward facing sensors (sonars 5 and 6) while the rear facing sensors
(sonars 11 and 12) are actually quite stable (refer to Figure 4.11). The distances re-
ported by the front sonars, especially in the region −40 m ≤ x ≤ 0 m, report distance
55
|
Colorado School of Mines
|
Figure 4.11: 45◦ angled sonar sensor plots for Nov. 9th Edgar mine test.
readings up to three meters above the average value. The systematic positive error
seems to indicate that the ultrasonic sound wave is not being reflected back to the
sensor reliably. However, because the angled sensor in the rear are reporting reliably,
electric noise on the sensor cables could also be the cause.
The results of the side-facing sensors are far more reliable and, with some excep-
tions, perfectly acceptable considering the variation in the mine walls being sensed.
Please note that the readings on the far left side of the plot may have captured Mi-
neSENTRY teams members at the end of the test. The readings for the right side of
the robot (top, blue) have very little noise and in most regions all three sonars (4,
8, and 14) fall within an approximatley 0.5 meter range or less. These results are
acceptable considering the natural variation of the mine wall is of the same order of
magnitude. As shown in Figure 4.12, the right facing sonars produce very consis-
tent readings with occasional spikes on the order of 1-2 m. These spikes are most
likely caused by electrical noise and may be easily eliminated with software filtering.
Unfortunately, the dsPIC microcontroller only samples the sonar analog output once
per sonar update, thus increasing the susceptibility to electrical noise despite the low
57
|
Colorado School of Mines
|
Figure 4.12: Right side sonar sensor plots for Nov. 9th Edgar mine test.
pass RC filters on the signal lines.
The results for the left facing sonars (2, 10, and 13) are similar (Figure 4.13)
except for the large amount of noise observed on sonar 10. The tendency for sonar
10 to return lower values is most likely due to the sensor being pointed at a slight
downward angle. The MaxBotix LV-MaxSonar(cid:13)R-EZ3 sensors’ feature a beam pattern
that can detect objects within a ≈ 40 cm radius of the sensors center line. Thus a
slight downward angle may allow the sensor to detect rocks on the tunnel floor or the
tunnel floor itself. Furthermore, the tunnel cross section is somewhat rounded and
in many locations the width of the tunnel is narrower near the floor. It is expected
that a simple mechanical adjustment could greatly improve the reliability of sonar
10’s readings. If this is the case, the variation in the left facing sonars are acceptable
as well.
Theforwardandrearfacingsonarspresentsomeproblems. Formostofthetesting
the forward facing sonars (sonars 1 and 3) read a full 6 meters, but the rear facing
sonars (7 and 9) were highly noisy (Figure 4.14). Although electrical noise may be
present, the primary cause for the noisy readings is due to the sensing pattern for the
58
|
Colorado School of Mines
|
Figure 4.13: Left side sonar sensor plots for Nov. 9th Edgar mine test.
forward/rear facing sensors. The original design intent was to use the forward and
rear facing sensors for collision avoidance (i.e. to allow the vehicle to detect an object
in its path). To fulfill this requirement, sonars with a wider sensing pattern were
selected that could detect any object across the width of the AMR. Unfortunately,
the sensors selected have too broad a sensing pattern (≈ 1 m radius from sensor
center line) and often detect the floor or nearby walls. Tests within the hallways of
Brown Hall indicated the front/rear facing sensors could detect door frames or even
fire alarms on the walls. It would be beneficial to replace these sonars with the same
model used for the side-facing sonars and, if needed, use one broad range sonar on
the front and back of the vehicle for collision detection.
4.4 Summary
TofullytesttheconceptssetforthbytheMineSENTRYresearchproject,guidance
controllers for the AMR were designed and implemented. Although the controllers
operated as intended within building hallways, the controllers performed inconsis-
tently within Edgar mine’s rough, unordered tunnels. To a degree, each of the three
59
|
Colorado School of Mines
|
CHAPTER 5
PROPOSED CONTROLLER
The unreliable performance of the guidance controllers presented in Chapter 4
proved that a simple guidance controller is not sufficient for successful AMR nav-
igation. A more reliable controller able to navigate the more complex areas of the
mine is necessary. However, the simple wall-following controllers’ successes within the
straight sections of the mine should not be ignored. The guidance controller proposed
in this chapter extends the wall-following approach of the Chapter 4 controllers and
provides a unique approach to the navigation discussed in Chapter 2.
5.1 Controller Approach
This chapter describes the development of a guidance (or path-tracking) controller
that does not require highly accurate map data. Since active mines are typically
documented with structurally detailed survey maps, a controller of this type would be
very useful for quick deployment in an emergency situation. The controller proposed
in this chapter attempts to address these requirements with a unique approach to the
path-tracking problem.
Recent research focused on mobile robot navigation typically relies on an abun-
dance of sensor data to perform mapping and localization (SLAM). With accurate
mapandlocalizationdata,therobotcanplanapaththroughtwoorthree-dimensional
space and follow that path using a path-tracking controller. In Probabilistic Robotics,
localization is described as the process of referencing the robot’s position with respect
to an external reference frame [14]. This is the primary difference between current
techniquesandthepathtrackingcontrollerpresentedinthischapter. Tobemorepre-
cise, rather than referencing the map, path, and robot to a common reference frame,
the path and robot position are referenced directly to the environment in which it is
61
|
Colorado School of Mines
|
Figure 5.1: Example of the variable resolution path sampling for a path around a
simple right turn.
the environment. The purpose of the cost function is to eliminate path samples that
are indistinguishable from their nearest neighbors. In regions where the features
are unchanging (e.g. a long tunnel) only one path sample is required to accurately
describe that region. The design of the cost function necessarily depends on the
sensors, robot, and the intended operating environment. The reduced set of path
samples form a list of setpoints—so termed because they are the sensor range values
the controller is attempting to achieve. To be clear, setpoints are path samples that
are used to describe the path, whereas any individual path sample may or may not
be used.
5.2 Structured Approach Using Singular Value Decomposition
To utilize the structure of the environment for control purposes, a relationship
between the sensor reading, the environment, and the robot’s motion is necessary.
Consider Figure 5.2, in which the robot is following a predefined path and has not
deviated from the path. The robot’s distance to the wall is defined as d and the
w
orientation of the wall is given by φ. To determine the sensor reading, as explained
above, it is useful to place a coordinate system (reference frame) on the sensor with
the y-axis aligned with the sensing direction. This coordinate system is referred to
63
|
Colorado School of Mines
|
y
s
Wall
x
s
x
w
y
r
y
w
x
r
d
w
Figure 5.3: Reference frame placement for the robot, sensor, and wall.
To apply this to the case shown in Figure 5.2 a reference frame is placed on the
wall, the robot, and the sensor (Figure 5.3). It is important to note that all the
parameters given in Figure 5.2 are measured with respect to the robot’s reference
frame; therefore, it is convenient to find all the transformation matrices for mapping
to the robot reference frame and then use (5.3) to reverse the mapping. Thus we
have:
ST = ST R T (5.5)
W R W
To represent the wall in the sensor’s FOR, both a point and direction are required.
Although the point used is arbitrary, it is convenient to use the origin of the wall’s
reference frame which is obtained using (5.6).
x 0
y = S T 0 (5.6)
W
1 1
The direction is determined by using the rotation matrix S R. Because the x-axis
W
of the wall’s FOR lies parallel to the wall, the direction of the wall (as measured in
the sensor’s FOR) is found by:
(cid:20) (cid:21) (cid:20) (cid:21)
n 1
x,w = S R (5.7)
n W 0
y,w
65
|
Colorado School of Mines
|
sensor’s FOR is only dependent on the mounting position and orientation of the
sensor. The sensor reading when the robot deviates from the path is then:
d −s cos(δθ−φ)−δxcos(φ)+s sin(δθ−φ)−δysin(φ)
w x y
S = (5.10)
dM
cos(β +δθ−φ)
Now we must consider how this information is used to generate control outputs.
Following the approach outlined in Section 5.1, it is desired that the controller utilize
control setpoints, where each control setpoint is a list of the expected sensor values.
One possible approach is to use the difference between the expected, S , and mea-
dE
sured, S , sensor reading to determine the robot’s deviation from the defined path.
dM
Taking the difference between the expected and measured sensor reading yields:
c = cos(β)+cos(β −2φ)
a
c = sin(β)−sin(β −2φ)
b
c = 2(s sin(β)−s cos(β)−d sin(β −φ))
c x y w
c = 2d cos(β −φ)
d w
c = 1+cos(2(β −φ))
f
c = sin(2(β −φ))
g
c δx+c δy +c sin(δθ)+c cos(δθ)−c
a b c d d
S −S = (5.11)
dM dE
c sin(δθ)−c cos(δθ)
g f
At this point some interesting structure is noticeable. The coefficients c are de-
pendent on two things: sensor configuration and the path/environment description.
Now consider that S − S is equivalent to a standard control error signal (with
dM dE
opposite sign). The setpoint contains the expected sensor reading, S , when the
dE
robot follows the planned path. The measured sensor reading, S , are the sensor
dM
readings measured by the robot taken at some point which may or may not be on
the planned path. By setting S −S = e and rearranging (5.11) the following is
dM dE
obtained:
c δx+c δy +(c −c e)sin(δθ)+(c +c e)cos(δθ) = c (5.12)
a b c g d f d
67
|
Colorado School of Mines
|
The equation is formatted to better show the structure. Provided the path/envi-
ronment data is determined from map data, the only unknowns in the equation are
δx, δy, and δθ. Given that this equation is only for one sensor, it is apparent that
at least three sensors are necessary to resolve all the unknowns. However, this is
not necessarily a complete description as more than one wall is required for all three
variables to be solved. If only one wall is available only two variables can be com-
puted, δθ and either δx or δy. Often only δx will be determinant because the path is
typically parallel or near parallel to the wall (when only one wall or two parallel walls
are available). Note that these equations represent a wall as an infinite line, so any
bend or corner is minimally represented as two or more ‘walls’ within the equations.
Most frequently the robot will have more than two range-finding sensors and operate
within environments with at least two distinct walls. Thus, there is almost always
enough information to solve for these three unknowns.
To solve for these unknowns using all available sensor data, (5.12) is written for
each sensor. The corresponding system of equations may be presented in matrix form
as A(cid:48)x(cid:126)(cid:48) = b(cid:126)(cid:48):
A(cid:48) x(cid:126)(cid:48) = b(cid:126)(cid:48)
c a,1 c b,1 c h,1(e 2) c i,2(e 2) δx c d,1
c a,2 c b,2 c h,2(e 2) c i,2(e 2) δy c d,2
. . . . = . (5.13)
. . . . . . . . sin(δθ) . .
c c c (e ) c (e ) cos(δθ) c
a,n b,n h,n n i,n n d,n
c (e) = c −c e
h c g
c (e) = c +c e
i d f
noting that the number of sensors, n, must meet the requirement n ≥ 2. Methods
for computing the minimum least squares solution for problems of this type are well
known; however, this problem has non-linear constraints on the solution imposed by
cos(δθ) and sin(δθ). Rather than solve the non-linear problem, a straight forward
68
|
Colorado School of Mines
|
5.3 Algorithm Application
The equations developed in Section 5.2 are a mathematical description of the rela-
tionship between the sensor readings, robot pose (i.e. position and orientation), and
the environment with some simplifying assumptions. The equations were developed
with the intention of determining the pose of the robot, directly relative to the desired
location within the environment. Thus the following text will focus on applying these
equations to determine the robot’s pose and the efficacy of these methods for control
purposes.
5.3.1 Data Sources
Using equations presented in Section 5.2 to determine the robot’s pose requires
several data sources for each sensor: the sensor parameters (s , s , and β), the en-
x y
vironment data (d and φ), the expected sensor reading (S ), and the measured
w dE
sensor reading (S ). The sensor parameters are static for the AMR because each
dM
sonar is mounted in a fixed position and direction which is common for many mobile
robots. The variables describing the environment, d and φ, can be calculated from
w
setpoint, sensor, or map data. The expected sensor reading, S , is stored in the
dE
setpoint. Finally, the measured sensor reading, S , is provided by the AMR while
dM
it is navigating and sensing the environment. Of these data sources, it is prudent to
further discuss the environment variables due to the varied approaches to determining
their values.
Environment Data
Thevariablesusedtodescribeeachwalltherobotsenses,d andφ,arenotdirectly
w
observable nor is there any easy methods to measure them. As previously mentioned,
these values can be determined either from setpoint, sensor, or map data. Accord-
70
|
Colorado School of Mines
|
estimate/measurement of φ. By using live sensor data to compute φ we can elimi-
nate errors caused by discrepancies between the map and environment. The errors in
measuring φ from real-time sensor data is easier to quantify than the errors incurred
from the map.
Lastly, d and φ can be generated directly from the map data. In early versions
w
of the code, only the sensor data was determined from the map; but computing the
d and φ using the calcdwphi function is a rather indirect means of using the map to
w
determine the wall information. Thus the path sampling code also generates the val-
ues for d and φ and saves them with the setpoint information. As with the estimates
w
obtained from expected sensor values, this approach is susceptible to differences be-
tween the map and actual environment.
Position Estimation
AsdescribedinSection5.2, theangleδθ issampledbeforecomputingthedisplace-
ment values δx and δy with least squares minimization. How this angle is determined
is very important, because it can have a large effect on the accuracy of the position
estimate. In general, two techniques were considered for determining the angle. The
first technique is to sample a range of angles and uses the angle which yields the
smallest (cid:96) norm (5.16). Because sampling δθ requires the minimization problem to
2
be solved multiple times, this technique requires significantly more processing power.
To reduce the processing time required, the second technique determines δθ di-
rectly by comparing the φ values produced by the expected sensor readings (setpoint)
with the values obtained using real-time sensor readings. This is done by using the
calcdwphi function on both the expected and live sensor data. The result is two φ
estimates for each ‘wall’ the sensors see, one referenced from the desired position and
one from the robot’s current position. An estimate of δθ is obtained by subtracting
the observed (current) φ from the desired φ values and averaging the results.
72
|
Colorado School of Mines
|
Once an estimate of the robot’s pose (δx, δy, δθ) is known, a pose controller
is implemented to correct for the position error. The pose controller may be any
controller of suitable type (e.g. PID, fuzzy logic, et cetera). It is worth mentioning
that, from the perspective of the pose controller, the variables δx and δθ are the
error signals. The pose controller adjusts the robot’s speed and steering to drive
these error signals to zero. At this point it seems that the algorithm presented
in Section 5.2 is a type of localization algorithm, and in many ways it is similar.
But unlike localization algorithms, the robot’s position, the map, and the path are
not referenced to a common coordinate system. This algorithm localizes the robot
directly with respect to the desired path, which is itself referenced to the environment
(or walls) via the expected sensor readings. Thus this algorithm is a generalized wall-
following algorithm where the desired sensor values are provided, in a similar manner
to the wall-following controllers described in Chapter 4. Unlike those controllers,
the proposed algorithm does not distinguish between left, right, and center-following
techniques.
5.3.2 Algorithm Implementations
The different methods for calculating the environment variables and robot ori-
entation were developed while coding and testing the algorithm. For example, the
slow computation time associated with sampling the robot’s orientation compelled
the author to find a better means of determining δθ. This search resulted in the φ
comparison technique, which utilizes the calcdwphi function. When the speed of the
(cid:126) (cid:126)
codeimprovedremarkably, theauthorconsideredusingthed andφ(thevectornota-
w
tion denotes the d and φ for all range-finding sensors) computed from the live sensor
w
data instead of the map generated values. Thus, several different implementations of
the algorithm were tested with different results. Three primary implementations of
the algorithm were tested in depth and are described herein.
73
|
Colorado School of Mines
|
Figure 5.8: Simple turn simulation used for baseline comparisons.
the desired path. The pose controller used in each simulation uses a proportional-
integral-derivative (PID) controller which closes the loop around the angular (δθ) and
displacement (δx and δy) errors estimated by the algorithm. The pose controller was
initially tuned using the ‘true’ errors computed by the simulator. A proportional
controller reduces the simulated robot’s speed when large errors occur and completes
the pose controller. Because the estimated δθ affects the estimates of both δx and
δy, the different methods for determining δθ should be carefully analyzed.
5.4.1 Angle Estimation
Thereweretwomethodstestedforestimatingtheangleerror,δθ. Thefirstmethod
(cid:126)
samples different angles to determine which best minimizes ||A(cid:126)x−b|| . The second
2
(cid:126) (cid:126)
method compares φ and φ computed with the calcdwphi function. Figure 5.8
E M
shows the simulation used for the baseline comparison and Figure 5.9 shows the
angle estimated for the sampling and φ comparison methods, respectively.
From the graphs it can be seen that the φ comparison method produces more
accurate δθ estimates for the simulation shown. In contrast, the sampling method
underestimates the initial large displacement until the robot is within 15◦ of the
desired orientation. Within a range of ±15◦ degrees the sampling method yields
77
|
Colorado School of Mines
|
Sampling Based δθ Estimation
0.8
Actual δθ
0.6
Estimated δθ
0.4 b
0.2
a
0.0
0.2
c
0.4
0.6
0 5 10 15 20
Time (sec)
Theta Sampling at (a) Theta Sampling at (b) Theta Sampling at (c)
16 14 25
14 12 20
12 10
8 15
10 6 10
8 4
6 2 5
4 0 0
100 50 0 50 100 100 50 0 50 100 100 50 0 50 100
Angle (degrees) Angle (degrees) Angle (degrees)
Figure 5.10: Least squares minimization results shown at different locations.
reasonable results except for three large spikes. These spikes in δθ estimates coincide
with changes in sensor readings going in or out of saturation; in this case, the front
sensors pointing east and northeast (assuming the front of the robot is north). In
order to better understand why the theta sampling method yields these results, a
closer look at the least squares minimization at the sampled angles is beneficial.
As shown in Figure 5.10, the results of the LS minimization at each point provide
(cid:126)
a great deal of insight. At point a, the minimization of ||A(cid:126)x−b|| is more sensitive to
2
angles in the direction of the offset. This is likely due to the increase of the incident
angle for the majority of the sensors. It is also notable that the minimum is not
definitive, the cusp of the graph is much wider than the better estimation shown at
point b.
79
)m(
b
xA
2||
−
||
)snaidar(
elgnA
)m(
b
xA
2||
−
||
)m(
b
xA
2||
−
||
|
Colorado School of Mines
|
Theminimizationplotforpointb isdistinctandthesensitivityoftheminimization
is nearly equal across the range of angles. It has been observed that the sensitivity of
the plot on either side of the minimum is related to the sign of the minimum angle. As
the minimum angle changes, the v-shape of the graph tilts. Thus when the minimum
angle is positive, the minimization graph becomes less sensitive to angular changes in
the negative direction and more sensitive to angles in the positive direction. However,
this does not necessarily hold true in environments with more complex geometry.
The minimization plot at point c has two nearly equivalent minimums. The
appropriate theta estimate should be approximately zero degrees, and it should be
noted that the minimums appear to straddle this value. The minimization graphs
at these spikes all have multiple minimum values, and although there is typically a
distinct global minimum the theta estimate is poor regardless. These results indicate
that the solution with the lowest (cid:96) norm, with respect to the orientation, δθ, is not
2
necessarily the best solution.
On the other hand, the φ comparison method uses a weighted average which
is focused on reducing the effect of extreme δθ estimates. A non-weighted average
causes the δθ estimation to be sensitive to changes in sensor readings that result from
objects entering or exiting the sensor’s range. For example, Figure 5.11 shows the δθ
estimation results when direct averaging is used.
As shown in Figure 5.11, as the vehicle turns to the correct heading, the wall on
the left abruptly enters the range of sonar 5 (refer to Figure 5.12 for sonar positions).
Shortly thereafter, the wall on the right exits sonar 3’s range. Between these events,
a non-existent wall is assumed to be between these two sensors. This results in an
extreme φ estimates which skews the δθ average. The other eight sensors, however,
M
detect existing walls that yield coinciding angular estimates. By weighting the angle
estimates appropriately the δθ estimate is improved over direct averaging.
Now that the performance of the two δθ estimation methods has been discussed, a
80
|
Colorado School of Mines
|
S S
S 3 1 S
5 6
S S
2 4
S S
13 14
S S
10 8
S S
11 12
S S
7 9
Figure 5.12: Approximate Sonar Positions on the AMR
close look at the computation of δx and δy is required to further explain the proposed
controller’s process.
5.4.2 Estimating Translation
A path-tracking controller cannot function without knowledge of the translational
distance to the desired path. As defined in Section 5.2, the perpendicular distance
to the desired path is expressed as δx and the distance traveled along the path (with
respecttoapathsample)isrepresentedasδy. Aninherentproblemexistswhiletrying
to estimate δy within a tunnel environment which is known as the ‘infinite hallway’
problem. As viewed by the range-finding sensors, there are few if any distinct features
which can be used to determine progress along the path (δy). Because every point
along the hallway (tunnel) appears the same to the sensors, they are not sufficient to
determinedistancealongthepath. Duetothisproblem, manylocalizationalgorithms
depend heavily on dead-reckoning methods in these regions. An inherent feature of
the algorithm presented in this chapter is its ability to quantify the reliability of the
δy estimate.
82
|
Colorado School of Mines
|
TheSVDdecompositionofmatrixAyieldsthesingularitymatrixΣ, whichinthis
case is a 2x2 diagonal matrix containing the singular values of A. These two values
indicate how close the solution of equation 5.16 is to a singularity. Furthermore, each
of these singular values relates to how sensitive the solution is to errors within the
(cid:126)
vector b. Finally, the matrix V indicates how sensitive the estimates of δx and δy are
to σ and σ respectively. The closer σ or σ is to zero, the closer the solution is to
1 2 1 2
a singularity. For example, if the robot is traveling straight down an infinite hallway,
the matrices V and Σ will appear as:
(cid:20) (cid:21) (cid:20) (cid:21)
−1 0 σ 0
V = , Σ = 1 (5.18)
0 1 0 σ ≈ 0
2
where σ is a positive non-zero number. However, because the matrix V is a diagonal
1
matrix for this situation, the solution for δx is independent of σ , and thus decoupled
2
from the solution to δy, which is indeterminate (infinite solutions). The least squares
solution is determined by replacing the 1/σ place within Σ† with zero and only
2
solving for δx. By using the values within Σ and V, the ability of the algorithm to
yield stable estimates of δx and δy can be determined. Due to the intended use of
the algorithm for control within mine tunnels, only the value for δx will typically be
determinant.
Figure 5.13 shows the δx estimation corresponding to the δθ estimation shown
in Figure 5.9(a). In this example, δθ is estimated by sampling. The initial error of
approximately 200 mm is caused by the sampling method’s underestimation of δθ
initially. The first and last ‘spikes’ in the δx estimate are caused by the δθ estimate
spikesatapproximately12secand18sec,respectively. Thesmallerrorspikesbetween
12 sec and 18 sec are expected to be caused by a variety of error sources which will
be discussed in Section 5.4.3.
The primary δx estimation errors incurred using Method 1 are due to errors in
estimatingδθ; however, accurateδθ estimatesdonotguaranteeaccurateδxestimates.
83
|
Colorado School of Mines
|
Error in δx Estimation
300
200
100
0
100
200
300
0 5 10 15 20
Time (sec)
Error in δθ Estimation
0.8
0.6
0.4
0.2
0.0
0.2
0.4
0.6
0 5 10 15 20
Time (sec)
Figure 5.13: δx Estimation Error using Method 1
Method 2 produces the δx estimate shown in Figure 5.14. Although the δθ estimate
error is approximately zero for the first 12 seconds, the δx estimate varies between
±0.15 meters in the first four seconds. Initially the algorithm estimates the vehicle
is 0.1 meters right of the desired position, this indicates a bias towards the sensors
aiming towards the left of the vehicle, which are currently reporting larger distances
on average due to the vehicle pose. This behavior only occurs if the error in the
robot’s pose has an angular component. If the robot is displaced ±1 meter to the
right or to the left of desired, the δx estimate will be correct until the robot incurs
an angular displacement while trying to get back on path. The abrupt changes in the
error from positive to negative correspond to two of the sensors exiting saturation
(indicating a wall entered the sensors’ range). The first change corresponds to sensor
12 exiting saturation. This results in an imbalance between left/right facing sensors
being used by the algorithm, with one more right facing sensor than the left. The
algorithm’s bias shifts to the right facing sensors until sensor 5 exits saturation.
Method 3 utilizes the real-time sensor data to determine the values of d and φ.
w
84
)mm(
rorrE
)snaidar(
rorrE
|
Colorado School of Mines
|
Error in δx Estimation
400
300
200
100
0
100
200
0 5 10 15 20
Time (sec)
Error in δθ Estimation
0.2
0.1
0.0
0.1
0.2
0 5 10 15 20
Time (sec)
Figure 5.14: δx Estimation Error using Method 2
Thus, the robot estimates the structure from its current position then determines the
best action to drive the sensor readings towards the desired values (based upon the
map). Figure 5.15 shows the δx and δθ estimation error using this method in the
simple turn simulation.
It is interesting to note that results of Method 3 do appear better than the first
twomethods. Theinitialerrorisless, brieflyapproaching−0.15metersinthehallway
before the turn, and limited to a maximum of -0.25 meters during the turn. There is
a great deal of similarity in the δx estimate plots between Method 2 and 3; however,
the corresponding errors are attenuated using Method 3. As mentioned in Section
5.3, the output of Method 3 must be adjusted using (5.17) which incorporates both
δx and δy. It may seem unusual to incorporate δy in the calculation, because it was
indeterminate using the first two methods within the straight hallway section. One
might expect that the attenuation is due to δy being estimated as zero, resulting in
a loss of information. However in Method 3, V takes the form:
85
)mm(
rorrE
)snaidar(
rorrE
|
Colorado School of Mines
|
Sonar 4
6
5
4
3
2
1
0
0 5 10 15 20
Time (sec)
Sonar 14
6
5
4
3
2
1
0
0 5 10 15 20
Time (sec)
Figure 5.16: Simulated sensor readings for sonar 4 and 14 in the simple turn simula-
tion.
5.4.3 Error Sources
As was previously mentioned, there are a number of small estimation errors that
occur near the turn. This is because the algorithm was defined assuming straight line
paths and walls. Because the path around the turn is an arc, it must be approximated
by a greater number of path samples. The ‘true’ angular error computed by the
simulator is defined against the setpoint being used rather than the actual path. This
is why the plots are not continuous functions as one would expect—when the setpoint
changes the desired angle jumps. Another problem arises because the sensor values
change drastically within this region due to the low sampling rate.
Figure 5.16, shows the sensor values for sensors 4 and 14 (both facing right). Dur-
ing the turn the incident angle between these sensors and the wall changes quickly
resulting in large changes in the distance seen by the sensor. The simulation ac-
curately reflects the update rate and order of the ultrasonic sensors. New sensor
readings are taken every 50 ms with each sensor updating once every 200 ms. At
any one update, the algorithm is using anywhere between 10-12 old and 2-4 updated
87
)m(
egnaR
)m(
egnaR
|
Colorado School of Mines
|
sensor values. As seen in Figure 5.16, some of the sensor readings change as much as
1 meter between updates and regularly change as much as 0.1-0.2 meters while the
robot is turning (between 12-15 seconds). Considering the errors within the sensor
readings used by the algorithm, the error in the δx estimation, while still undesirable,
is not unreasonable.
Although each Method presented produced different estimation results, each was
able to successfully navigate the turn utilizing the pose controller described in Sec-
tion 5.4. The simple turn simulation was chosen specifically to display some of the
differences between Methods, but was not the only environment/path simulated. Ap-
pendix C.4 shows other environments used while developing the proposed controller.
In all of these simulations, the path was first sampled and specific path samples were
selected to be used as setpoints. How these setpoints are selected including the effect
on the performance of the proposed controller is an important factor that must be
addressed before a complete implementation can be achieved.
5.4.4 Setpoint Frequency
The Methods presented above have been compared using the accuracy of the δx
and δθ estimation within simulation. Because the ultimate goal of the algorithm
is to provide path-tracking data, it could be described as a type of feedback filter
within the control loop. Because a wide variety of control algorithms (e.g. PID, fuzzy
logic, et cetera) could be employed, the focus thus far has been to generate accurate
feedback data. However, all of the experiments thus far assume the algorithm is
being given the correct setpoints that define the path it is tracking—the role of the
localization algorithm. If the algorithm requires frequent and accurate setpoints to
perform, accurate localization becomes the primary barrier to implementation. If the
algorithm can perform with infrequent or inaccurate setpoints, the requirements of a
localization algorithm are greatly reduced.
88
|
Colorado School of Mines
|
Consider Figure 5.17 which compares the robot’s simulated path with three differ-
ent setpoints, each using the Method 2. The simulations in the first column use the
same number of setpoints as the previous simulations (n=19), the second shows n=7
setpoints, and the last shows n=1 setpoints (the robot is never provided the second
setpointshown). Interestinglytherobotinitiallyfails tonavigatethe turnusingseven
setpoints, but succeeds using only one setpoint. Figure 5.17(g) and Figure 5.17(h)
in the first two columns shows the same simulation with a slight modification: when
the robot reaches the turn, the algorithm is fed the next setpoint rather than the
current setpoint. By providing the robot some look-ahead information, the overshoot
is nearly eliminated in both the n=19 and n=7 simulations. There is little difference
between the path taken between the two simulations; however, the accuracy of the δx
and δθ estimations in the n=7 simulation suffers. This reflects the sparse setpoints’
inaccurate representation of the environment. The pose estimation for the n=1 simu-
lation is almost useless, only the region before and after the turn are accurate because
only these regions match the setpoint used. But, the ability to navigate the turn with
only one setpoint suggests the possibility to navigate with very infrequent setpoints.
Lastly consider Figure 5.18, which shows two simulations of a much more complex
environment: a hallway on the first level of the Brown Hall building on the Colorado
School of Mines campus. The first simulation uses a total of 89 setpoints and the
second simulation uses 16 setpoints (with look-ahead at the turn). In both cases the
robot successfully navigated the environment. In each case the robot will barely clip
the opposing wall at the turn due to overshoot, but the principle cause in this case
is the path and environment. The hallway itself is small compared to the AMR, and
the defined path leaves little room for error in either defining or following the path.
From these simulations it is clear that the algorithm can be tolerant to sparse,
and thus often inaccurate, setpoints if some adjustments are made to provide look-
ahead information in particular regions. The ability to navigate with sparse setpoints
89
|
Colorado School of Mines
|
benefits the MineSENTRY project. The lower the setpoint frequency, the smaller the
burden on the localization algorithm. Consequently, the map detail and computing
power required is also reduced. Lower-detail maps will be easier to update on-the-fly
which is likely to be necessary in a mine disaster situation. Furthermore, reduced
computing power allows the use of smaller, low-wattage processing technology. The
benefits of this are not immediately apparent unless the logistics of the MineSEN-
TRY system are considered. The deeper communications are extended into the mine
the greater the number of AMRs required. It would be difficult to transport and
launch a large number of golf-cart-sized robots. However, smaller robots can carry
less on-board power which also limits available processing power. The extent to which
the proposed algorithm can fulfill these statements will largely depend on the per-
formance of a full implementation. For example, the effect the setpoints have on the
performance of the simulated algorithm/controller requires further investigation. An
improved cost function for choosing setpoints could greatly improve the performance
of the proposed algorithm, but will require far more testing and analysis beyond the
scope of this thesis.
5.5 Summary
The algorithm presented is a unique approach to the path-tracking control prob-
lem due to the unique combination of the environment, path, and sensor information
into a single equation. The solution to this equation provides the tracking infor-
mation necessary for the robot to follow the path within a mine environment. Un-
like the guidance controllers of Chapter 4, the proposed algorithm is a generalized
wall-following approach that does not specifically distinguish between left, right, and
center-following techniques.
Several methods of implementing the proposed algorithm were tested in simula-
tion. The results indicate that the algorithm can provide accurate pose estimation
92
|
Colorado School of Mines
|
CHAPTER 6
CONCLUSIONS
The MineSENTRY research project set out to develop a multi-robot system that
solves the problems associated with wireless communication in mining environments.
To create this system, a single Autonomous Mobile Relay was created using a golf
cart as the vehicle platform. AMR construction required a number of subsystems to
be designed, built, and tested. Proof-of-concept testing, conducted at the Colorado
School of Mines’ Edgar mine, confirmed the effectiveness of the MineSENTRY design,
but the AMR guidance controllers were lacking. A unique controller was developed
as a proposed solution to the guidance problem. The controller was tested using
a variety of simulations. Although the simulations yield promising results, further
development is necessary before the controller can be implemented on the AMR.
6.1 Recommendations for Future Work
With respect to the AMR robotic platform detailed in Chapter 3, the modified
golf cart performed admirably well as a test platform. The primary recommendations
for improving the AMR system are itemized below:
• Improve protection from the environment
– Cover steering servo system.
– Paint/cover steel components.
– Seal electronic/sensor enclosures.
• Improve actuator assemblies
– Add easily accessible lever for steering disconnect.
– Replace brake actuator with higher power unit.
95
|
Colorado School of Mines
|
• Improve sonar reliability.
As shown in Chapter 4, the sonars can provide very reliable data within the mine.
It is expected rather simple adjustments could be made to greatly improve reliability.
The controller proposed in Chapter 5 requires some work before a full implemen-
tation can be realized. However, relatively little work is required before live exper-
imentation can begin. It is recommended that future work focus on the following
topics:
• Improving the pose estimation during turns.
• Improving sensor-based calculations for d and φ (calcdwphi function).
w
• Analyze the effect of setpoint frequency and an associated method for choosing
setpoints.
• Developing a method to choose setpoint during navigation.
• Testing how poor map data affects controller.
In particular, a means of choosing the correct setpoint is necessary for any large-
scale experiments. Small-scale experiments could be performed nearly immediately
for single setpoint paths, although some code alteration would be required before the
controller could be used outside of the simulator. As a final recommendation, efforts
to streamline and optimize the AMR control software would be very beneficial.
6.2 Implications of Research
The controller proposed in Chapter 5 has potential to provide mobile robot nav-
igation at a lower cost. By incorporating the environment structure into the con-
troller the need for accurate localization is reduced. In turn, this lowers the need
for expensive sensor systems and high performance computing. With respect to the
96
|
Colorado School of Mines
|
for the VC to fully enter autonomous control mode. The AC’s query packet requests
sensor data, and the command packet requests a control action as well as sensor data.
The VC responds to both the query and command packets with a sensor response
packet, which includes the sensor data requested. The emergency stop packets for
both controllers are unique.
If the AC sends an emergency stop (set E-stop) packet to the VC, the VC will
cut throttle and allow the AMR to coast. The VC will return a status packet in
response. The AC uses another emergency stop (release E-stop) packet to release the
E-stop condition. The VC’s emergency stop packet is the only packet the VC will
send without prompting. The VC sends an emergency response packet to the AC to
indicate that a mechanical E-stop on the AMR has been pressed. The VC reacts the
same way to both a software (AC) E-stop and a mechanical E-stop. The E-stop can
only be released by releasing the source of the E-stop (i.e. Mechanical E-stops cannot
release software E-stops or vice-versa).
The data is transferred via an RS-232 synchronous data stream with 8 bits, 1
start bit, 1 stop bit, and no parity at 57600 Baud. All data is represented with ASCII
characters – numbers are represented in hexadecimal format followed by the carriage
return character. Please note, the following tables are common to the MineSENTRY
project, and thus may appear in other MineSENTRY documentation and Theses.
111
|
Colorado School of Mines
|
ABSTRACT
Autonomous Haulage Systems (AHS) in mining began trials in the 1990’s but have only
become widely accepted in the last five years. Mine operators with AHS deployments are
reporting greater than 20% increases in utilization of trucks and reductions of hourly operating
costs as compared to manned fleets. Future deployments of AHS could be improved through the
implementation of a more complex and systematic approach to safety risk assessment than is
currently utilized within the mining industry. This dissertation describes the development of a
methodology for the safe selection and/or reconfiguration of an AHS’s Safety Instrumented
Systems, which follows the process prescribed by IEC 61508 but was adapted to the specific
situation of mining. The derived new methodology is accomplished by developing a generic fault
tree to define exposures and failure modes of an AHS. Representative scenarios simulating
various mining environments and situations are utilized to define the typical exposures in an
AHS deployment. Through the implementation of a systematic safety risk analysis specifically
tailored for AHS, acceptable levels of risk can be achieved while reducing undue process
interruption associated with safety systems. This novel analysis includes determination of the
maximum tolerable risk, likelihood of an event, and the required safety systems to mitigate those
risks. A theoretical determination of the safety ratings of the systems utilized in the commercial
deployments was undertaken by comparing the AHS deployments to autonomous automobiles.
Final analysis presents an idealized autonomous haul truck safety system to meet the demands of
the model mine presented in the functional safety analysis of an AHS.
iii
|
Colorado School of Mines
|
LIST OF ACRONYMS AND ABBREVIATIONS
ADAS Advanced Driver Assistance Systems
AEB Automatic Emergency Braking
AHS Autonomous Haulage System
AHT Autonomous Haulage Truck
ALARP As Low As Reasonably Possible
AOZ Autonomous Operating Zone
ASAM Autonomous or Semi-Autonomous Machine
ASAMS Autonomous or Semi-Autonomous Machine System
ASI Autonomous Solutions Inc.
ASIL Automotive Safety Integrity Level
BPCS Basic Process Control System
CAN Control Area Network
CDC Center for Disease Control
DMIRS Department of Mines, Industry Regulation and Safety
ECU Engine Control Unit
EMESRT Earth Moving Equipment Safety Round Table
EN European Standards
ESR Electronically Scanning Radar
Euro NCAP European New Car Assessment Programme
FTA Fault Tree Analysis
GMG Global Mining Guidelines Group
GNSS Global Navigation Satellite System
GPS Global Positioning System
HARA Hazard Analysis and Risk Assessment
HME Heavy mining equipment
IEC International Electrotechnical Commission
IMU Inertial Measurement Unit
IPL Independent Protection Layer
ISO International Organization of Standardization
LiDAR Light Detection and Ranging
LTE Long Term Evolution
LTI Lost Time Incident
LO Local Object
LV Light vehicle
MMM Manned Monitored Machines
MSHA Mine Safety and Health Administration
MTFR Maximum Tolerable Failure Rate
NHTSA National Highway Traffic Safety Administration
NIOSH National Institute for Occupational Safety and Health
ODS Object Detection System
OEM Original Equipment Manufacturer
OHS Occupational Health and Safety
PFD Probability of Failure on Demand
PL Protection Layer
ix
|
Colorado School of Mines
|
INTRODUCTION
Automation is an emerging key value driver within the mining industry. Automation was
used extensively within mineral processing and metallurgical functions prior to even the use of
solid-state electronics. However, Autonomous Haulage Systems (AHSs) only began field testing
in open pit mines in the early 1990s. Commercial adoption occurred nearly 20 years later in
2008. Technologies for vehicle autonomy advanced rapidly in recent years and effectively
moved the automation of mining fleets within reach of many mining operations.
Onboard systems of an Autonomous Haulage Truck (AHT) differ very little from any
autonomous vehicle, including a Google Car or Tesla’s system. An autonomous vehicle system
requires a vehicle control unit that takes in sensory perception, facilitates communication, and
outputs the appropriate vehicle controls to the actuators within the vehicle system. Typical
sensors utilized in autonomous vehicles include radio detection and ranging (Radar), light
detection and ranging (LiDAR), high precision Global Navigation Satellite System (GNSS),
cameras, odometers, and inertial monitoring units (IMUs).
Typically, actuators utilized in autonomous vehicles are already onboard a manned
vehicle. They are accessed via a Controller Area Network (CAN) bus, an internationally
accepted standard designed to allow microcontrollers and devices to communicate with each
other in applications without a host computer. Most modern on and off-road vehicles are
primarily drive-by-wire devices. They take inputs from a sensor that the operator/driver controls
and then electronically controls vehicle functions such as steering, throttle, and braking.
AHSs are more than trucks and onboard sensors. AHSs consists of: mining trucks
equipped with commercial and proprietary electronic devices; software to command, control, and
track vehicle movements and interactions; a communications network with complete coverage to
all active mining areas; high precision GNSS transponders; and a team of control room operators
and site support staff to manage vehicles, devices, and the network.
Integration of AHSs at a mine is a complex task that includes many of the safety hazards
of a conventional surface mine as well as the specific safety risks associated with AHSs. Prior to
1
|
Colorado School of Mines
|
implementation of an AHS, risk analysis and risk mitigation must be undertaken during
development of the system. Risk analysis and mitigation must continually occur during
commercial operation and improvement of AHSs.
Today, AHSs are operating safely and effectively within the Pilbara District of Western
Australia as well as at sites in Chile and Canada. Safety incident and accident data show very
few events and are primarily limited to human errors, such as failing to yield to traffic or failure
to follow operational procedures (Department of Mines and Petroleum 2017).
While AHSs show production improvements exceeding 30% over manned fleets
(Fotescue Metals Group 2016), reliance on the current safety instrumented systems (SIS)
introduces process interruptions that limit productivity improvements. For instance, the primary
engineering control of Caterpillar’s Site Awareness – the process where vehicles communicate
their bearing and velocity to all other vehicles – requires uninterrupted network communication.
This wireless connection is often difficult and expensive to maintain in a mining environment.
The principle wireless connection difficulty is the complex terrain associated with mining
that creates areas shadowed from wireless coverage. Process interruptions due to network
connectivity issues could be mitigated by restructuring SISs. Other process interruptions may
occur due to false detections from the object detection systems (ODS).
The current hierarchy of controls is unchanged from the first commercial introduction of
AHSs. Hierarchy of Controls is a system applied across many industries to categorize, minimize,
and/or eliminate hazard exposure (National Institute for Occupational Safety and Health 2016). It
is possible that the legacy system hierarchy could be restructured and SISs could be removed or
modified dependent on site specific conditions to reduce process delay impacts associated with
poor SIS performance while maintaining appropriate risk reduction factors.
An existing example of a system hierarchy reduction can already be seen when
comparing Caterpillar and Komatsu AHS trucks. Due to licensing constraints, the LiDAR
utilized by Caterpillar is not available for Komatsu to use in a mining application. Komatsu
determined their AHS trucks meet their internal safety requirements without the use of LiDAR
for object detection.
2
|
Colorado School of Mines
|
To safely reconfigure an AHS’s safety systems, the process prescribed by International
Electrotechnical Commission (IEC) 61508 Standard “Functional Safety of
Electrical/Electronic/Programmable Electronic Safety-Related Systems” can be applied to
analytically set safety integrity targets for a mining operation or portions of the operation. From
these targets, the appropriate safety systems can be selected based on the anticipated failure
modes from risk analysis.
Safety Integrity Level (SIL) rating methodology can then be utilized to ensure that safety
systems provide sufficient risk reduction factors to meet safety integrity targets. SIL is rated in
probability of failure on demand (PFD) and is expressed in terms such as 1 in 1,000. A SIL
rating is a ranking of a safety systems performance. The SIL rating system has four discrete
levels: SIL 1, SIL 2, SIL 3, and SIL 4. The higher the SIL rating, the higher the safety level and
the lower the PFD. This methodology is commonly used within highly automated industries,
such as automotive, transportation, medical, manufacturing, power generation, mineral
processing, and chemical refining.
Reduction or simplification of SISs within AHSs will not likely be driven by equipment
manufacturers and must be driven by the mining company. Original equipment manufacturers’
(OEM) business models are not driven by production improvements in the same way that a
mining company is. In fact, based on experience within the AHS sector, it is likely that reduction
of safety systems will be actively resisted by the OEM’s. A rigorous analysis of the SISs must be
managed by a mining company, with the support of the OEM, to drive future productivity gains
or implementations of AHSs into more complex mining environments.
1.1 Problem Statement
The expansion of worldwide AHS will require implementation of a more complex and
systematic approach to safety risk assessment than is currently available through the process
described by International Organization for Standardization (ISO) 31000 “Risk Management.”
The addition of AHS components to the haulage, loading, and ancillary fleets must be treated
with rigor to ensure safety as well as acceptable performance. Commercial developments of AHS
have taken the approach of continually adding layers of SISs to make AHS as safe as possible. It
is likely the combined SIL rating of those systems exceeds the demands of a mining
3
|
Colorado School of Mines
|
environment. It is also likely some SISs are unnecessarily redundant with limited additional
protection, yet these SISs add process interruptions due to false positives in the SIS. Through the
implementation of a systematic safety risk analysis specifically tailored for AHS, acceptable
levels of risk reduction can be achieved while reducing undue process interruption associated
with SISs.
1.2 Thesis Statement
IEC 61508 was developed during the transition of the process industry from human
control to computer control. This transition added significant complexity and potential risks that
could not be quantified through traditional risk assessment methods. IEC 61508 prescribed a
method to analyze the systematic risk of programmable safety related systems. IEC 61508 is
currently adapted through specific standards for process, automotive, railway, medical device,
manufacturing, and nuclear industries. In this dissertation, the process prescribed by IEC 61508
Standard is adapted to analytically set the maximum tolerable risk and safety integrity targets for
a mining operation.
1.3 Research Objectives
The primary research objective is the development of a methodology for the safe
selection and/or reconfiguration of an AHS’s SISs, which follows the process prescribed by IEC
61508. The derived new methodology is accomplished by developing a generic fault tree to
define exposures and failure modes of an AHS. Representative scenarios simulating various
mining environments and situations are utilized to define the typical exposures in an AHS
deployment. The methodology requires the following steps be taken:
1) Determine the maximum tolerable risk for a mine site.
2) Analyze the exposure of the typical roles within an AHS deployment. From these
exposures, safety integrity targets can be calculated for each scenario; and
3) From these targets, the appropriate SIS(s) can be selected based on anticipated failure
modes from risk analysis and provide sufficient risk reduction factors that meet the safety
integrity targets.
4
|
Colorado School of Mines
|
1.4 Research Methodology
The proposed research methodology is presented below.
1) Review of the current regulations and standards that may apply to AHS deployments.
2) Perform a risk analysis following the adapted IEC 61508 methodology of common
exposure within a mine site utilizing an AHS.
3) Risk analysis to be performed utilizing the fault tree analysis method.
4) Determine appropriate safety integrity targets to mitigate risks to reach maximum
tolerable failure rate as defined by IEC 61508.
5) Theoretically determine the limits of the SIL rating of existing SIS utilized in AHS.
6) Evaluate existing systems utilizing IEC 61508.
7) Develop of an idealized SISs for AHS.
The methodology developed through this research can be utilized by a mining company
to evaluate a mine or a portion of a mine prior to the development and deployment of an AHS by
an OEM to optimize both its safety and productivity based on site specific conditions and the
technology deployed by the OEM. The method can also be utilized on existing deployments to
improve productivity within an autonomous operating zone (AOZ) by reducing impacts from SIS
without negatively impacting the safety requirements of the site.
1.5 Originality of Work
While AHSs were trialed for over two decades, widespread adoption of AHS is just
beginning. Independent layers of protection were introduced by OEMs throughout the
commercial development of AHS but are rarely removed. The original contribution associated
with this research is the adaptation of a systematic approach to risk analysis utilized in other
industries and is now applied to AHS. While this approach is not new to other industries, it was
previously not developed for and currently is not publicly utilized within AHS. By conducting a
rigorous analysis of the SISs with the support of OEMs, a mining company can drive future
productivity gains or implementations of AHSs into more difficult mining environments.
5
|
Colorado School of Mines
|
1.6 Dissertation Organization
This dissertation is organized to provide the required AHS background prior to adapting
IEC 61508 to create a generic fault tree for an AHS installation and ultimately define a potential
alternate safety hierarchy for an AHS deployment.
Chapter 1: Introduction
Chapter 1 develops the problem statement, thesis statement, research objectives and
process of the research. A discussion of the originality of work is included.
Chapter 2: Regulations and Standards
Chapter 2 explores the existing regulations and standards utilized to govern AHS
deployments. Most AHS deployments are implemented under guidelines issued by the State of
Western Australia, but the recent deployments in Alberta, Canada are under actual regulations
issued by the Country’s Occupational Health and Safety department.
Chapter 3: Safety Integrity Targets of a Model Mine
Chapter 3 calculates the required safety integrity targets of a model mine through the
adapted IEC 61508 process. The maximum tolerable risk for an AOZ is determined. Calculations
of the exposure for various roles within an AOZ are utilized to calculate a sitewide exposure.
From these calculations, the safety integrity targets of the AOZ are determined.
Chapter 4: Theoretical SIL Ratings of Current AHS Deployments
Chapter 4 provides a theoretical calculation of the SIL ratings of the current AHS systems
again following IEC 61508. As actual data for each mine site’s systems remains proprietary, the
mining systems are compared to systems utilized in the autonomous automobile industry to
provide theoretical limits to the safety of the AHS deployments.
Chapter 5: Idealized AHT System Architecture
Chapter 5 defines an idealized AHT system architecture to meet safety integrity targets of
a model mine. The idealized system architecture looks to maintain appropriate levels of safety
while minimizing the potential for process interruptions due to false positives from the SISs.
6
|
Colorado School of Mines
|
REGULATIONS AND STANDARDS
Australia, Chile, and Canada are the only countries with commercial-scale AHS
deployments. Unlike the United States, mine safety in those countries are governed by state level
regulations. Western Australia has approximately 80% of the worldwide autonomous
deployments, but the Department of Mines, Industry Relations and Safety (DMIRS) only
provides a guideline for compliance rather than enforceable regulations. Alberta, Canada’s
Occupational Health and Safety (OHS) recently developed regulations are the first enforceable
regulations for AHS and are the first to reference international standards. Worldwide pressure
from industry groups like Global Mining Group (GMG) show the need for the adoption of
international standards such as Functional Safety.
With the roll-out of the Alberta OHS regulations, international standards are becoming an
enforceable document. Those standards already govern many of the aspects of the base haul
trucks. The addition of functional safety into the enforceable standards will enable a quantitative
or semi-quantitative process to evaluate a very complex system. Key definitions and processes
from within those standards are explored through their application on AHS.
2.1 Layers of Protection
Woven into an AHS operation are several safety systems or controls arranged in the
Hierarch of Controls format (National Institute for Occupational Safety and Health 2016). This
hierarchy concept assumes that the control methods at the top of Figure 2.1 are potentially more
effective and protective than those at the bottom which are more reliant on human behavior
rather than systematic controls.
In the case of AHS, elimination of risks associated with human error is one of the greatest
benefits. Through the removal of truck operators, the risk to those operators is eliminated.
However, loading unit operators, ancillary unit operators, drilling operators, and technical
services staff still remain within the AOZ. Safety impacts to the maintenance employees does
not change within the maintenance workshop. However, field repairs will expose them to AOZ
risks. Therefore, it is recommended by the DMIRS that additional primary controls, such as
8
|
Colorado School of Mines
|
elimination and substitution, be put in place to reduce risks to these employees (Department of
Mines and Petroleum 2014). Typical controls may include remove the risk by isolating or
providing alternative access for personnel not directly involved with the autonomous activity;
reduce the likelihood of the risk by restricting functions and activities to authorized personnel
and additional technology.
Figure 2.1 NIOSH Hierarchy of Controls (National Institute for Occupational Safety and Health
2016).
Engineering controls remain the second control tier to mitigate risk. Engineering controls
are arranged as layers of protection from most effective to least effective as perceived by system
designers. Current AHS implementations utilize the layers of protection in Figure 2.2. An AHS
provides protection via the object detection systems on each AHT, situational awareness that
provides positioning and orientation(pose) information of all machines within the AOZ, and the
ability to stop some or all the machines within an AOZ. If a hazard is not mitigated by site
procedures that eliminate the risk, layers of protection provided by AOZ access control systems,
situational awareness, object detection systems, and All-Stop systems should provide an
acceptable level of risk. If those layers do not provide sufficient risk reduction, administrative
controls such as site procedures or PPE in the form of the remote stop system should be added.
9
|
Colorado School of Mines
|
Figure 2.2. AHS Layers of Protection increasing left to right.
2.2 DMIRS Code of Practice
The Western Australia Mines Safety and Inspection Act 1994 (the Act) governs mining
within the state. A Code of Practice, as issued by the DMIRS, is a practical guide to achieving
the health and safety standards of the Act (Department of Mines and Petroleum 2014). The code
applies to those with a duty of care in the circumstances described in the code. Following the
code of practice, in most cases, would achieve compliance with the regulations. “Safe mobile
autonomous mining in Western Australia” (the Code) is designed to provide guidance on: mobile
autonomous and semi-autonomous systems used in surface and underground mines; developing
and evaluation of safe work procedures for those systems; the control of autonomous loaders,
trucks and other mobile equipment, such as drills and dozers; and identifying the unique risk
profiles in relation to mobile autonomous mining systems.
Rather than the prescriptive rules applied by the United States Mine Safety and Health
Administration (MSHA), the DMIRS utilizes risk management techniques as required by the
1994 Mines Safety and Inspections Act (State of Western Australia 1994). ISO 31000 defines
10
|
Colorado School of Mines
|
risk as the probability and consequence of occurrence of loss, injury, or illness. Analysis
prescribed under ISO 31000 analyzes the frequency of the risk, as well as the consequences of
the risk.
2.3 Alberta OHS Approval Process of AHS
In January of 2019, Alberta OHS issued a draft version of their approval process for
AHS. This standard was developed through cooperation with Western Australian AHS sites and
specifically calls out ISO 17757 “Earth-Moving Machinery and Mining - Autonomous and Semi-
Autonomous Machine System Safety.” The approval process is site specific and may vary based
on unique parameters of the mine site, technology, systems, and site operator’s experience with
AHS system.
Unlike the DMIRS Guide, the OHS process is a set of specific regulations rather than a
guide. Within these regulations are prescriptive requirements of international standards that must
be met. A required portion of the application is a compliance letter either from the manufacturer
or an Alberta Professional Engineer indicating that the applicable requirements of ISO 17757 are
met.
2.4 United States AHS Deployments Under MSHA
Currently, only one AHS trial is under the jurisdiction of the MSHA. MSHA does not
have any published regulations or guidelines for AHS. The current trial is regulated under an
Experimental Permit issued to Barrick in 2018 at their Arturo Joint Venture operation in Nevada.
One of the earliest Caterpillar trials occurred under MSHA jurisdiction at the Navajo Kayenta
Coal Mine in 2014.
2.5 Relevant International Standards
Standards are generally not mandatory in comparison to regulations unless that standard
is referenced in a regulation, such as ISO 17757 in the Alberta OHS Regulation. Additional
standards required for all manned operations will still be required for AHS sites. An example of
a required standard for manned sites that would also be required for an AHS is European
Standard (EN) ISO 12100 “Safety of Machinery – General Principles for Design – Risk
Assessment and Risk Reduction” which describes the requirements safe design principles.
11
|
Colorado School of Mines
|
ISO and IEC standards follow a hierarchical system, where the tip of the pyramid (Type
A standards) represents the basic design and analysis principles. Figure 2.3 is a graphical
representation of the hierarchy of standards that often is utilized by ISO and IEC, Figure 2.3. The
middle of the pyramid is built from group safety standards (Type B standards) that provide
standards for general safety aspects (Type B1 standards) and special protective devices (Type B2
standards). The base of the pyramid is built from specialist standards (Type C standards) that
specify safety features for individual machine families. Standards are developed by a committee
often made up of consortia of national and regional regulators, business within a marketplace,
and non-governmental agencies such as Australasian Institute of Mining and Metallurgy
(AusIMM), Society for Mining, Metallurgy and Exploration (SME). The definition of a standard
from ISO is a document established by consensus and approved by a recognized body, that
provides, for common and repeated use, rules, guidelines or characteristics for activities or their
results, aimed at the achievement of the optimum degree of order in a given context
(International Organization for Standardization 2020).
Figure 2.3. Hierarchical Organization of the IEC, ISO, EN Standards. Adpated from common
sources.
Table 2.1 lists many of the standards that would apply to an AHS and their
corresponding Hierarchical Organization Type classification. Each standard is relevant to risk
12
|
Colorado School of Mines
|
analysis required for the deployment of AHS at a mine site and should not be considered an
exhaustive list. Additional standards are also applied to manned and AHS sites.
Table 2.1 Relevant Standards.
Standard Title Type
ISO 31000 Risk Management Package A
IEC 60204 Safety of Machinery – Electrical Equipment of Machines B1
IEC 61025 Fault Tree Analysis A
IEC 61310 Safety of Machinery – Indication, Marking and Actuation B1
IEC 61508 Functional Safety of Electrical/Electronic/Programmable Electronic A
Safety-Related Systems
IEC 61511 Functional Safety – Safety Instrumented Systems for the Process B
Industry Sector
IEC 62061 Safety of Machinery – Functional Safety of Safety-Related Electrical, C
Electronic and Programmable Electronic Systems
ISO 9001 Quality Management Systems – Requirements A
ISO 12100 Safety of Machinery – General Principles for Design – Risk A
Assessment and Risk Reduction
ISO 13849 Safety of Machinery – Safety-Related Parts of Control Systems B1
ISO 14121 Safety of Machinery – Principles for Risk Assessment A
ISO 15998 Earth-Moving Machinery – Machine-Control Systems (MCS) Using B1
Electronic Components – Performance Criteria and Tests for
Functional Safety
ISO 16001 Earth-Moving Machinery – Object Detection Systems and Visibility C
Aids – Performance Requirements and Tests
ISO 17757 Earth-Moving Machinery and Mining – Autonomous and Semi- C
Autonomous Machine System Safety
ISO 19014 Earth-Moving Machinery – Functional Safety C
SO 20474-1 Earth-Moving Machinery – Safety – Part 1: General Requirements B1
ISO 21815 Earth-Moving Machinery – Collision Warning and Avoidance C
ISO 26262 Road Vehicles – Functional Safety A
2.6 ISO 17757: Autonomous Machine System Safety
The recently approved ISO 17757 formalizes the definitions and requirements of an
autonomous mining machine. This standard is explicitly referenced in the Alberta OHS
regulations for AHS. ISO 17757 requires a risk assessment process for Autonomous and Semi-
Autonomous Machine System (ASAMS), which conforms to the principals of ISO 12100. All
identified risks shall be mitigated to acceptable risk levels as part of the risk assessment process.
ISO 17757 also requires that safety-related parts of control systems shall comply with the
13
|
Colorado School of Mines
|
appropriate functional safety performance level. Examples include: ISO 13849 “Safety of
Machinery – Safety-Related Parts of Control Systems;” ISO 19014 “Earth-Moving Machinery –
Functional Safety;” IEC 62061 “Safety of Machinery – Functional Safety of Safety-Related
Electrical, Electronic and Programmable Electronic Systems;” or IEC 61508.
2.7 IEC Functional Safety Standards
Functional Safety is defined by IEC 61508 as “the safety parameters control systems
provide to an overall process or plant.” Functional safety standards were developed to improve
confidence in safety systems. This requirement is due to the complexity in modern safety
systems that are now predominantly electrical, electronic, or programmable systems. The process
plant environment bred many of the development standards used today in autonomous systems.
This new approach to reducing risk deviates from previous standards that were prescriptive in
nature by focusing on quantitative risk reduction. The functional safety approach greatly differs
from the prescriptive approach taken by MSHA regulators with the mandate of no allowable risk.
Functional safety engineering seeks to identify specific hazardous failures that can lead to
serious consequences. It then establishes a maximum tolerable frequency target for each mode of
failure. If the failure of a piece of equipment contributes to the identified hazards, it is referred to
as “safety related.” A “safety function” is further defined as the function of safety related
equipment which maintains a safe state or brings it to a safe state in response to a hazard.
Functional safety addresses both random hardware failures and systematic failures. The
former is often quantified and assessed in terms of failure rates while the later cannot be easily
quantified and must rely on conformance to acceptable development standards. Systematic
failures most often occur when software is involved in the safety related function.
Basic Process Control System
IEC 61511 “Functional safety - Safety instrumented systems for the process industry
sector” defines a basic process control system (BPCS) as the system that continuously monitors
and controls the process in a day-to-day plant operation. A BPCS is typically considered as an
independent protection layer (IPL) preceding the SIS.
14
|
Colorado School of Mines
|
IEC 61511 Part F.9 states the criteria to qualify a Protection Layer (PL) as an IPL are the
protection provided reduces the identified risk by a large amount, for instance 10-1, and the
protective function is provided with a high degree of availability (90% or greater). An IPL
should be designed solely to prevent or to mitigate the consequences of one potentially
hazardous event (i.e. a runaway reaction, release of toxic material, a loss of containment, or a
fire). Multiple causes may lead to the same hazardous event; therefore, multiple event scenarios
may initiate action of one IPL. An IPL must also be designed to facilitate regular validation of
the protective functions also known as auditability. Only those protection layers that meet the
tests of availability, specificity, independence, dependability, and auditability are classified as
independent protection layers.
Typically, a BPCS is a relatively weak IPL due to limited redundancy in components,
limited built-in testing capability, and limited security against unauthorized changes to the
internal programming. The latter of these may have the greatest impact to effectiveness due to
human error. IEC 61511 limits the combined PFD to not less than 1 × 10-1 for all the BPCS IPLs
that can be applied to a unique initiating event-consequence pair. However, some companies use
a PFD equal to 1 × 10-1 for each BPCS IPL if analysis indicates that the configuration,
maintenance, and regular testing of the BPCS ensures that each IPL BPCS is truly independent.
For the analysis presented in this study, the PFD of the BPCS is assumed to be 100 due to the
complexity of the system.
Safety Instrumented System
A safety system provides functional safety if the safety system – including logic solvers,
sensors, and actuators – achieves a tolerable level. IEC defines tolerable risk (International
Electrotechnical Commission 2016) as, “The aim of functional safety is to bring risk down to a
tolerable level and to reduce its negative impact; however, there is no such thing as zero risk.
Functional safety measures risk by how likely it is that a given event will occur and how severe it
would be; in other words: how much harm it could cause.”
A SIS is a system designed to prevent or mitigate hazardous events by taking the process
to a safe state when specific conditions are violated. A SIS is comprised of at least one Safety
Instrumented Function (SIF) composed of various combinations of logic solvers, sensors, and
15
|
Colorado School of Mines
|
actuators. SIFs must be functionally independent from the BPCS. Signals can be shared between
the SIF and BPCS only if the PFD is not affected. Typically, SISs utilize logic solvers with
multiple redundant processors, power supplies, signal paths, sensors, and actuators. Ideally, this
redundancy is achieved by more than one or duplicate components. An example of non-duplicate
redundant sensors is Caterpillar’s use of Radar and LiDAR in their perception system.
Additionally, each SIS will use self-diagnostics to detect and communicate sensor, logic solver,
and final control element faults.
Each SIF will have a SIL rating. A SIL rating is a ranking of a safety systems
performance. It is rated as PFD and can be expressed in terms such as 1 in 1,000. The SIL rating
system has four discrete levels: SIL 1, SIL 2, SIL 3, and SIL 4, as shown in Table 2.2.
Table 2.2: SIL Level, Risk Reduction Factor, Probability of Failure on Demand. Annotated from
(International Electrotechnical Commission 2016).
Risk Reduction
SIL Level PFD Rigor
Factor
SIL 4 100,000 to 10,000 10-5 to 10-4 State of the art and usually avoided.
Less rigorous than SIL 4, but still
SIL 3 10,000 to 1,000 10-4 to 10-3
requiring sophisticated techniques.
Requires good design and operating
SIL 2 1,000 to 100 10-3 to 10-2 practice such as in ISO 9001 “Quality
Management Systems – Requirements.”
The minimum level requiring good design
SIL 1 100 to 10 10-2 to 10-1
and practice.
The higher the SIL rating, the higher the safety level or the lower the probability of
failure on demand. Individual components do not have SIL ratings; instead, components are
determined to be suitable for different SIL environments. Use of components suitable to a SIL
rating does not ensure that a system has that SIL rating. The actual SIL rating is dependent on the
implementation of SIL appropriate logic solvers, sensors, and actuators.
The selection of a SIL rating for a SIS utilized in a process is based on risk analysis of the
situation. This risk analysis follows the same process as that described in the DMIRS’s code.
Risk analysis will result in a risk profile for the process. If that risk is too high, a risk reduction
and mitigation plan must be prepared that could include SISs. The SIL ratings of the SISs would
16
|
Colorado School of Mines
|
be dependent on the desired risk reduction factor. For example, a process industry company may
only accept SIS designs up to SIL 2. If the risk analysis indicated a requirement for a SIL 3 SIS,
a redesign of the process to lower the intrinsic risk would occur rather than implementing a SIL 3
SIS. This change in process would be during the elimination or substitution step of the hierarchy
of controls.
A typical SIS/SIF/SIL example case is a pressure vessel containing a flammable liquid. If
this pressure vessel’s process control system fails and allows the vessel to be subjected to an
overpressure condition, the SIS will act to prevent or mitigate the hazardous condition resulting
from the overpressure event. The SIS may include a pressure transducer, a logic solver to control
the system behavior, and a solenoid valve to vent the vessel into a safe location. If the risk
reduction factor required by the risk analysis is a factor of 100, then a SIL 2 level SIF would be
specified. This one SIF may constitute the entire SIS, or the SIS may be composed of multiple
SIFs to mitigate additional unacceptable process risks.
Human Independent Protection Layers
Human IPLs rely on operators to take action to prevent an undesired consequence in
response to an event. It is worth noting that most well-known major incidents, such as Three
Mile Island and Chernobyl, involved complex systems interacting with humans. The
effectiveness of humans in performing routine tasks under low or high stress conditions is
limited. Overall, human performance is less reliable than engineering controls. The Center for
Chemical Processes and Safety has extensively studied human failure rates. Routine tasks
typically have PFDs in the 10-2 to 10-3 range, while complicated non-routine tasks under timing
limitations have PFDs approaching 100 (Center for Chemical Process Safety 2001).
Maximum Tolerable Risk
Counter to many of the catch phrases used in the mine health and safety community,
there is no such thing as zero risk. Public perception and those involved in the health and safety
of mines often call for zero risk; however, this is not practical. Table 2.3 demonstrates the
everyday risk of death from various causes. Data is compiled from Center for Disease Control
(CDC), DMIRS, and MSHA.
17
|
Colorado School of Mines
|
Table 2.3: Probability of Risk of Death. Compiled from CDC, MSHA, and DMIRS data souces.
Cause Probability Source
All causes (mid-life including medical) 7 × 10-3 (Centers for
Disease Control
and Prevention
2020)
Unintentional Injuries 5 × 10-4 (Centers for
Disease Control
and Prevention
2020)
Traffic Accidents 1 × 10-4 (Centers for
Disease Control
and Prevention
2020)
Working at a USA Coal Mine (2018) 1 × 10-4 (Mine Safety and
Health
Adminstration
2019)
Work at a USA mine (2018) 8 × 10-5 (Mine Safety and
Health
Adminstration
2019)
Working at a Western Australia mine (2018) 3 × 10-5 (Department of
Mines, Industry
Regulation and
Safety 2019)
Natural disasters 6 × 10-6 (Centers for
Disease Control
and Prevention
2020)
The tolerable degree of risk varies based on degree of control of the circumstances, the
voluntary or involuntary nature of risk, and the number of persons at risk. However, 10-5 per
annum (pa) is often considered a broadly acceptable level of risk within mining and industrial
activities, such as oil & gas. Tolerable risks in the 10-6 range are typically only utilized for
scenarios that present a risk to communities outside of an operational area.
Maximum Tolerable Failure Rate
For simple systems, the maximum tolerable failure rate is calculated by dividing the
maximum tolerable risk by external IPLs. Typically, most of these IPLs reduce the exposure rate
18
|
Colorado School of Mines
|
of an individual. Reduction in staffing, AOZ access control, and site procedures are the reducing
factors specific to AHS. For example, factors that might be experienced at an AHS site are
shown in Table 2.4.
Table 2.4: Factors used to calculate maximum tolerable failure rate.
Factor Value Source
7000 operating hours/ 8760 total hours per
Portion of time system can offer risk 80%
year
2000 working hours/8760 total hours per
Person at risk 23%
year
Probability of Fatality from Incident 80% Judgement
Daily exposure to an AHT (<30m Estimated exposure based on operational
5%
separation) conditions
Based on Table 2.4, a requirement of a maximum 1.4 × 10-3 pa PFD is calculated in
Equation 2.1. The calculated PFD results in a safety integrity level requirement of SIL 3. This
means that a single SIL 3 safety system is required to meet the tolerable degree of risk for an
AHT. However, a single SIL 3 system is not a preferred solution due to the difficultly in creating
those individual systems; however, a combination of SIL 1 and SIL 2 systems can achieve the
risk reduction factor of a SIL 3 system. All current AHS implementations have numerous layers
of protection.
1 × 10 pa (2.1)
= 1.4 10 pa
(0.80 × 0.23 ×− 05.75 × 0.05)
−3
𝑥𝑥
As Low As Reasonably Possible (ALARP)
It is insufficient to end functional safety analysis at the determination of a SIL target
based on an acceptable level of risk. It is also necessary to establish if further improvements to
the system can be undertaken to reach as low as reasonably possible (ALARP) risk target.
Ideally, a system would reach a state of safety where the risk lies below the broadly acceptable
risk limit of 10-6.
19
|
Colorado School of Mines
|
An example of an ALARP calculation is presented below:
1) A $2,000,000.00 USD cost per life saved target is used in a particular industry;
2) A maximum tolerable risk target of 10-4 pa was set for a particular hazard that is
likely to cause two fatalities;
3) The proposed system was assessed and a predicted risk of 8 × 10-5 pa was
obtained. Given that the negligible risk is taken as 10-6 pa, the application of
ALARP is required;
4) For a cost of $6,000.00 United States Dollars (USD), additional instrumentation
and redundancy will reduce the risk to just above the negligible region (2 × 10-6
pa);
5) The plant life is 30 years;
6) A gross disproportionality factor of 10 is used;
7) The cost per life saved criteria is becomes 10 × $2,000,000 USD = $20 M USD;
8) The cost per life saved is calculated by the cost of the proposal divided by the
number of lives saved of the plant life, as follows: $6,000 [(8 × 10
−5
2 × 10 ) ×2 ×30] = $1.3 M USD
⁄ −
−6
Based on this calculation, the cost per saved life is less than the cost per life saved
criteria; therefore, the safety improvement should be implemented. The application of ALARP to
a mine model should evaluate more than just the cost of the system. The cost of additional IPLs
is negligible to the cost per human life; however, the process interruptions due to false positives
does have a significant impact. Typical operating costs of an AHT varies between $500/hr. and
$800/hr. (USD). Typically, process interruptions from a loss of communications or a false
detection of the ODS results in a chain reaction across the AHT fleet that shuts down the entire
fleet until the interruption is resolved. In a small fleet of 20 trucks, process interruption could
cost $10,000-$16,000 USD per hour. In more a common sized fleet of 50 trucks, process
interruption could cost $25,000-$40,000 USD per hour.
Meeting the Requirements of IEC 61508
Based on the Alberta OHS regulations on AHS and movement within the regulatory
environment of Western Australia, it is likely that AHSs will at some point be required to
21
|
Colorado School of Mines
|
demonstrate conformance to the requirements of IEC 61508. An active topic within the Global
Mining Group Guidelines focuses on the application of Functional Safety to AHS (Global
Mining Group 2019). Their problem statement, “The industry is not aligned. Available
international standards applicable for mine autonomy are not clearly defined and the
requirements for managing functional safety are therefore unclear,” demonstrates the lack of
clarity on the application of the standard. However, Alberta OHS has already included ISO
17757 in their regulations for AHS.
Certification under IEC 61508 takes two forms: certification of the organization creating
the product or system and certification of the system. It is not credible to expect certification of a
product without verification of the organization producing the product. Many of the companies
associated with AHS cleared or could clear the burden of ISO 9001 “Quality Management
Systems.” However, the situation with IEC 61508 is different and less well developed. Self-
certification is not precluded in IEC 61508.
IEC 61508 includes a provision for a system to be “Proven in Use.” As an alternative to
the rigorous process of Parts 1 and 2, statistical data from field applications may be used to
satisfy IEC 61508. Due to their large number of trucks, Caterpillar and Komatsu could argue that
their AHSs achieved a ‘proven in use’ demonstration.
In November 2018, Caterpillar announced they hauled one billion tonnes and traveled 35
million kilometers under AHS (Caterpillar 2018). Assuming the average speed of their trucks is
20 km/hr, the result is 1.75 million hours of operation. AHS trucks operate approximately 7,000
hours per year, which means that Caterpillar has approximately 250 years of operation without a
lost time incident (LTI) or fatality. Because truck numbers are climbing and a majority of
operating hours occurred in the last few years, it is likely that Caterpillar could claim a SIL 3
rating within the next couple of years under the “Proven in Use” path.
While many of the requirements to meet IEC 61508 are the responsibility of the
manufacturer, the implementation strategy and supporting documentation are critical to the
success of the deployment as well as the potential certification of the system. Implementation
philosophies, restrictions, and limitations will contribute to an effective and practical
implementation of AHS and are also critical in the certification of a system under IEC 61508. An
22
|
Colorado School of Mines
|
adaptation of a Functional Safety “V” diagram presented in ISO 26262 “Road Vehicles
Functional Safety” shows the stages of the Functional Safety process that should be the sole or
shared responsibility of the mine site (Figure 2.5).
2.8 IEC 61025: Fault Tree Analysis (FTA)
The more complex scenario of an AHT is better represented by a fault tree. Fault Tree
Analysis (FTA) is a deductive failure analysis that uses Boolean logic to combine a series of
lower level events. FTA is commonly used in high hazard industries, such as aerospace, nuclear
power, chemical process, pharmaceutical, and petrochemical. IEC 61025 “Fault Tree Analysis”
is the international standard that covers FTA.
The undesired event is taken as the top event of the logic tree. FTA works backwards
from this event to determine the causes of the event and failures of the layers of protection of the
system. FTA assumes failure of the functionality of a product or process and then identifies all
potential root causes of the assumed failure. Each system or subsystem failure is evaluated one at
a time and can combine multiple causes by identifying causal chains. Results of the FTA are
schematically represented in the form of a tree of fault nodes. At each level of the tree,
combinations of fault modes are described by logical Boolean operators (AND, OR, etc.). Basic
symbols are shown and described in Figure 2.6.
23
|
Colorado School of Mines
|
( ) = ( B) = P(A) × P(B) (2.4)
An OR gate𝑃𝑃 co𝐴𝐴rr𝑎𝑎es𝑎𝑎p𝑎𝑎on𝐵𝐵ds to 𝑃𝑃a s𝐴𝐴et∩ union:
( ) = ( B) = P(A) + P(B) - ( B) (2.5)
𝑃𝑃As𝐴𝐴 th𝑜𝑜e𝑜𝑜 f𝐵𝐵ailure𝑃𝑃 pr𝐴𝐴ob∪abilities used in FTA𝑃𝑃 ten𝐴𝐴d∩ to be less than 0.1, ( B) becomes a
very small term and the output can be conservatively approximated by using an assumption that
𝑃𝑃 𝐴𝐴∩
the inputs are mutually exclusives events, therefore:
( ) = ( B) = P(A) + P(B) (2.6)
Figur𝑃𝑃e 2𝐴𝐴.7𝑜𝑜 r𝑜𝑜ep𝐵𝐵resen𝑃𝑃ts 𝐴𝐴a g∪eneric single event fault tree and layers of protection of an AHT.
The layers of protection of an AHT are broken out to the terms defined by ISO 17757. In
addition to the safety systems identified by ISO 17757, BPCS from the pose system was added.
The pose system provides a double redundant process to verify location via dual antennas,
odometry, and an IMU. While this would not qualify as a SIS, it certainly provides a layer of
protection. A generic event is also represented within the diagram.
2.9 Earth Moving Equipment Safety Round Table
The Earth Moving Equipment Safety Round Table (EMESRT) was formed in 2005 and
was driven by the desire to fill the knowledge gap between customers and OEMs with regard to
understanding operational and maintenance risks. EMESRT is a global initiative involving major
mining companies that made valuable risk evaluation frameworks for vehicle interaction systems
that are relevant to the implementation of AHS.
The Vehicle Interaction Systems document is a performance requirement developed to
augment the interpretation of the EMESRT design philosophy and machine operation controls
within various potential unwanted event scenarios. Performance requirement objectives serve to
prevent a person or equipment (machine or vehicle) event resulting in injury or equipment
damage including equipment to person, equipment to equipment, equipment to environment, and
loss of control of equipment.
26
|
Colorado School of Mines
|
SAFETY INTEGRITY TARGETS OF A MODEL MINE
By rigorously defining safety integrity targets of an AOZ, the safety requirements of the
AHS and AHTs can be defined. Those requirements can be utilized to select the appropriate
SISs to provide sufficient risk reduction without burdening the system with redundant or
unnecessarily complicated systems. The process prescribed in IEC 61508 has been adapted to
meet the demands of an AOZ. The first step described in IEC 61508 is to analytically set a safety
integrity target for the model mine. This target will generally align with the industry standard of
10-5 fatalities per annum. The maximum tolerable failure rate is then determined by applying the
risk reductions associated with external IPLs. The primary external IPL is the limited exposure
that all personnel within the AOZ experience. Exposure values are developed by evaluating the
risks for various roles within the AOZ. The resultant maximum tolerable failure rate is then
utilized to calculate the required SIL for the AHT safety system(s).
3.1 Risk in the AOZ
A generic fault tree, Figure 2.7, is divided in half with one side representing the internal
layers of protection while the other half focuses on the exposure to fault conditions or conditions
that could result in a fatality in an AOZ. The latter half of the fault tree sets the integrity target of
functional safety analysis which is the focus of IEC 61508 Part 1. Fault events are evaluated
relative to the risk they present to an individual. This allows for evaluation of the fault risk by
role within the AOZ. The fatality risk associated with AHS is limited to AHT-to-MMM
collisions in this study. There is some minimal risk associated with AHS and personnel not in a
vehicle; however, the rates of exposure are likely one to two orders of magnitude lower than
vehicle collisions due to site procedures and limited permitted activities outside of a vehicle.
Risk of collision only occurs when an MMM and an AHT are near each other (<30 m) and in a
position for the AHT to collide with a manned machine. The EMESRT surface vehicle
interaction scenarios provide a set of vehicle interactions that may result in an incident (Figure
2.8).
Potential interactions anticipated in an AOZ are show in Table 3.1. Interactions are
limited to plausible scenarios that occur or have the potential to occur in an AOZ. Four
31
|
Colorado School of Mines
|
interaction scenarios are anticipated to occur: oncoming, dovetailing, intersections, and HME-
AHT interactions during work. Oncoming interactions occur when an AHT passes an MMM in
the opposite direction in an adjacent lane. Dovetailing interactions occur within the same lane
and are limited to AHTs approaching too closely from the rear of an MMM. Intersection
interactions occur when an MMM fails to comply with the hierarchy of vehicles utilized within
the AOZ. The hierarchy of vehicles typically means that the larger/heavier vehicle has the right
of way. When this procedure is utilized within an AOZ, all AHTs are located at the top of the
hierarchy regardless of their actual size. The last interaction type occurs while an HME is
performing its function. This interaction is limited to the loading units, dozers, and motor
graders. Scenarios are compared to the EMESRT list and include the anticipated machine type
that would be exposed to the scenario. Primarily exposures occur within the static lanes of the
AOZ, such as the haul roads and ramps. Dynamic lane zones (loading or dumping) are typically
off-limits to manned machines with the exception of loading units and dozers. This exclusion
removes many of the risks associated with the AHTs and for all MMM other than loading units
and dozers.
Table 3.1: Interaction Scenarios in an AOZ.
Machine Type Exposed EMESRT
Scenario Definition
to Scenario Scenario
Oncoming Oncoming traffic in an All L1, L5, L8, C1,
adjacent lane V4
Dovetailing Traffic within the All L4, C2, V4
same lane approaching
from the rear
Intersections MMM failing to All T1, T2, T3, T4,
comply with hierarchy V4
of vehicles at an
intersection
HME-AHT HME working within Loading Unit, Dozer, L2, L3, L6, L7,
their operating area. Motor Grader C3, V4
3.2 Maximum Tolerable Risk
The maximum tolerable risk is the value a mining company should determine as part of
their corporate policy. Heavy industry, which includes mining, generally sets the maximum
32
|
Colorado School of Mines
|
• clearing/confirming AHT object detections;
• introduction to manned machine systems;
• safe work procedures in the AOZ (loading, dumping, maintenance, etc.); and
• manual operation of an AHT.
Typical training to reach this level of competence takes multiple weeks. Training reduces
exposure to AHT hazards by first reducing the number of people allowed within the AOZ and
secondly, by allowing only highly training individuals within the AOZ. Typically, an AOZ sees a
significant reduction in staffing in comparison to a manned site. This reduction is not only
limited to truck operators but includes a reduction in non-critical employees within the AOZ as
compared to a manned site.
A calculation of Maximum Tolerable Failure Rate (MTFR) will be calculated only for the
classes of workers expected to be within the AOZ. Each class has a calculated exposure time
based on their role and responsibility. The following classes will be investigated for specific
risks, exposure time, and consequences:
• AHS Pit Supervisor;
• Loading Unit Operator;
• Motor Grader Operator;
• Water Cart/Truck Operator;
• Dozer Operator (Dump);
• Dozer Operator (Loading Unit Clean-up);
• AHS Pit Technician & Mechanic/Fitter;
• AHS Service Technician and Wi-Fi Service Technician; and
• Serviceman (Fuel bay).
There are other roles that occur within an AOZ; however, their exposure is considerably
lower than the roles listed above and might include mine services, technical services, and other
support functions. These roles are generally escorted within the AOZ or allowed only to operate
within exclusions zones within the AOZ.
34
|
Colorado School of Mines
|
Autonomous and manned mines have operating procedures in place to limit human
exposure to large mobile equipment. A common term for this type of rule is the 50/30 rule and is
enforced when a LV or a person on foot intends to approach a piece of heavy mining equipment
(HME). Prior to entering within 50 meters of an HME, personnel must establish positive two-
way communication with the HME operator to indicate their request to enter within 50 meters of
the HME. The LV or person must wait until they are granted permission to approach by the
HME operator. Other than where segregated by a physical barrier and prior to entering within 30
meters of an HME, positive two-way communication must be established, the HME must be
parked, all ground engaging tools are grounded, and the operator’s hands must be away from the
controls of the HME. Similar procedures exist for autonomous vehicles but require positive two-
way communication with the control room and direct evidence of the change in the AHT’s
operating state as shown by its signal lights.
A few large mining companies have implemented LV only roads to remove the potential
for interactions between LVs and HME. This adoption obviously has a significant impact on the
exposure of LV occupants to the risk of AHS. Even without separated lanes, the interactions
around AHTs are heavily controlled. Static lanes have more limited Safe Work Procedures
(SWPs) compared to the dynamic areas due to their increased complexity and risks.
A major contributor to external IPLs at a mine site is SWPs. SWPs originated in Australia
and now serve as risk management protocol throughout multiple industries within the country.
Outside Australia, SWP are often referred to as Standard Operating Procedures (SOP). A SWP is
a step-by-step description of a process or task when deviation from the task could result in a loss,
i.e. damage to equipment, injury, or fatality. The risk control document is unique and is created
by teams within a mine site to describe the safest and most efficient way to perform a task.
3.4 Cause of Event
Human error stands out as the most probable cause of an incident within a manned or
AHS mine. If loss of control events are excluded from the DMIRS AHS incident data, all
reported incidents are a result of human error alone and sometimes accompany a mechanical
failure (Department of Mines, Industry Regulation and Safety 2019). DMIRS data indicates that
notifiable incidents occur at a rate of 25-45 incidents per million hours worked. Assuming that
35
|
Colorado School of Mines
|
an average miner works 2000 hours per annum, the average of the DMIRS range converts to an
annual incident rate of 7.0×10-2 pa.
The Center for Chemical Process and Safety recommends a range of 10-1 to 10-3 per
opportunity with the most commonly used value of 10-2 for analysis (Center for Chemical
Process Safety 2001). Data used in the fault tree analysis is in units of per annum. Conversion
from per opportunity to per annum requires that an estimate of the number of times per year that
a demand is placed on the operator. The number of incidents requiring human intervention is
difficult to predict; however, the combination of annual incident rate and the rate of human PFD
can be represented as the annual incident rate. Therefore, 7.0×10-2 pa was utilized for the fault
tree analysis presented here.
While no metal on metal incident was recorded as a result of a loss of control incidents, it
should be included in the fault tree analysis as a potential incident type. Generally, DMIRS loss
of control incidents are related to a loss of traction due to wet roads. In all incidents, the BPCS
brought the vehicle to safe stop. In some cases, the stop occurred outside the assigned travel lane
but within a few meters of the travel lane. From 2017-2019, an average of 2.3 loss of control
incidents occurred within AHSs of Western Australia. During that period, more than 200 AHTs
were in operation within the Pilbara. Based on these assumptions, the annual incident rater per
AHT is approximately 1.1×10-2 pa. This value was utilized for the fault tree analysis.
While no incident related to a failure of the BPCS has been reported, there is still a
potential for a fatality related to a BPCS failure. IEC 61511 limits the combined PFD to not less
than 1 × 10-1 pa for all the BPCS IPLs that can be applied to a unique initiating event-
consequence pair. However, some companies use a PFD of 1 × 10-1 for each BPCS IPL only if
analysis indicates that the configuration, maintenance, and regular testing of the BPCS ensures
that each IPL BPCS is truly independent. Therefore, 1.0×10-1 pa was utilized for the fault tree
analysis.
3.5 Definition of the Model AOZ
A model AOZ has been developed following the path described in IEC 61508. First the
maximum tolerable risk of the AOZ is defined utilizing industry norms. A simplified scheme of a
fleet comprised of sub-fleets is then imagined representing a typical mine. This scheme is
36
|
Colorado School of Mines
|
utilized to determine cause and probability of a failure/fatality. The predominant driver of risk
within the AOZ is the exposure time of an individual to the AHTs. These exposures occur in two
prime scenarios: brief passing within travel lanes and longer exposures at loading and dumping
points. While the loading and dumping points have much longer exposure intervals, their
probability of fatality due to collision is much lower than interactions between light vehicles and
AHTs within the travel lanes. The risks to individuals are combined into a resultant factor, that
when combined with the maximum tolerable risk, represents the SIL requirements of the safety
system(s).
3.6 Model Assumptions
To evaluate the exposure rates within an AOZ, the ideal solution would be time-in-
motion studies of an active AOZ or a simulation of a complex AOZ within traffic modeling
software. An ideal study would generate an average exposure time where a manned monitored
machine is within 30 meters of an AHT and able to collide with that machine. Unfortunately, that
data and process are not readily available; but a simple mine model can be developed to assess a
typical AOZ. To simplify the calculation of exposure rate, it is assumed that sub-fleets do not
interact with each other. A single sub-fleet is represented in Table 3.2. It should be noted that
half units are shared between sub-fleets. The result of this assumption is that exposures for motor
grader, water cart/truck, and service truck operators assumes full utilization, but the machine is
shared across multiple sub-fleets with similar exposures. If multiple sub-fleets are considered,
the impact from their addition will vary from zero if the sub fleets do not interact within the same
haulage cycle to a multiple of the sub-fleets using the entirely the same haulage cycle.
37
|
Colorado School of Mines
|
Table 3.2: Model AOZ Sub-Fleet.
Machine Type Units
Haul Truck 5
Shovel/Excavator 1
Dozer 1
Wheel Loader/Dozer 1
Motor Grader 0.5
Water Truck/Cart 0.5
Service Truck 0.5
LV (Supervisor) 1
LV (Pit Tech) 2
LV (Wi-Fi Tech) 1
A simplified model of a waste haul route from a loading unit to a dump is utilized to
define the haul path of an AHT (Figure 3.2). The model mine assumes that an AHT is loaded
and then travels out of the loading zone. The AHT then progresses along the pit floor, up the pit
ramp, across a surface haul road, up a dump ramp, across the dump bench, and into the dump
zone. The truck then dumps its load and returns along the same path to the loading unit. The
model only assumes waste haulage; however, the risk profile of a haul to a run-of-mine stockpile
or crusher is a reduction in exposure due to typically shorter distance traveled and the removal of
a dozer within the dump zone.
Dump
Zone
Empty AHT
Dump
Bench
Surface
Road
Loading
Zone
Full AHT
Pit Floor
38
|
Colorado School of Mines
|
Table 3.3: Model AOZ Segments.
Segment Length (m) Speed (kph) Opposing Speed (kph) Duration(s)
Load NA NA NA 240
Loading Area 200 30 30 24
Pit Floor 300 50 50 22
Pit Ramp 1,000 15 30 240
Surface Road 500 50 50 36
Dump Ramp 300 15 30 72
Dump Bench 300 50 50 22
Dump Area 200 30 30 24
Dump NA NA NA 180
Dump Area 200 30 30 24
Dump Bench 300 50 50 22
Dump Ramp 300 30 15 36
Surface Road 500 50 50 36
Pit Ramp 1,000 30 15 120
Pit Floor 300 50 50 22
Loading Area 200 30 30 24
AHT Frequency is utilized to calculate the number of interactions between an AHT and
an MMM. Three types of AHT-MMM exposures occur within the model: oncoming traffic,
dovetail traffic, and AHT-MMM working interactions. The latter type of exposure is modeled in
two ways: loading or dozing and tramming or grading. Loading or dozing is modeled by
multiplying the AHT frequency, the total shifts that an operator is onsite in a year, length of shift,
and the time that a machine is within 30 meters until it comes to a stop. The exposure from an
AHT backing towards a RO is then reduced by the risk of fatality and portion of a shift that an
operator is present in the machine.
=
𝐹𝐹𝑎𝑎𝜆𝜆𝑎𝑎𝐶𝐶𝑇𝑇𝜆𝜆𝐶𝐶𝑊𝑊𝑒𝑒𝑇𝑇𝑆𝑆ℎ𝜆𝜆𝑒𝑒𝑎𝑎𝐸𝐸𝑥𝑥𝐸𝐸𝑜𝑜𝑚𝑚𝐸𝐸𝑜𝑜𝑒𝑒𝐿𝐿𝐷𝐷𝐷𝐷𝑆𝑆𝐷𝐷𝐷𝐷𝐿𝐿𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐿𝐿
× × ×
(3.5)
×
𝐴𝐴𝐴𝐴𝑇𝑇𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝑚𝑚𝑇𝑇𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝑈𝑈𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝐷𝐷𝐿𝐿𝑆𝑆𝑇𝑇𝐷𝐷𝑆𝑆𝐿𝐿 𝑇𝑇𝑇𝑇𝑆𝑆𝑒𝑒𝐵𝐵𝐷𝐷𝐹𝐹𝐵𝐵𝐷𝐷𝑆𝑆
𝑅𝑅𝑇𝑇𝑚𝑚𝑅𝑅𝐹𝐹𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝜆𝜆𝐹𝐹 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝐿𝐿𝐿𝐿𝐷𝐷𝐿𝐿𝜆𝜆ℎ
Tramming of loading units and dozers is handled in the same way that a motor grader
working within a segment is handled. By multiplying the frequency of AHTs and the time spent
within the haulage path of the AHTs, a total time of exposure can be calculated. The buildup of
40
|
Colorado School of Mines
|
the exposure is nearly identical to the previous equation with the backup time replaced by an
average exposure time within the haulage cycle. Average exposure time is a weighted average of
the AHS exposure by segment times. The average is weighted by the segment duration.
=
𝐹𝐹𝑎𝑎𝜆𝜆𝑎𝑎𝐶𝐶𝑇𝑇𝜆𝜆𝐶𝐶𝑊𝑊𝑒𝑒𝑇𝑇𝑆𝑆ℎ𝜆𝜆𝑒𝑒𝑎𝑎𝐸𝐸𝑥𝑥𝐸𝐸𝑜𝑜𝑚𝑚𝐸𝐸𝑜𝑜𝑒𝑒𝑇𝑇𝐷𝐷𝐷𝐷𝑆𝑆𝑆𝑆𝐷𝐷𝐷𝐷𝐿𝐿𝐷𝐷𝐷𝐷𝐺𝐺𝐷𝐷𝐷𝐷𝑆𝑆𝐷𝐷𝐷𝐷𝐿𝐿
× × ×
(3.6)
×
𝐴𝐴𝐴𝐴𝑇𝑇𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝑚𝑚𝑇𝑇𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝑈𝑈𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝐷𝐷𝐿𝐿𝑆𝑆𝑇𝑇𝐷𝐷𝑆𝑆𝐿𝐿 𝐴𝐴𝐴𝐴𝑇𝑇𝑊𝑊𝐿𝐿𝐷𝐷𝐿𝐿ℎ𝜆𝜆𝐿𝐿𝑆𝑆𝐴𝐴𝐴𝐴𝐿𝐿𝐷𝐷𝐷𝐷𝐿𝐿𝐿𝐿𝐸𝐸𝐸𝐸𝑆𝑆𝐷𝐷𝐸𝐸𝐷𝐷𝐷𝐷𝐿𝐿
𝑅𝑅𝑇𝑇𝑚𝑚𝑅𝑅𝐹𝐹𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝜆𝜆𝐹𝐹 𝑆𝑆ℎ𝑇𝑇𝑖𝑖𝜆𝜆𝐿𝐿𝐿𝐿𝐷𝐷𝐿𝐿𝜆𝜆ℎ
The remaining exposures are related to travel time spent within the haulage cycle. When
in a haulage cycle, oncoming exposure is calculated in a similar way to the other exposures. It is
calculated by multiplying the frequency of AHT interactions by the exposure time within a
segment. Because the AHT and RO are both moving, the issue of dilation due to the Doppler
Effect must be addressed. Conceptually, the AHT frequency behaves as if it is the emitter or
source frequency in a doppler calculation. The speed of the medium in the doppler calculation is
the opposing speed of the segment. The RO is exposed to the source frequency and is travelling
towards the source frequency at the speed of the segment. The result of this calculation is an
increased relative AHT frequency. The total exposure for the segment is then calculated by
multiplying the relative AHT frequency by the duration of the segment and the AHT exposure by
segment. The sum of the exposure segments results in the total exposure per haulage cycle. And
the total exposure is calculated by the total number of cycles expected for each role.
=
𝐴𝐴𝐴𝐴𝑇𝑇𝑅𝑅𝐿𝐿𝑇𝑇𝐷𝐷𝜆𝜆𝐷𝐷𝐴𝐴𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹
+
× (3.7)
𝑆𝑆𝑒𝑒𝑆𝑆𝑆𝑆𝑒𝑒𝑎𝑎𝜆𝜆𝑂𝑂𝑆𝑆𝑆𝑆𝐷𝐷𝐸𝐸𝐸𝐸𝐷𝐷𝐷𝐷𝐿𝐿𝑆𝑆𝑆𝑆𝐿𝐿𝐿𝐿𝑆𝑆 𝑆𝑆 𝑒𝑒𝑆𝑆𝑆𝑆𝑒𝑒𝑎𝑎𝜆𝜆𝑆𝑆𝑆𝑆𝐿𝐿𝐿𝐿𝑆𝑆
𝐴𝐴𝐴𝐴𝑇𝑇𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹
𝑆𝑆𝑒𝑒𝑆𝑆𝑆𝑆𝑒𝑒𝑎𝑎𝜆𝜆𝑂𝑂𝑆𝑆𝑆𝑆𝐷𝐷𝐸𝐸𝐸𝐸𝐷𝐷𝐷𝐷𝐿𝐿𝑆𝑆𝑆𝑆𝐿𝐿𝐿𝐿𝑆𝑆
41
|
Colorado School of Mines
|
=
𝐹𝐹𝑎𝑎𝜆𝜆𝑎𝑎𝐶𝐶𝑇𝑇𝜆𝜆𝐶𝐶𝑊𝑊𝑒𝑒𝑇𝑇𝑆𝑆ℎ𝜆𝜆𝑒𝑒𝑎𝑎𝐸𝐸𝑥𝑥𝐸𝐸𝑜𝑜𝑚𝑚𝐸𝐸𝑜𝑜𝑒𝑒𝑂𝑂𝐷𝐷𝐹𝐹𝐷𝐷𝑆𝑆𝐷𝐷𝐷𝐷𝐿𝐿
× ×
×𝑇𝑇𝑜𝑜𝜆𝜆𝑎𝑎𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑒𝑒𝑚𝑚 (3.8)
𝐴𝐴𝐴𝐴𝑇𝑇𝑅𝑅𝐿𝐿𝑇𝑇𝐷𝐷𝜆𝜆𝐷𝐷𝐴𝐴𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹 𝐴𝐴𝐴𝐴𝑇𝑇𝐸𝐸𝐸𝐸𝑆𝑆𝐷𝐷𝐸𝐸𝐷𝐷𝐷𝐷𝐿𝐿𝑏𝑏𝐹𝐹𝑆𝑆𝐿𝐿𝐿𝐿𝑆𝑆𝐿𝐿𝐷𝐷𝜆𝜆 𝑆𝑆𝑒𝑒𝑆𝑆𝑆𝑆𝑒𝑒𝑎𝑎𝜆𝜆𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝜆𝜆𝐷𝐷𝐷𝐷𝐷𝐷
�
𝑅𝑅𝑇𝑇𝑚𝑚𝑅𝑅𝐹𝐹𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝜆𝜆𝐹𝐹
Dovetailing exposure is calculated by determining the average distribution of AHTs
within the haulage cycle and multiplying it by the weighted average exposure from AHTs. This
represents the portion of a haulage cycle within 30 meters of an AHT. The total cycles of a role
are then multiplied by this exposure to determine the total exposure which is then reduce by the
risk of fatality.
=
𝐹𝐹𝑎𝑎𝜆𝜆𝑎𝑎𝐶𝐶𝑇𝑇𝜆𝜆𝐶𝐶𝑊𝑊𝑒𝑒𝑇𝑇𝑆𝑆ℎ𝜆𝜆𝑒𝑒𝑎𝑎𝐸𝐸𝑥𝑥𝐸𝐸𝑜𝑜𝑚𝑚𝐸𝐸𝑜𝑜𝑒𝑒𝐷𝐷𝐷𝐷𝐴𝐴𝐿𝐿𝜆𝜆𝐷𝐷𝐷𝐷𝑇𝑇𝐷𝐷𝐷𝐷𝐿𝐿
× × ×
(3.9)
𝑇𝑇𝑜𝑜𝜆𝜆𝑎𝑎𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑒𝑒𝑚𝑚 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑒𝑒𝑇𝑇𝑇𝑇𝑆𝑆𝑒𝑒 𝐴𝐴𝐴𝐴𝑇𝑇𝐹𝐹𝐷𝐷𝐿𝐿𝐹𝐹𝐷𝐷𝐿𝐿𝐷𝐷𝐹𝐹𝐹𝐹 𝐴𝐴𝐴𝐴𝑇𝑇𝑊𝑊𝐿𝐿𝐷𝐷𝐿𝐿ℎ𝜆𝜆𝐿𝐿𝑆𝑆𝐴𝐴𝐴𝐴𝐿𝐿𝐷𝐷𝐷𝐷𝐿𝐿𝐿𝐿𝐸𝐸𝐸𝐸𝑆𝑆𝐷𝐷𝐸𝐸𝐷𝐷𝐷𝐷𝐿𝐿
𝑅𝑅𝑇𝑇𝑚𝑚𝑅𝑅𝐹𝐹𝐷𝐷𝜆𝜆𝐷𝐷𝑇𝑇𝐷𝐷𝜆𝜆𝐹𝐹
Intersections are included on the fault tree but were not included in this model. This
model represents a simplified closed loop system. The addition of a more complex system of
roads will increase the exposures experienced within an AOZ; however, the increase is not
anticipated to be significant. The impact to exposure can be explored by envisioning a duplicate
parallel set of segments and their impact on exposures. This scenario would double the number
of exposures within the AOZ. The opposite hypothetical scenario would be a set of segments 90
degrees to the initial modeled segments. This would result in only one crossing. This single
crossing could be represented as a single 30-meter-long segment with double the frequency of
AHT interactions. The overall impact to AHT exposure in the 90-degree case would be minute
in comparison to all interactions experienced along the entire haul route. Therefore, it was
assumed that intersections do not materially affect the AHT exposures.
The combination of the cause of event or incident and the fatality weighted exposure
within the AOZ can be expressed as a fault tree Figure 3.3. The fault tree is specific for each
role and should be calculated to determine the minimum level of risk mitigation necessary to
reach the maximum tolerable risk for the site.
42
|
Colorado School of Mines
|
short while the AHT passes a manned machine. The exception is during the loading or dumping
phase of the haulage cycle; however, this exposure is offset by the reduced risk of interactions
between AHTs and loading units or dozers due to size and area geometries. All calculations
result in the fatality weighted exposure for the role.
Loading Unit Operator
Role: Loading unit operators operate shovels, excavators, and large front-end loaders
within the mine. Typically, shovels and excavators operate within a relatively limited area while
front end loaders can more quickly traverse the mine site.
Exposure: The loader’s principal exposure to the AHTs is during loading operations.
They also experience limited exposure during their travel from the loading unit during shift
changes and breaks. It is assumed that a loading unit operator makes four trips through the AOZ
per shift for shift change and breaks. The chance of a fatality if an incident occurs while
traveling in a LV is greater than while operating an HME, but the exposure period is much
lower. During production, the operator would be exposed to EMESRT scenario class L2. While
tramming a loading unit to another location, the operator would be exposed to the following
EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and T4. Tramming is
expected to happen a few times per month and was assumed to represent 5% of an average shift.
Consequence: For excavators loading from a bench above the AHT, the consequence is
very limited due to difficulty in reaching the excavator. For shovels loading on the same bench
as an AHT, the consequence is also very limited due to size differences. An AHT does not even
reach the cab of the shovel. It is assumed that the chance of a fatality due to an AHT and a
loading unit collision is 5%. DMIRS and MSHA fatalities databases indicate incident rates were
non-existent or too infrequent to provide useful insight. The consequence of a collision while
being transported in a LV to the loading unit would likely be a fatality. The assumed chance of
fatality is 90%.
Motor Grader Operator
Role: The motor grader operator’s primary task is maintenance of the haul roads. In static
lane sections of the AOZ, the operator will maintain sections of the haul road adjacent to AHTs.
48
|
Colorado School of Mines
|
The lane segment adjacent to the motor grader will be treated as a single lane to allow the
machine to pass maintenance activities.
Exposure: Motor grader operators are exposed to the risk of AHTs passing them in the
opposing lane and AHTs overtaking them from behind. During maintenance operations, the
operator may be exposed to the following EMESRT scenario classes: L1, L4, L5, L6, L7, C1,
C2, C3, T2, T3, and T4. During tramming to another location, the operator may be exposed to
the following EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and T4.
During a typical shift, it is assumed that motor grader operators spend 60% of their time
maintaining static autonomous lanes and make four trips through the AOZ tramming between
work areas.
Consequence: The consequence of a collision of an AHT with a motor grader is much
lower than with a LV, but higher than a loading unit. It is assumed that the chance of a fatality
due to an AHT and a motor grader collision is 50%.
Water Cart/Truck Operator
Role: The water truck/cart operator role is to apply water to the roadways to minimize
dust and improve road reliability. Typically, a water truck/cart collects a load from outside the
AOZ, or from within an exclusion zone within the AOZ, and then applies water to the static
lanes of the AOZ. If they were required to apply water to a dynamic zone within the AOZ, an
exclusion boundary would be utilized to isolate the water truck/cart from any AHTs.
Exposure: Water truck/cart operators are primarily exposed to AHTs while they water
the static lanes of an AOZ. During operations, the operator would be exposed to the following
EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and T4 (Figure 2.8). It is
assumed that during a shift a water cart/truck completes twenty haulage cycles per average shift.
Consequence: The consequence of a collision of an AHT with a water cart/truck is much
lower than with a LV, but higher than a loading unit. It is assumed that the chance of a fatality
due to an AHT and a water cart/truck collision is 20%.
49
|
Colorado School of Mines
|
Dozer Operator (Dump)
Role: Dozer operators located at a dump are tasked with spotting dump loads, pushing
off loads, and maintaining the dump area. Depending on specific site operating procedures,
dozers may operate within the dynamic lane area of a dump or may be precluded from entering
an active dump area.
Exposure: If dozers are allowed to operate within the dynamic lane portion of the dump,
the operator’s principal exposure to the AHTs is during dumping operations. They also have
limited exposure during travel to and from their dozer during shift changes and breaks. During
production, the operator would be exposed to the following EMESRT scenario classes: L1, L2,
L4, L5, L6, L7, C1, C2, and C3. During tramming of a dozer to another location, the operator
would be exposed to the following EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2,
T2, T3, and T4; but this scenario may only happen a few times per month.
Consequence: The consequence of a collision of an AHT with a dozer is much lower
than with a LV, but higher than a loading unit. It is assumed that the chance of a fatality due to
an AHT and a dozer collision is 20%.
Wheel Loader/Dozer Operator (Clean-up)
Role: Wheel Loader/Dozer operators tasked with clean-up near a loading area generally
operate within an exclusion zone. The clean-up dozer typically operates on the opposite side of
the loading unit. They are utilized to clean-up the area utilized by trucks being loaded by the
shovel.
Exposure: When a dozer trams from one location to another, the operator would be
exposed to the following EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and
T4 (Figure 2.8). It is assumed that a clean-up dozer makes four LV trips through the AOZ,
spends 10% of an average shift tramming, and 20% of its working time exposed to passing
AHTs.
Consequence: The consequence of a collision of an AHT with a dozer is much lower
than with a LV, but higher than a loading unit. It is assumed that the chance of a fatality due to
an AHT and a dozer collision is 20%.
50
|
Colorado School of Mines
|
AHS Pit Supervisor, Pit Technician, and Wi-Fi Technician
Role: All three roles continually patrol the AOZ to address and correct issues within the
AOZ. They also often function as a shuttle driver for HME operators for shift changes, breaks,
and hot seating.
Exposure: The principal exposure of these roles to AHTs is during their drive within or
between locations. When an they drive from one location to another, they would be exposed to
the following EMESRT scenario classes: L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and T4 (Figure
2.8). These roles experience minimal risk while parked and observing a production area. They
are following procedure and are maintaining a sufficient buffer or barrier from haulage ways. It
is assumed that these roles each make ten full haulage cycles worth of travel each shift. This is
half the number of trips of a water cart/truck and about a third of the number of trips that each
AHT makes.
Consequence: The consequence of a collision would likely be a fatality due to the
location of the supervisor in a LV. The assumed chance of fatality is 90%.
Fitter/Mechanic (Service Truck)
Role: The fitter/mechanic will travel throughout the AOZ servicing and repairing HME.
Actual servicing occurs within an exclusion zone of the AOZ and greatly reduces the exposure to
AHT risks.
Exposure: A fitter/mechanic’s principal exposure to the AHTs is during their drive
between service locations. They would be exposed to the following EMESRT scenario classes:
L1, L2, L4, L5, L6, L7, C1, C2, T2, T3, and T4 (Figure 2.8). They have minimal risk while
parked as long as they follow procedure and are maintaining a sufficient buffer or barrier from
the haulage ways. It is assumed that these roles each make 5 full haulage cycles worth of travel
each shift.
Consequence: The consequence of a collision would likely be a fatality due to the
fitter/mechanic being in an LV. The assumed chance of fatality is 90%. However,
fitter/mechanics are rarely operating outside an exclusion zone.
51
|
Colorado School of Mines
|
3.9 Calculated Exposures and SIL Requirements by Role
Table 3.5 details the factors and exposures by role within the AOZ. All LV-only-use
from AHS Supervisors, Pit Technicians, and Wi-Fi technicians were combined. From the
exposure rates, the cause of event rates, and a maximum tolerable failure rate of 10-5 pa,
maximum PFDs was calculated for each primary role within the AOZ. Additional roles were not
calculated; however, their exposure hours would be much less than the listed primary roles.
Additionally, a SIL requirement based on the broadly acceptable risk (10-6 pa) was calculated for
each role. The results are shown in Table 3.4.
Table 3.4: Required SIL based on Role.
Machine Type Annual Risk Max. Failure Rate Required SIL ALARP SIL
Shovel/Excavator 5.0×10-4 7.1×10-2 SIL 2 SIL 3
Motor Grader 3.5×10-4 3.4×10-2 SIL 2 SIL 3
Water Truck/Cart 1.3×10-4 1.3×10-1 SIL 1 SIL 2
Dump Dozer 5.0×10-4 6.4×10-2 SIL 2 SIL 3
Cleanup Dozer 5.0×10-4 6.4×10-2 SIL 2 SIL 3
LV Only 2.9×10-4 5.7×10-2 SIL 2 SIL 3
Service Truck 1.5×10-4 3.4×10-2 SIL 2 SIL 3
3.10 Modifications to an AOZ to Reduce Integrity Targets
Modifications to an AOZ could reduce the SIL requirements. The furthest extreme would
be complete removal of personnel from the AOZ and would effectively remove the risk to life.
Risk would still be associated with equipment or environmental damage. A more reasonable
change would be stricter limits on the exposure of manned machines to AHTs. If the rate of
exposure is reduced sufficiently, the required SIL could be reduced to a SIL 1. This could allow
for the removal of one or more SIS on the AHT. The more subtle change to risk associated with
removal of LVs from autonomous zones would not reduce the required SIL level but may reduce
the complexity of achieving a SIL 2 rating with the existing SIS.
52
|
Colorado School of Mines
|
THEORETICAL SIL OF CURRENT AHS
At this time, no manufacturers of AHSs produce a SIL rated system; however, it is likely
SIL rated systems will become common place and will eventually be required. Alberta OH&S
already requires that AHS meet ISO 17757 and within ISO 17757, which states all safety related
functions must comply with functional safety standards (International Organization for
Standarization 2019). As more regulatory bodies adopt ISO 17757, AHSs will need to become
compliant with functional safety standards. A mine should be able to define its safety integrity
targets (SIT) as part of the typical specifications during the request for proposal process.
Manufacturers will select from their available hardware and software to meet those requirements.
While Caterpillar and Komatsu AHS do not have a rated system, sufficient publicly
available information is available on the Caterpillar and Komatsu systems to provide a
preliminary evaluation of their hardware and architecture to determine if their systems could be
SIL rated. To be SIL rated, a device must satisfy the quantitative random hardware failure rate,
or PFD, as well as the qualitative requirements that are not easily quantified. Smith and Simpson
summarize the qualitative requirements of the four SIL ratings (Smith and Simpson 2017):
• SIL 1 is relatively easy to achieve especially if ISO 9001 practices apply throughout the
design.
• SIL 2 is not dramatically harder than SIL 1 to achieve although clearly involves more
review and tests and thus, has an increased cost. If ISO 9001 practices apply throughout
the design, it should not be difficult to achieve.
• SIL 3 involves a significantly more substantial increment of effort and competence than
is the case from SIL 1 to SIL 2. Specific examples include the need to revalidate the
system following design changes and the increased need for operator training. Cost and
time will be significant factors and the choice of vendors will be more limited by lack of
ability to provide SIL 3 designs.
• SIL 4 involves state-of-the-art practices including “formal methods” in design. Cost will
be extremely high and competence in all techniques is required. There is a considerable
body of opinion that SIL 4 should be avoided and additional levels of protection should
be preferred.
54
|
Colorado School of Mines
|
It is unlikely that any of the SIS utilized by an AHS would qualify as a SIL 3 or SIL 4,
but it is also unnecessary. Due to the low exposure rates within the AOZ, a combination of SIL 1
and SIL 2 systems should provide sufficient risk reduction.
4.1 Layers of Protection in the Current Deployments
Layers of protection of current AHS installations were addressed in Section 2.1. Layers
of protection are not necessarily SIS. While manufacturers may argue that all the layers of
protection are safety systems, AOZ access control, site procedures, and situational awareness
would likely not meet the requirements of IEC 61508 and IEC 61511. A revised AHS layers of
protection can be seen in Figure 4.1.The AOZ access control system and site procedures would
be classified as an external independent layer of protection. The impact to fault tree analysis are
reflected in exposure factor, fatality factor, and shift factor. The division of the resultant factor
and the maximum tolerable failure rate results in safety integrity targets from the last chapter.
The remaining layers of protection are represented in the right side of a typical fault tree
analysis. It should be noted that situational awareness is critical to the BPCS and is not
considered a SIS, Figure 4.2.
Figure 4.1. AHS Layers of Protection showing external IPLs in green. SIS in gray.
55
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.