University
stringclasses 19
values | Text
stringlengths 458
20.7k
|
|---|---|
Colorado School of Mines | Figure 3.21: Equivalent Atkinson’s friction factor as a function of wall roughness height
Without any adjustments and a roughness height of 0m, the airway friction factor closely
resembles a smooth airway. Based on the values listed in literature, a wall roughness height of
0.15m-0.50m (0.49-1.64ft) represents a reasonable value for the moderately to highly obstructed
coal mine entries presented in Table 3.11 and Table 3.12. For this study, a roughness height value
of 0.3m (1ft) was chosen to represent the wall roughness condition in a longwall coal mine. It was
also found that the result without the addition of inflation layers near the wall cannot accurately
resolve the effect of wall roughness near the wall. The resulting pressure drop is higher than it
should be, which then reflected on the calculated friction factor value. However, considering the
difference in the result and the number of cell required to mesh it, the currently used mesh size
0.36m for the mine entry is deemed reasonable.
3.8 Solution Approach and Convergence
For a steady-state simulation, good convergence is achieved when there are no longer any
significant changes in the solution with further iterations. To determine convergence, FLUENT
recommends checking the following:
Discrete conservation equations are solved to a specific tolerance;
Overall mass, momentum, energy and scalar balances;
57 |
Colorado School of Mines | A decrease in residuals by at least three orders of magnitude;
An energy residual decrease by six orders of magnitude;
A species residual decrease by five orders of magnitude; and
Monitor relevant key variables
Following this guideline, the convergence criteria used for all models in this study is as
follows:
Continuity: 1x10-3
Momentum: 1x10-4
Turbulence parameter, Kappa and Epsilon: 1x10-3
Energy: 1x10-6
Methane species: 1x10-5
To achieve good convergence, the initial solutions were solved for the turbulent case, followed
by the energy equations. Several surface monitors were set at various locations in the model, which
include the monitoring of:
Airflow quantity at longwall shield 9 and 161;
Airflow quantity passing through the headgate ventilation curtain;
Methane concentrations at the outlet
Figure 3.22 and Figure 3.23 show the comparison of convergence results for the base case model
using velocity and pressure boundary conditions. The residuals appear steady for species and
energy but fluctuate for momentum equations, especially for the pressure boundary case. In both
cases, there is no longer significant change in the monitored key variables with further iterations
and the overall mass balance is achieved, thus the convergence is considered to be reached. Based
on the residual value for each parameter, the pressure boundary case achieve a better overall
convergence compared to the velocity boundary case.
58 |
Colorado School of Mines | 4 CHAPTER 4
ANALYSIS OF RESULTS
4.1 Airflow distribution across longwall face
4.1.1 Normal ventilation conditions
In normal ventilation, 100,000 cfm of fresh air is delivered from the headgate entry; 25,000
cfm is assumed to leak through the headgate curtains and 10,000 cfm of air is returned through the
belt entry, resulting in 65,000 cfm of air delivered to the longwall face. The tailgate entry and three
bleeder entries are each set to supply 10,000 cfm of air. The result of the CFD simulation for
airflow distribution across the face in normal ventilation conditions is shown in Figure 4.1. An
elevation of 2.3ft from the mine floor was chosen for this, as it clearly shows the air leakage
through the back of the shields and the airflow blockage by the face equipment, while the
measurement of 5ft from the mine floor was chosen to show the location of the shearer tailgate
drum, which has the potential to ignite the air-methane mixtures at the tailgate corner once the
tailgate ventilation changes.
Airflow distribution at the headgate side of the face shows that the stageloader, headgate drive
and crusher at the headgate intersection provide an obstruction for airflow trying to enter the face.
The majority of that flow passes over these obstacles through a reduced opening, resulting in an
increase of airflow velocity in this area. The pathlines of leaked fresh air around the headgate
corner, which is presented in Figure 4.2, also shows that some of the fresh air supplied to the face
is leaking into the gob through the face curtain and through the gaps behind the shields.
Near the face’s headgate corner, a small portion of this leaked fresh air flows through the high
permeability area along the gob’s headgate side, enters back into the headgate entry and mixes
with the fresh air that passed though the headgate curtain. The remainder of leaked fresh air toward
the tailgate flows either along the gob void behind the shields of the face or through the gob directly
toward the back of the panel, as shown in Figure 4.3.
62 |
Colorado School of Mines | The results show that, near the headgate of the face, most of the supplied fresh air is
concentrated near the coal face, especially prior to shield 6 where the face ventilation curtain is
set; however, as airflow passed shield 6, the fresh air started to leak through the gaps behind the
shields. By the time it reached the tailgate side of the face, the majority of supplied air flows behind
the shields parallel to the face, leaving only small quantity of fresh air to flow inside the last couple
of shields where the shearer is located. Figure 4.5 shows the leakage rate across the longwall face
in normal ventilation conditions.
Figure 4.5: Air quantity distribution along the longwall face before the roof fall
Airflow distribution across the face shows there was minimum leakage between shield 1 and
6 due to the face ventilation curtain. However, significant leakage occurred starting at shield 7
until the last shield. Out of the 65,000 cfm supplied to the face, only around 10,000 cfm remained
on the face at the tailgate. This amount of airflow is likely insufficient to dilute the methane along
the face and can cause an accumulation of methane at the shearer. The extent of air leakage is
highly affected by the caving condition behind the shields, the pressure drop across the face and
the ventilation system used. In progressively sealed gobs, the fresh air that leaked into the gob at
the headgate corner and along the face will be pulled back into the face through the shield gaps as
65 |
Colorado School of Mines | it approaches the tailgate, as demonstrated by Yuan et al. (2012) and Saki (2015). In bleeder
systems, the air that leaks into the gob does not return to the face, but rather flows directly towards
the bleeder fan at the back.
4.1.2 Tailgate entry blocked by a roof fall inby the longwall face
After a tight roof fall blocks the tailgate entry inby the face, stoppings outby the face must be
opened to allow ventilation access from the face to the bleeder entries. As a result, airflow from
the face is now directed toward the nearest open crosscut outby the face. Figure 4.6 shows a
comparison of airflow distribution around the tailgate area before and after the roof fall.
Figure 4.6: Comparison of tailgate conditions before (left) and after (right) a roof fall occurred
The roof fall also increased the mine resistance, especially on the tailgate side. This has
resulted in changes in airflow quantity flowing from the tailgate entries. The opening of the
stopping outby the face provided a new, less restrictive airway for fresh air supplied from the
tailgate entry, which resulted in a significant increase of airflow quantity from that point. At the
same time, the fresh air supplied through tailgate entry numbers 2 and 3 were reduced due to this
pressure change. Figure 4.7 shows the new air quantity flowing from the tailgate entries after the
roof fall occurred.
66 |
Colorado School of Mines | Figure 4.11: Methane distribution around longwall face tailgate before a roof fall, units in mole
fraction
Pre-fall, the close-up view of the tailgate conditions show that the face air sweeps the corner
of the gob on the tailgate side and prevents methane accumulation in this area. The results also
show that, although the fresh air leaks across the face has significantly reduced air quantity at the
tailgate corner, this leaked air help preventing methane buildup in the gob area immediately behind
the longwall face.
4.2.2 Tailgate entry blocked by the roof fall inby the longwall face
Figure 4.12 shows the methane distribution around the tailgate corner after the roof fall
blocked the tailgate entry. The roof fall in the tailgate entry inby the face forces a large air current
to flow to the newly opened crosscut outby the face. In addition, methane from inside the gob will
be pulled toward the tailgate corner. The close-up view of the tailgate area shows that contaminated
air traveling behind the shields returns to the face near the tailgate and then travels outby the face,
passing close to the tailgate-side shearer drum. This ventilation condition can create explosion
hazards if the shearer manages to interact with the shearer drum while cutting the coal face at the
tailgate corner. The roof fall’s impact on methane accumulations and mixing patterns in the tailgate
area depend largely on the tightness of the fall and the amount of methane outgassing from the gob
at the tailgate corner, as demonstrated in a study by Brune and Sapko (2012).
70 |
Colorado School of Mines | Figure 4.13: Comparison of methane distribution inside the gob
The roof fall at the tailgate entry hindered fresh air from diluting the methane at the tailgate
gob corner, thus allowing the methane inside the gob to accumulate around that area.
The impact of the roof fall on the methane distribution inside the gob can also be analyzed by
observing the formation and movements of EGZs in the gob. Figure 4.14 shows a comparison of
the formation of EGZ inside the gob before and after the roof fall occurred at 5ft above the floor.
Before the roof fall occurred, the EGZ is located at the back of the gob and far from the
longwall face; after the roof fall blocked the tailgate entry, that area is now filled with near-
explosive gas mixtures and the EGZ is being pulled closer to the longwall face, where it can be
ignited by the shearer’s cutting action. The extent of the EGZ movement will vary based on gob
characteristics and the amount of methane originated from the gob.
72 |
Colorado School of Mines | 5 CHAPTER 5
RECOMMENDATIONS FOR ADDITIONAL METHANE MONITORING
The Willow Creek and UBB mine explosions highlight the need for additional monitoring
locations to detect methane outgassing from the gob due to changes in tailgate ventilation,
barometric pressure drops and other factors. Based on the literature and studies done on this topic,
in order to be considered effective, the proposed monitoring locations should satisfy the following
criteria:
Give early indication of changes in tailgate ventilation;
Give early indication of methane outgassing from the gob;
A fixed location, moved along with the face advance; and
Located in area that is not prone to damage, or which can be covered with coal dust or soaked
by water sprays
The simulation results show that the area most prone to methane accumulation is located
around the tailgate side of the longwall face. The ventilation in this area is highly affected by the
caving condition behind the shields and is sensitive to changes in the tailgate setup, thus the need
to be properly monitored. Figure 5.1 and Figure 5.2 each show the airflow and methane distribution
near the tailgate corner after the roof fall.
Figure 5.1: Pathlines showing airflow distribution near the tailgate after a roof fall, units in
ft/min
76 |
Colorado School of Mines | Figure 5.2: Pathlines showing methane distribution near the tailgate after a roof fall
The results clearly show the possibility of methane from the gob to enter the longwall face,
bypassing the methane monitor mounted on the shearer body and tailgate drive, and interacting
with the shearer tailgate drum. Therefore, the proposed additional methane monitoring location
must also resolve this issue.
Based on the observed airflow and methane patterns near the tailgate corner, the best location
to install the additional methane monitor should be close to the shield gaps at the back of the last
few shields near the tailgate corner. The proposed locations for the methane monitor are presented
in Figure 5.3. These locations satisfy the requirements listed above and do not require significant
modification the existing longwall shields model. These locations move along with the advance of
the face, are not prone to damage or covered with coal dust, and should be able to give early
indications of potential methane entering from the gob into the longwall face.
77 |
Colorado School of Mines | 6 CHAPTER 6
CONCLUSIONS
The use of computational fluid dynamics (CFD) can provide a detailed interaction between
air flow at the longwall face and leaks in the voids behind the shields. In addition, the impact of a
roof fall in the immediate tailgate entry can be modeled and visualized. The resulting airflow
profiles and methane-air mixture conditions can provide a better understanding of tailgate
ventilation hazards caused by the accumulation of methane near the tailgate corner.
The results confirm that the tailgate corner of the longwall face is a critical area that is prone
to face ignitions and needs to be properly monitored. Caving conditions behind the shields have a
significant impact on the ventilation conditions of the tailgate corner area. Poor gob caving can
lead to insufficient fresh air at the tailgate corner and also make this area prone to methane
accumulations and explosion hazards.
Roof falls in the tailgate entry inby the face can create an explosion hazard. Not only does the
roof fall hinder the ability of face air to dilute methane accumulations in the tailgate corner, it can
also deflect contaminated air from the gob back into the face and pull the EGZ inside the gob
closer to the face. Roof control at the tailgate and ventilation monitoring there are, therefore,
important to preventing longwall face ignitions.
The current practice of mounting methane monitors at the tailgate drive and shearer body is
not effective for detecting methane outgassing from the gob to the face. Additional monitoring
location is deemed necessary to provide an early indication for such an event.
Maintaining effective bleeder ventilation is key in ensuring the safety of the longwall
operation. Adequate face ventilation, good roof control and proper methane monitoring locations
can all help minimize the explosion risk around the tailgate corner of the longwall face.
79 |
Colorado School of Mines | ABSTRACT
Duringemergency minerescue operations, life–threatening situationsand commu-
nication problems are frequent and recurring issues. Autonomous robotic vehicles
can assist rescue workers by exploring areas that may be unsafe for rescue teams
to enter and by establishing a wireless communication network. A key trait of an
autonomous vehicle is the ability to map and localize relative to the surrounding en-
vironment, using artificial or natural landmarks. One technique for localization in a
mineistousetheminewall’suniquefeaturesaslandmarks. Traditionally, acommon
sensor choice for detecting these natural landmarks is a scanning laser; however, laser
sensors are unreliable in the conditions that are common in mine emergency situa-
tions (e.g. thick smoke). Thus, localization and mapping algorithms using ultrasonic
range finders are examined; these sensors are impervious to thick smoke and can,
therefore, accurately range in emergency mine rescue scenarios. This thesis focuses
on the design of an autonomous robotic vehicle, along with the development of a
localization algorithm using ultrasonic range finders. Simulations of the localization
algorithm lead to the conclusion that localization using ultrasonic range finders is
possible if the mine wall texture exceeds, or is distinguishable from, the sensor noise.
iii |
Colorado School of Mines | CHAPTER 1
INTRODUCTION
Robotics technology has seen many advancements in the past 20 years; com-
puters are becoming faster, allowing robots to become more intelligent and more
autonomous. One type of robotics technology is tele–operated vehicles. These vehi-
cles travel in areas that are unfavorable to humans, serving as an invaluable tool in
environments ranging from search and rescue in rubble from a collapsed building to
exploring the hostile environment on Mars. Another application of these intelligent
machines can be in emergency response operations in subterranean environments,
where the air may be toxic, the ground may be flooded, or an explosion may be
imminent.
1.1 Motivation
Mining is a dangerous occupation. Since 1976, in the United States alone, there
have been 20 recorded mining disasters (defined as 5 or more fatalities) in both coal
and metal/nonmetal mines, resulting in 215 fatalities [9]. Despite advances in mine
safety, 30% of the fatalities have occurred in the past decade.
Arecentexampleofaminingdisasteroccurredin2006, whenanexplosiontrapped
13 miners at the Sago Mine in Sago, West Virginia [27]. Toxic carbon monoxide levels
in the mine delayed rescue attempts until 12 hours after the explosion. Even after
the toxic gases diffused, rescue teams proceeded with caution, as they continually
tested for hazards such as water seeps, explosive gas concentrations, and unsafe roof
conditions. Due to multiple safety concerns, rescue workers took an additional 30
hours after the initial 12 hour delay to reach the 12 victims and single survivor,
located approximately two miles from the mine entrance [26]. On average, rescue
operations proceeded at a rate of 250 feet per hour (two miles/12 hours), leaving
1 |
Colorado School of Mines | room for improvement.
In dangerous situations such as these, robotic vehicles can be invaluable tools for
exploring in areas that may be unsafe for rescue teams to enter; twenty-four hours
after the initial explosion at the Sago Mine, the Mine Safety and Health Adminis-
tration (MSHA) deployed an explorer robot. Equipped with a gas analyzer and live
video feedback, the explorer robot could have expedited rescue efforts if it had not
encountered technical difficulties [37].
Even in less–dangerous situations, robots can work alongside rescue workers, as-
sisting in communications or carrying equipment. According to the Sago Mine Ac-
cident Investigation Report, communication was a major problem plaguing rescue
operations: “The handheld radios become less reliable as the distance between users
isincreasedorwhentheusersarenotindirectlineofsightofeachother”[26]. Indeed,
radio frequency power attenuation around bends in a mine can be more than 20 times
greater than line-of-site attenuation [22]. Communication was such a problem that
losing radio contact “caused the outby team member to repeatedly wade through the
water while trying to maintain communication with the inby rescuers and outby res-
cuers” [26]. Robotic vehicles can assist rescue workers by establishing an automatic,
self-configuring communication network [22].
In this thesis, we explore one aspect of the various technologies needed to demon-
strate an effective robotic communication network for a mine emergency rescue sce-
nario. In the remainder of the chapter, the overall project is described, then a poten-
tial scenario outline to resolve the communication issue is proposed. Then, the next
section introduces an important aspect needed to implement the scenario outline:
localization. The final section of this chapter provides an outline of the thesis contri-
butions, which includes part of the vehicle design, as well as a localization algorithm.
2 |
Colorado School of Mines | 1.2 MineSENTRY Description
A project in the Center for Automation, Robotics, and Distributed Intelligence
(CARDi) Research Center, Mine Safety and Rescue Through Sensing Networks and
Robotics Technology (MineSENTRY) is a research program at the Colorado School
of Mines, with a focus on developing emergency first response robotic vehicles for
mine rescue efforts and wireless communication in subterranean environments.
1.3 Scenario Outline
In the event of a mining disaster, such as an explosion or a cave-in, MineSENTRY
proposesawirelesstetheringsolutiontoexplorethemineandre-establishcommunica-
tion. To illustrate the wireless tethering, consider the scenario depicted in Figure 1.1.
Figure 1.1: Autonomous Mobile Radio (AMR) Tethering [24]
ABobcatfront-endloaderleadsacaravanofAutonomousMobileRadios(AMRs)into
themine. EachAMR isequipped withan arrayofsensors forautonomous navigation,
along with a radio capable of relaying signals from a base station command center to
the Bobcat leader. Using these vehicles, an operator can clear blockages from a safe
location by tele-operating the Bobcat using live video feedback.
The tethering occurs in the sequence illustrated in Figure 1.2. In the MineSEN-
TRY scenario, the leader would be the tele–operated Bobcat front-end loader, the
AMR is an autonomous vehicle carrying a radio, and the base station could be either
3 |
Colorado School of Mines | fixed or mobile near the entrance of the mine. All vehicles begin near the base sta-
tion, with identical received signal strength (RSS). When the Leader moves, the RSS
among the radio nodes becomes unbalanced. The AMR detects this imbalance, and
automatically navigates forward until the RSS is approximately equal among radio
nodes. Equal RSS is not necessarily equivalent to equal distance. This tethering
experiment can be extended to any number of AMRs in between the base station and
the Bobcat; the AMRs would still configure themselves for equal RSS among all radio
nodes.
Figure 1.2: AMR Tethering Sequence [24]
1.4 Autonomous Mobile Radio (AMR) Localization
In order to navigate in a mine, as depicted in the AMR tethering sequence in
Figure 1.2, autonomous vehicles rely on intelligent computer algorithms that make
decisions based on exteroceptive and propreoceptive sensor readings. One such algo-
rithm is a localization algorithm; localization is the problem of estimating the pose
of a robot relative to a map—the map may be known a-priori, or the robot may use
a Simultaneous Localization and Mapping (SLAM) algorithm. The simplest form of
localization is dead reckoning, where a robotic vehicle begins with an initial position
4 |
Colorado School of Mines | and propagates that position forward based on odometry and heading information
(from a gyro or steering angle sensor, for example). Dead reckoning has inherent
uncertainty and drift, requiring periodic correction using exteroceptive sensors.
For example, localization in a mine using scanning laser sensors for drift cor-
rection has already been demonstrated [20]. However, due to reasons detailed in
Chapter 2, ultrasonic range finders are more appropriate in emergency situations.
Though many studies utilize ultrasonic sensors for SLAM algorithms in an indoor
environment, localization using ultrasonic sensors in a mining environment is a novel
concept. Ultrasonic sensors may be able to replace scanning laser sensors in drift–
correcting localization algorithms, which would prove useful during mine disaster
rescue operations.
1.5 Thesis Contributions
This thesis, along with work by Hulbert [22] and work by Meehan [13], provide the
documentation for the entire MineSENTRY research project; the three theses contain
all the information required for replicating and improving upon the MineSENTRY
experiments. This thesis focuses on a localization algorithm for the AMR using
ultrasonic range finders, including:
1. The electronics and sensors involved in converting a stock EZ–GO golf cart into
an AMR, particularly covering:
(a) Sensors
(b) Circuit Schematics
(c) Wiring
(d) Robotic-level code
2. The development of a computer simulator to test the localization algorithm on
artificial data
5 |
Colorado School of Mines | CHAPTER 2
BACKGROUND
Before diving into vehicle construction and localization algorithm development,
previous research efforts are explored. With a focus on vehicle localization algorithm
development, this chapter highlights some of the research efforts in mining rescue
robotic vehicles, as well as localization algorithms in both indoor and mine environ-
ments, using both ultrasonic and scanning laser sensors. Furthermore, this chapter
examines the use of ultrasonic range finders versus scanning sensors in Section 2.3.
Beginning with some challenges of a mining environment for robotic vehicles,
this chapter first describes some of the notable vehicles designed to overcome those
challenges. Many of the vehicles were either tele–operated or are not designed for
emergency situations (e.g. exploratory robotic vehicles), so were not equipped with
theappropriatesensorsforthisproject. Otherresearcheffortsfillinthegapbytesting
various sensors for emergency rescue operations, specifically comparing scanning laser
sensors with ultrasonic range finders.
The remaining sections cover the previous efforts on localization and mapping
algorithms. First, the algorithms that utilize scanning laser sensors in a mining envi-
ronment are explored, followed by the algorithms using the same sensor in an indoor
environment. The last section describes a SLAM algorithm for ultrasonic sensors in
an indoor environment. Each particular sensor and environment combination pro-
vides unique challenges, but none satisfy the constraints of this particular project:
ultrasonic sensors in an underground mine environment. Combined, the various re-
search efforts have the potential to produce a robotic vehicle platform capable of
traversing a mine and localizing using sensors that are appropriate for a mine rescue
operation.
7 |
Colorado School of Mines | 2.1 Mining Environment
Mines present a particular challenge for robotic vehicles. Many mines have un–
compacted surfaces and large rocks, which may be insurmountable obstacles for small
vehicles. Even for larger vehicles, the uneven terrain causes errors in dead reckoning
sensors, such as wheel odometry; tire slipping and rocks result in an overestima-
tion of distance traveled, and tire skidding results in an underestimation of distance
traveled. Many autonomous vehicles incorporate additional sensors such as Global
Positioning Satellite (GPS) receivers or magnetometers to correct the dead reckoning
drift. However, both GPS receivers and magnetometers are useless in mines; iron
deposits common in many mines skew the earth’s magnetic field and satellite signals
cannot penetrate into most mines. Furthermore, water and mud wreak havoc on elec-
tronics and sensors, and smoke and toxic gases can be present during mine disasters,
particularly after explosions.
Coal mines present an additional challenge; the dust in the air is flammable, so a
smallsparkfromonboardelectronicsmaytriggeranexplosion. Particularlyimportant
in coal mines, mining regulations require that all electronics on a vehicle entering a
mine must be either intrinsically safe (defined by the International Electrical Code)
or in an explosion-proof container (as defined by the National Electric Code).
No vehicle is suitable for all mining environments and disasters; size is one of
many dichotomous factors. Large vehicles cannot enter a mine after a roof collapse,
and a small vehicle may have difficulty traversing the rough terrain. Other problems
come into play, such as explosion-proofing, water-proofing, and using sensors robust
to mud, smoke, water, and other conditions common in a mine disaster scenario.
2.2 Mine Rescue Vehicles
Despite the numerous challenges, there have been several attempts at using robots
during mine rescues. For example, the Mine Safety and Health Administration’s
8 |
Colorado School of Mines | (MSHA) ANDROS Wolverine robot (Figure 2.1), nicknamed V2, was deployed dur-
ing the Sago Mine rescue operation. V2 is a bomb squad robot customized to be
mine permissible, with explosion-proof motors, navigation and surveillance cameras,
lighting, atmospheric detectors, vision capability, a two-way communication device,
and a manipulator arm [33].
Figure 2.1: Romotec ANDROS Wolverine Robot [33]
At 50 inches tall, and weighing 1200 pounds, V2 can only be deployed via surface
entry. DuringtheSagominerescue, therobottraveledapproximately5600feetbefore
it ran off the rail, causing the left tread to fall off and a deflated left rear wheel. The
robot failed party because of Human-Robot Interaction (HRI) issues; a single human
operator had difficulty navigating the robot [25].
On the opposite end of the robot size spectrum is the Inuktun Mine Cavern
Crawler (MCC), designed for borehole entry (Figure 2.2). During a mine disaster,
drilling boreholes is common practice in order to determine the air quality and insert
a borehole camera. At approximately eight inches wide, the MCC is a small tethered
robot, capable of fitting into boreholes measuring 83 inches and larger. The MCC
8
was deployed after a cave-in at the Crandall Canyon Mine, located 120 miles south of
Salt Lake City, Utah. The MCC encountered multiple problems when attempting to
enter the mine; namely, physical barriers blocked the robot, and water, mixed with
9 |
Colorado School of Mines | Groundhog navigates by analyzing local three-dimensional scans with regard to
traversable paths, and two-dimensional scan matching keeps the vehicle localized.
Under operator control, Groundhog successfully mapped multiple mines including the
Bruceton Research Mine and the Florence mine, both near Pittsburgh, Pensylvania.
However, duringitsmaidenvoyageunderautonomousnavigationattheMathiesmine
(also near Pittsuburgh), Groundhog experienced software difficulties. After failing to
establish wireless communication, the wireless link was manually reset and the vehicle
was driven out of the mine under manual control. Though not perfect, Groundhog is
a promising step toward autonomous navigation and Simultaneous Localization and
Mapping (SLAM) in mines.
Despite numerous achievements, none of these robotic vehicles were completely
successful. The world of robotics in mine rescue, particularly autonomous rescue
vehicles, is still new, with plenty of room for improvement. These robotic vehicles
highlight the need for sensors robust to water and other environmental variables,
mobile power and wireless communication development, and improved robot-human
interaction (such as autonomy).
2.3 Sensor Selection
One challenge that was not addressed in the previous section was equipping
the MineSENTRY AMR with sensors appropriate for emergency situations. The
GUARDIANS (Group of Unmanned Assistant Robots Deployed In Aggregative Nav-
igation by Scent) project addresses the sensor selection dilemma; they developed a
swarm robotics technique to assist firefighters [31], whom experience similar condi-
tions to those present in a mine disaster such as thick smoke, toxic gases, falling
material, and possible explosions. The small swarming robots use sensors robust to
thick smoke to complement the laser sensors; experiments from this project confirm
that laser range finders are prone to failure in the presence of smoke, depending on
11 |
Colorado School of Mines | the particulate concentration [30]. Figure 2.4 is one of the experiments.
Figure 2.4: Laser Range Finder Measurements in Smoke [30]
During this experiment, an object was placed two meters away. Starting at the one
minute mark, smoke was injected into a container surrounding the laser range finder.
During the remaining 28 minutes of the test, the smoke was released through an
opening, increasing the visibility in the test chamber. Even though the target object
remained a constant two meters away, the range readings dropped dramatically after
introducing smoke. Accurate range measurements did not return until the smoke
dissipated, confirming that smoke strongly affects a laser range finder’s ability to
make accurate measurements.
Another study expanded on the GUARDIANS project findings, directly compar-
ing a sonar range finder to a laser range finder, as shown in Figure 2.5. Range is
recorded as a function of smoke density, with a target placed two meters away. The
results support findings from the GUARDIANS experiments; the presence of smoke
yields inaccurate range measurements from laser sensors. Although ultrasonic sensors
are less precise, the experiments also confirm that smoke does not significantly affect
ultrasonic range finders.
12 |
Colorado School of Mines | Figure 2.5: Sonar and Laser performance with increasing smoke density [34]
The current industry standard for corridor detection and localization on mining
equipment is a scanning laser sensor. However, as the aforementioned experiments
confirms, thick dust and smoke can block the 800 nm wavelength light used in these
sensors. Since thick dust and smoke are often present in emergency situations, scan-
ning laser sensors are not suitable for this project. On the other hand, ultrasonic
sensors use a much longer wavelength than laser sensors (8×106 nm compared to 800
nm), and therefore, can still accurately range in thick smoke and dust.
Another characteristic of laser sensors is they have a narrow beam width; a laser
sensor measures the distance to a single point (of negligible area). While a point
measurement is advantageous in many regards, it can cause problems. For example,
a slight vehicle tilt can alter readings from a laser. In contrast, ultrasonic sensors have
an adjustable beam width (by replacing the sensor, from about 15◦ to 45◦) and return
a single measurement from a relatively large target area—the very nature of ultra-
sonic sensors provide a low-pass filter, robust to vehicle tilting, but correspondingly,
resolution is reduced.
Ultrasonic ranging sensors have numerous advantages over laser ranging sensors;
however, ultrasonic sensors are not without their disadvantages. One problem with
thesonarsensorsisthattheiraccuracyisontheorderofafewcentimeters(depending
13 |
Colorado School of Mines | on the frequency used), which causes problems when trying to build an accurate
environment model. Another problem with the sonar sensors is they have a limited
angle of incidence: angles greater than the critical angle tend to reflect away from
the sensor, yielding unreliable readings in indoor environments. However, in theory,
the mine environment helps alleviate this issue, because the diffuse mine walls scatter
acoustic waves, allowing the sensor to easily detect the wall (see Section 3.2.4).
Havingdiscussedpreviousresearcheffortsonemergencyrescuevehiclesandsensor
selection, the remaining sections highlight the localization algorithms that run behind
the scenes, utilizing ultrasonic range finders and scanning laser sensors to control
various vehicle platforms in either an indoor or underground mine environment. The
sensorandenvironmentcombinationsaredelineatedintofourcategories, eachwithan
associated localization algorithm: scanning lasers in a mining environment, scanning
lasers in an indoor environment, ultrasonic range finders in an indoor environment,
and ultrasonic range finders in an underground mine environment. The former three
scenarios are discussed in the next section and the reminder of this thesis is dedicated
to the fourth.
2.4 Previous Localization and Simultaneous Localization and Mapping
(SLAM) Overview
Localization has two distinct parts: reference guidance and dead reckoning [4].
Reference guidance refers to algorithms, such as a Kalman filter or particle filter
techniques, that use exteroceptive sensor data (e.g. laser or ultrasonic range finders)
to search the surrounding environment for landmarks (unique environment features).
The major drawback of reference guidance involves data association, or linking the
exteroceptive data with the correct landmark. Dead reckoning comes to the rescue by
givinganinitialpositionestimate,eliminatingtheimprobablelandmarks. Fromthere,
the exteroceptive algorithms finely tune the robot’s position, eliminating the drift
14 |
Colorado School of Mines | from dead reckoning. Thus, reference guidance and dead reckoning algorithms are
complementary. Exteroceptive choice becomes important; reference guidance works
best with precise and accurate data, though even the best sensors are ineffective if
they are unreliable or fail to function correctly in their expected environment.
General SLAM algorithms are prevalent; however, localization algorithms specific
to a mine environment are somewhat less common. We are unaware of any work on
using ultrasonic rangers for robot localization in a mining environment. Some papers
use an artificial infrastructure, such as beacons [14] [29], reflectors (at LKAB’s Kiruna
Mine) [38], or surveying lasers (built into the mine) [35] for localization. However, an
artificial infrastructure does not exist in all mines, and even if one does exist, it may
not be intact after a mine disaster. To resolve these issues, many papers developed
localization algorithms that use natural mine features, such as walls [20], intersections
[1], and geometric beacons [17].
2.5 Map Building and Localization Using Scanning Laser Sensors in a
Mine Environment
When smoke, dust, or other visible particulates are not a concern, many research
efforts use scanning laser sensors, which supply plenty of data for scan matching and
localization algorithms. Scanning laser sensors are very precise, so they can capture
fine details of the surrounding environment; laser precision allows for algorithms such
astheIterativeClosestPoint(ICP)matchingalgorithm. Thoughthesesensorsarenot
used on the MineSENTRY project, some of the algorithms still apply. The following
research efforts use multiple scanning laser sensors for exteroceptive sensors, either
mounted front–and–back or orthogonally on the robotic vehicle.
A research paper from M¨akel¨a describes a means of localization in a mine [20] us-
ing scanning laser sensors; Load-Haul Dump (LHD) vehicles autonomously navigate
through a mine using the wall’s profile to correct for dead-reckoning drift. The algo-
15 |
Colorado School of Mines | rithm uses a scanning laser to measure the wall profile and compare that profile to
a pre-made “environment model,” which is built from range data compiled while an
operator manually drives a route. Assuming the wall profile is uniquely identifiable at
every location (a “fingerprint”), wall profile measurements can be used to determine
and correct vehicle location within a reasonable uncertainty using a Kalman-type al-
gorithm. Similar LHDs automate Sandvik Tamrock’s test mine in Tampere, Finland
[2].
The Groundhog vehicle (Section 2.2) also uses laser scan matching for localiza-
tion. However, Groundhogdoesnotuseana-priorimap; instead, ithas theadditional
difficulty of localizing while simultaneously creating a map of the mine (SLAM al-
gorithm) without using odometry—localization is entirely based on scan matching.
To accomplish this task, Thrun et al. used a Bayesian estimation technique, namely
Markov localization and a particle filter [8].
The SLAM algorithm begins with a scan matching algorithm that compares two
adjacent scans and identifies pairs of overlapping points. The algorithm then com-
putes the relative displacement and orientation by minimizing the distance between
all pairs of points. This matching algorithm is not perfect, and typically produces
errors in the form of parallel corridors (Figure 2.6).
Figure 2.6: Scan Matching, before erroneous path detection [36]
The parallel corridor is typical in a large scale cyclic environment, where a robotic
vehicleispronetohavinglargeuncertaintywhenre-traversingapathalreadytraveled;
the positional uncertainty causes the robot to make the correct data association with
the previous path, so the robot incorrectly assumes a parallel corridor. To account for
16 |
Colorado School of Mines | this issue, the Groundhog vehicle imposes Gaussian constraints on the displacement
and orientation computations of each successive scan, a representation known as
Markov random fields. When a conflict arises, such as an erroneous parallel hallway,
an iterative search questions past decisions in data association in order to increase
the probability of successive scans, resulting in the map in Figure 2.7. Ha¨hnel et al.
have a similar approach to address the data association issue in a large scale cyclic
environment; a probabilistic model of the residual errors from scan matching results
reduces uncertainty [11].
Figure 2.7: Scan Matching, after erroneous path detection [36]
In another research project, Bakambu et al. developed a system that has simi-
lar navigation capabilities as the Groundhog [1]. Bakambu’s robotic vehicle has two
modes: surveying and navigation. Surveying mode produces two and three dimen-
sional maps of the mine using artificial landmarks for localization and two orthogonal
scanning laser sensors for range measurements.
In navigation mode, a user sets a high level mission, in the form of waypoints on
an a–priori map (produced in surveying mode), and a motion planner reaches the
waypoints by translating the high-level missions into a set of consecutive navigation
actions. For localization landmarks, the robotic vehicle uses abrupt changes in the
local structure, such as corridors, intersections, and bays. To match the sensor read-
ings to landmarks, Bakambu uses point–to–line–segment matching (similar to Cox
[4]) instead of point–to–point matching (similar to Madhavan et al. [19]). Bakambu
notes that the matching algorithm degenerates in long drifts without distinguishable
features. Thus, matching results from these portions of the mine are ignored. This
17 |
Colorado School of Mines | problem is not a major concern for MineSENTRY purposes; radio attenuation occurs
around corners in a mine, so precise vehicle localization in long drifts (in between
nodes or adits) is not critical. In other words, until reaching an intersection or bend,
the AMR would only need to remain centered in the drift while moving forward or
backward, depending on RSS. Deciding which path to take upon reaching an inter-
section involves high-level path planning, beyond the scope of this thesis.
In another project using scanning laser sensors, Madhavan and Durrant developed
an outdoor map-building and localization algorithm that used a combination of an
iterative closest point (ICP) algorithm and an Extended Kalman Filter (EKF), to-
gether coined ICP–EKF [19] [18]. The ICP–EKF fuses Ackermann dynamics with
an ICP matching algorithm to localize an autonomous robotic vehicle relative to a
pre-made poly–line map (a map that consists of a series of connected line segments).
Figure 2.8 outlines the process.
Figure 2.8: Flow Diagram of the ICP–EKF Algorithm [18]
18 |
Colorado School of Mines | The ICP–EKF process is broken down into several steps. First dead reckoning
predicts the vehicle pose. Then, the ICP algorithm calculates the correspondence,
or how well the map fits the laser scan measurements. While proceeding forward,
the vehicle updates its position by fusing position estimates from Ackermann vehicle
dynamics (dead reckoning) with position estimates from correspondence data in an
EKF algorithm. The algorithm proceeds to the next scan iteration and the whole
process is repeated. See Madhavan et al. [19] for more details on the step-by-step
process.
Individually, Ackermann Dynamics and the ICP algorithm have shortcomings;
Ackermann Vehicle Dynamics is prone to drifting (like any other dead reckoning
method) and the ICP algorithm fails to form correspondences in areas without dis-
tinguishable landmarks (such as a long corridor or circular regions of a mine adit).
Fusing the two methodologies together dramatically reduces localization uncertainty,
allowing the robotic vehicle to accurately place itself on a map.
Localization using scanning laser sensors has been widely studied and is generally
considered a solved problem. Missing from the literature are localization algorithms
for sensors besides scanning lasers, for use in environments where laser-based sensors
fail. One such environment is a smoke-filled mine during an emergency rescue opera-
tion, driving the need for localization algorithm development using ultrasonics range
finders. Before addressing the issues that arise when using ultrasonic sensors, we first
expand upon the previous section by examining the use of scanning laser sensors in
an indoor environment.
2.6 Localization and Matching Algorithms in Indoor Environments
Many research efforts focus on localization and SLAM algorithms in an indoor
environment. Like sensor selection, environment plays a major role in localization
algorithmdevelopment; analgorithmdesignedforindoorsmostlikelywillnotworkfor
19 |
Colorado School of Mines | anundergroundminingenvironment. Thoughnoresearcheffortshavebothultrasonic
rangefindersandaminingenvironment, thealgorithmsdevelopedandlessonslearned
for indoor environments can partially apply to the MineSENTRY project.
2.6.1 Matching Using a Laser Sensor in an Indoor Environment
Adapted by several research efforts, including Bakambu et al. [1], Ingemar J.
Cox was one of the first to develop a point–to–line–segment matching algorithm [4].
Intended for a cost-effective means of navigation in an office environment, Blanche
is a non-holonomic vehicle that uses a single rotating infrared ranger and odometry
data to form a map of the surrounding environment. The robot gathers 180 range
points (per full revolution) from the ranging sensor and estimates the wall position
relative to a universal or local reference frame. Using an a-priori map, the robot can
then match its sensor readings to the map using an iterative least squares matching
algorithm. Beginning with the range data and environment map in Figure 2.9, Cox
developed a matching algorithm to correct the dead reckoning drift.
Figure 2.9: Blanche Iterative Least Squares Matching Algorithm Scenario [4]
Dead Reckoning drift has skewed the robot’s position estimate both in both carte-
sian position, (x,y), and orientation, θ; rotating the collected points clockwise and
20 |
Colorado School of Mines | moving the points down and left would match the wall. To correct the drift, Cox uses
an iterative matching algorithm, broken down into several discrete steps [4]:
1. For each point in the image, find the line segment in the model that
is nearest to the point. Call this the target.
2. Find the congruence that minimizes the total squared distance be-
tween the image points and their target lines.
3. Move the points by the congruence found in Step 2.
4. Repeat step 1-3 until the procedure converges. The composite of all
the step 3 congruences is the desired total congruence.
For every point in an image (collected range data), the algorithm finds the closest
line segment in the model (a–priori map). The algorithm then iteratively rotates and
translates the image until the image and the model line up as well as possible. Cox
concedes that the problem with least squares is its sensitivity to outliers. However,
eliminatingoutliersmakesthisalgorithmrobust,notingthat“minimizingtheabsolute
standard deviation would probably lead to a more inherently robust algorithm [found
in [12]]. However, it is computationally more expensive” [4].
An indoor environment assumption imposes a few restrictions not applicable to an
underground mining environment. However, these restrictions mostly apply to map
building. Matching the estimated map to the a-priori map still applies to a mining
environment, and hence, this algorithm can still prove useful for this project, despite
its indoor (and laser) origins.
Before a matching algorithm can be successful, ultrasonic sensors have numerous
disadvantagestoovercome; comparedtoascanninglasersensor, anarrayofultrasonic
sensors produces only a fraction of the data, at a slower rate, and with less precision.
Despite this fact, range measurements from ultrasonic sensors can still be used to
build maps and localize using a wall profile. However, the map produced and data
association have a larger uncertainty compared to those produced by laser range data.
21 |
Colorado School of Mines | The following section addresses the issues that arise when using an ultrasonic range
finder for localization.
2.6.2 SLAM Using Ultrasonic Ranging in an Indoor Environment
In an attempt to compensate for the relatively large uncertainty of ultrasonic
sensors, Crowley describes a means of localizing by forming line segments from three
consecutive ultrasonic readings [6]. Each line segment has an associated certainty,
called the ‘certainty factor;’ the certainty factor is initialized to a value of one and
any subsequent sensor data verifying the line segment adds one to the certainty factor
up to a maximum of five. This method eliminates stray sonar readings, but assumes
flat surfaces typically found in indoor environments. This method could be applied
to a mining environment to discover side passages, but would model the long tunnel
as a very long, flat wall, defeating the purpose of localization.
Crowley’s matching algorithm is primarily based on vehicle and line segment
(formed from ultrasonic sensor readings) uncertainties. Initially, for each measured
line segment, Crowley computes its orientation (θ), its perpendicular distance to the
origin (c), and uncertainty values for θ and c (σ and σ ). Before matching the line
θ c
segment to the composite model, the vehicle position uncertainties are added to the
line segment uncertainties, given by
σ2 = a2σ2 +2abσ2 +b2σ2 (2.1)
r x xy y
σ2 = σ2 +σ2, (2.2)
c,new c,old r
σ2 = σ2 +σ2, (2.3)
θ,new θ,old α
where a = sin(θ), b = −cos(θ), σ2 is the variance of the vehicle’s distance to the
r
origin, and σ2 is the variance of the vehicles orientation. If the observed segment
α
22 |
Colorado School of Mines | matches (within a computed uncertainty) the orientation, alignment, and overlap of
the composite model, the vehicle position and composite model are both updated.
Due to their wide beam width, ultrasonic sensors have more uncertainty in their
orientation than in their actual range measurement. Thus, Crowley simplifies error
propagation by approximating sensor covariance with one value: σ , the variance in
w
the orientation of the sensor. Crowley determined variance of the particular sensor
he used by experimental calibration: σ = 0.1+d∗tan−1(5◦), where d is the distance
w
reportedfromthesensor. Inasimilarfashion, theuncertaintyinverticaldisplacement
of a range measurement (due to vehicle tilting) can also be included in σ . In other
w
words, a wide beam width can be advantageous as ultrasonics sensors are robust to
vehicle tilting.
2.7 Other Related Work
A missing element from the previous sections was an explicit method for calculat-
ing the measurement covariance of an empirical model (i.e. the matching algorithm)
in a Kalman filter. This last section highlights a method for estimating a model’s
measured covariance, from a rather unexpected source.
Kolter et al. [15] developed a multi–model Linear Quadratic Regulator (LQR)
controller to autonomously slide a vehicle sideways into a parking space. The LQR
control interpolates between two separate models for vehicle dynamics: the first mod-
els the vehicle traversing in a straight line and the second models the vehicle while
sliding. The first model is well–described mathematically. However, due to stochastic
environmental parameters (e.g. friction), the second model is extremely difficult to
predict. Thus, an approximate model estimates the states while the vehicle is sliding;
the second model is a learned behavior from a human operator executing a single
sliding maneuver. Individually, each model fails to place the car accurately in the
stunt maneuver. However, when the models are fused together, the car is placed
23 |
Colorado School of Mines | consistently within two feet of the desired trajectory.
Fusing the two models in an LQR controller is analogous to fusing measured and
predicted data in an EKF algorithm. An important aspect of empirical modeling (be
it a car drifting sideways or observing a mine wall) is estimating the model’s true
covariance. In practice, the true covariance is impossible to obtain; Kolter et al. ap-
proximate the covariance of the second model by computing the error of the model’s
predictions and averaging over a small time window. This technique effectively low-
ers the covariance in neighboring points of the trained data set, making the sliding
vehicle more robust to deviations from the training set. Though the multi–model
LQR controller is not a mine rescue vehicle or a SLAM algorithm, the technique for
approximating the covariance matrix can be applied to the measurement update of
the EKF algorithm, making the matching algorithm more robust to sensor noise.
2.8 Background Summary
Despite the higher uncertainty and lower data rate of ultrasonic sensors compared
to laser sensors, their ability to accurately range in smoke, and therefore emergency
situations, overcomes the disadvantages. Though no research focuses on using ultra-
sonic sensors for localization in a mine, research efforts that developed localization
algorithms using scanning laser sensors, as well as research for indoor environments
and other related work, can help produce similar algorithms for ultrasonic range find-
ers. Before developing the localization algorithm, a vehicle is required for testing; the
next chapter covers the vehicle development.
24 |
Colorado School of Mines | CHAPTER 3
VEHICLE DESIGN
Using some of the ideas from Sections 2.2 and 2.3 in the previous chapter, this
chapter focuses on the electronics and sensors involved in converting a stock EZ-GO
TXT1 golf cart into the Autonomous Mobile Radio (AMR). The mechanical portion
and communication protocols are only briefly discussed: for further information, refer
to Hulbert [13]. For Information on the wireless tethering, refer to Meehan [22].
The overall purpose of the AMR is to relay a radio signal in an underground mine
during an emergency scenario. Thus, the vehicle must be large enough to carry a
radio and camera, rugged enough to traverse a typical underground mine, and have
the battery capacity to last the duration of the mine rescue operation. Furthermore,
the AMR must be capable of navigation without assistance, which requires numerous
sensors and computer algorithms. Each of these design requirements are addressed,
starting with the initial design decisions and moving on towards vehicle construction.
3.1 Design Alternatives
As with any design project, there are numerous design alternatives. Two of the
mostsignificantdesigndecisionsinvolvedroboticvehiclesizeandexteroceptivesensor
choice. Thefollowingbrieflydescribesthealternativestoeachofthetwomajordesign
choices, along with the reasoning behind the final decisions.
3.1.1 Vehicle Size
Because most design decisions depend on vehicle size, it is the first decision to
make. The vehicle size was broken down into two options: large (capable of trans-
porting a human passenger) and small (capable of transporting the necessary elec-
tronics only). Smaller vehicles have the advantage of fitting in tight places, mobility,
25 |
Colorado School of Mines | and portability. However, rocks, boulders, and even cart tracks become major obsta-
cles for these smaller vehicles. Additionally, without a tether, these vehicles have a
relatively short range because their ability to store energy (e.g. a battery) is limited.
Larger vehicles can easily overcome some of the obstacles that are problematic for
smaller vehicles. Even without tethering, larger vehicles have a relatively long range,
with the added ability to haul heavy equipment, injured miners, etc. However, these
vehicles can get stuck in narrow passages and areas with low ceilings.
In an early attempt to determine proper size, a small radio control vehicle was
tested at the Edgar Mine. Measuring 20” x 16” x 10” (Length x Width x Height) and
weighing about 10 pounds, the vehicle managed to traverse over most obstacles, but
struggledwhenithigh–centeredonacartrail; the3”groundclearancewasinsufficient
for the 6” rails.
Due to the rigors of a mine environment (Section 2.1), and the ability to transport
persons and equipment, larger vehicles are preferred in the mining industry. Thus
the vehicle of choice is a TXT1 EZ–GO golf cart. This vehicle has sufficient room
for necessary electronics and sensors, the radio and camera, six large 6V lead acid
batteries, and 2 passengers.
3.1.2 Sensor Choice
Another important decision driving the vehicle design is the sensor choice; based
on Section 2.3 in the previous chapter, the choice of sensors for emergency situations
is somewhat limited to ultrasonic sensors. There are a few interesting alternatives,
including a scent sensor (GUARDIANS project [31]); however, this sensor requires a
‘unique’ scent to follow, which is uncommon in a mine. Other algorithms use a hybrid
approach using both scanning lasers and ultrasonic sensors [34]. Though, if smoke
renders scanning laser sensors unusable, a localization algorithm without scanning
lasers (i.e. ultrasonic sensors alone) would be beneficial. Since laser range finders are
26 |
Colorado School of Mines | unreliableinadverseconditions, thevehiclemusthavetheabilitytorangeandlocalize
using ultrasonic sensors alone. Thus the AMR has an array of 14 ultrasonic sensors
surrounding the vehicle, providing both soft bumper and localization capabilities.
Given the design decision to use ultrasonic sensors, and the decision fromthe previous
section to use a large vehicle, the remainder of the chapter is dedicated to building
the vehicle platform.
3.2 AMR Construction
The MineSENTRY project developed a vehicle for the conditions at the Edgar
Mine in Idaho Springs, Colorado. Historically, the mine produced metals such as
silver, gold, lead, and copper. Today, the mine serves educational purposes, providing
a touring venue for public and school groups, a laboratory for underground research,
and a training facility for future engineers [28]. The mine primarily consists of dry,
loose rock; thus, for a proof–of–concept vehicle, explosions and water-proofing are not
a concern.
This section includes the documentation of the model numbers, calibration, and
placement of the sensors and electronic components used to make the AMR. The
vehicle platform is an EZ–GO TXT1 golf cart; thus, the terms ‘AMR’ and ‘golf cart’
are used interchangeably. The order of this section roughly corresponds to the order
in which the AMR was built. Beginning with the stock golf cart, the first step was to
addhardware, suchasactuationandsensors, includingsensorcharacterization. When
all of the hardware was completed, the golf cart was ready for a computer. The next
step was developing the software to run the hardware, which includes the computer
architecture, the microprocessor platform, and the microprocessor code. The final
step was to wire the computer to the hardware components, with careful regard to
isolation, voltage regulation, power distribution, and noise. The engineering safety
controls is the last subsection covered, as numerous safety concerns were discovered
27 |
Colorado School of Mines | Modulation (PWM) signal from the robotic controller into an analog voltage; a larger
PWM signal produces a smaller analog voltage signal, resulting in more power at the
wheels. This circuit is wired in parallel with the stock golf cart electronics (‘Throttle
Box’), so an operator can command more throttle at any time. If an emergency stop
is required, an operator can use the brake to slow the vehicle; the external brake
depression is detected and the throttle is turned off.
Figure 3.3: Throttle Actuation Circuitry
Furthermore, the computer detects throttle-position feedback from a built-in mi-
croswitch (see Figure 3.2); power is not activated unless the pedal is depressed, re-
quiring bypass circuitry, or a relay. Lastly, the direction (forward and reverse) is
activated by a switch on the golf cart’s front panel, requiring an additional set of
relays. A custom-made circuit board contains the throttle control circuitry, along
with the three optically isolated relays. See Section A.1 in the appendix for the over-
all circuit schematic and the color code for wiring into the golf cart computer. The
throttle microswitch relay is a single Panasonic Electric Works AQV212 IC (labeled
’AQV212’ on the schematic), and the FWD/REV switch relays are contained in the
single Panasonic Electric Works AQW212 IC (labeled ’AQW212’ on the schematic).
Similar to to the throttle control circuit, all relays are wired in parallel to the existing
switches on the golf cart, allowing an operator to override an open relay at any time.
30 |
Colorado School of Mines | The golf–cart computer defaults to reverse if both the forward and the reverse relays
are on simultaneously.
3.2.2 Propreoceptive Sensors
ThereareatotalofsixpropreoceptivesensorsontheAMR:aHallEffectsensoron
each wheel to measure odometry and calculate vehicle speed, a string potentiometer
on the steering arm to measure the Ackermann steering angle, and a five degree of
freedom Inertial Measuring Unit (IMU) to measure vehicle orientation. Combined,
these sensors provide steering angle, velocity, and yaw, all necessary to compute
Ackermann vehicle dynamics.
The Hall effect sensor on each wheel reads from 12 magnets, making the odometry
resolution118mm(tirecircumferencedividedbythenumberofmagnets). Thesensor
sends a 3.3V pulse to the microcontroller when a magnet passes, and in turn, the
microprocessor increments the odometer count by 118 mm if the forward switch on
the golf cart is engaged, or decrements the odometer if the reverse switch is engaged.
The speed is then updated with a simple difference equation and a lowpass averaging
filter:
d −d
n n−1
S = , (3.1)
n
t −t
n n−1
11
1 (cid:88)
S = S . (3.2)
average,n n−i
12
i=0
where n represents a fixed time interval, d is the current odometry reading in {m}
n
(theodometryreadingisupdatedaftereachmagnetpass), tistheelapsedtimein{s},
S is the vehicle’s speed in {m/s}, and S is a twelfth–order moving average filter
average
of S. If no magnet pass is detected during the fixed time interval (i.e. the vehicle is
moving slowly or stopped altogether), the numerator of Equation (3.1) becomes zero.
After 12 subsequent time intervals without a magnet detection, S becomes
average,n
31 |
Colorado School of Mines | zero.
For steering angle measurements, a Celesco SP1 string potentiometer is used.
This sensor is fixed to the steering rack and outputs a 0-3.3V proportional to its 0-6”
range, so the steering angle is approximately proportional to the string potentiome-
ter’s output voltage; Figure 3.4 is the voltage to steering angle calibration curve. The
conversion from voltage to steering angle is θ = −31.1 · V + 47. This calibration
was performed using two long straight edges: one aligned with the rear tire and one
aligned with the front tire. The angle between these two straight edges is the steer-
ing angle, and observing the voltage for every 5–10 degrees of steering throw gives
the data for a linear regression. Note that a positive steering angle is one that in-
duces a counter–clockwise rotation of the vehicle when viewed from above (i.e. left).
Calibration for all analog sensors can be found in Section A.2
Figure 3.4: String Potentiometer Voltage to Steering Angle Calibration
An Inertial Measurement Unit (IMU) detects the vehicle’s acceleration and ro-
tation. The IMU is a board from Sparkfun(cid:13)R Electronics. This board incorporates
32 |
Colorado School of Mines | units, which have a wider beam width for better obstacle detection, making the EZ1
model best suited as soft bumpers. The sensors oriented towards the wall are LV-
EZ3 units, which have a narrower beam width, and therefore, more accurate ranging
measurements compared to the EZ1; thus, the EZ3 model is well suited for building a
map of the environment. Refer to Table Table 3.1 for a comparison of EZ1 and EZ3
sensors. Furthermore, Section 3.2.4 elaborates on the beam width characteristics.
Table 3.1: List of Sensor Model Numbers
Sonar Sensor Model Beam Width Use
1, 3, 7, 9 Maxbotix LV-EZ0 Wide Soft Bumpers
All Others Maxbotix LV-EZ3 Narrow Wall Mapping
Each sensor has a 50 ms update rate, so triggering one sensor at a time would
take 0.7 seconds, corresponding to a 1.4 Hz update rate. When controlling a vehicle,
a quicker update rate is more desirable; to speed up the update rate, the 14 sensors
are broken up into four triggering groups. Each triggering group has between two
and four sonars; sensors within each group are triggered simultaneously. Refer to
Table 3.2 for the list of sonars in each triggering group.
Table 3.2: Sonar Triggering Groups
Trigger Group Sonars in Group
1 1, 2, 7, 8
2 3, 4, 9, 10
3 5, 6, 11, 12
4 13 and 14
Thesetriggeringgroupsaredaisychainedtogether. Apulsesignalfromtherobotic
controller sets off the chain. When each triggering group finishes collecting range
measurements, a pulse signal is sent to the next triggering group, until the final pulse
fromthefourthtriggeringgroupsignalstheroboticcontrollertostartanalogtodigital
conversion. Each triggering group takes approximately 50 ms to collect data, making
34 |
Colorado School of Mines | the total update rate 5 Hz. To avoid interference, each sensor is oriented 90◦ apart
from other sensors in the group, as illustrated in Figure 3.5. Measurements from
these sensors are used to build an environment model of mine walls.
Inadditiontothesonarsensors, fourSharpInfraredlong-rangesensorsaremount-
edatallfourcornersofthevehicleat45◦ angles. Theseinfraredsensorswereoriginally
intended to serve as backups to the diagonal sonar sensors (sensor numbers 5, 6, 11
and12),andtoprovideadditionalrangemeasurementsintheeventthattheultrasonic
sensors return inconsistent data during testing. However, these sensor were damaged
upon transportation to the mine, so they were not used.
3.2.4 Sonar Characterization
A single wall–profiling EZ3 sensor was tested for its mean and standard deviation
in both indoor and underground mine environments. During each test, the sensor was
first oriented directly at the wall (0◦), at one meter away (measured with measuring
tape). The sonar’s voltage measurements were recorded at 200 Hz for 10 seconds.
The manufacturer’s voltage conversion equation was used to report the range mea-
surements: d = V · [512/V ] · 0.0254, where V is the power supply voltage
measured cc cc
(3.3V), V is the sensor’s voltage output, and d is the range measurement in
measured
meters. For each environment, the data was recorded again at two meters away from
the wall, making a total of four experiments: indoors at one meter and two meters,
and in the mine at one meter and two meters.
The mean (µ ) and the standard deviation (σ ) of the range measurements for
0 0
the four experiments are reported in Table 3.3. With a mean near the measured
distance from the wall and a standard deviation under 3 cm, the sensors are fairly
accurate and precise. Because of the smoother walls, the standard deviation indoors
was slightly better compared to the mine.
35 |
Colorado School of Mines | Table 3.3: Sonar Characterization, θ = 0◦
Location Distance from Wall {m} µ {m} σ {m}
0 0
Indoors 1.0 0.96 0.014
Indoors 2.0 1.9 0.010
Edgar Mine 1.0 1.0 0.026
Edgar Mine 2.0 2.0 0.017
The sonar sensor was also characterized for its critical angle (θ ). Starting at
critical
0◦ (perpendicular to the wall), the sensor was rotated 10◦ at a time until it was no
longer able to detect the wall, at which point the sensor angle was reduced by 5◦ and
re–tested. The maximum angle the sensor can be rotated without losing the signal
from the wall is the critical angle (θ ) . The mean (µ ) and standard deviation
critical c
(σ ) of the range measurements for each scenario at the critical angle are given in
c
Table 3.4.
Table 3.4: Sonar Characterization, θ = θ
critical
Location Distance from Wall {m} θ {◦} µ {m} σ {m}
critical c c
Indoors 1.0 25 1.0 0.66
Indoors 2.0 25 2.2 0.98
Edgar Mine 1.0 45 1.2 0.47
Edgar Mine 2.0 25 1.9 0.18
The irregularities of the mine wall are nearly the same size as the sonar sensor’s
emitted wave packet, leading to a scattering effect of the acoustic wave. Theoretically,
the scattering effect would allow the sensor to detect the mine wall more easily,
meaning a higher critical angle and a lower standard deviation of the range data
the mine. In practice, the critical angle increased only when near the mine wall
(<1m), and σ was marginally lower in the mine compared to indoors, though σ
c c
was higher across the board compared to σ (Table 3.3). With the lower σ and
0 c
the potentially higher critical angle, the mine offered a slight advantage in sensor
performance compared to indoors. However, regardless of the sensor environment,
36 |
Colorado School of Mines | the ultrasonic sensors are unreliable at angles greater than 25◦.
A graphical representation of one sonar characterization experiment is illustrated
in Figure 3.6. This experiment took place at the Edgar Mine with the sonar range
finder rotated through various angles, two meters away from the mine wall. The
dotted lines show 10◦ increments, the solid black line is a rough approximation of the
mine wall, and the blue dots represent the sonar range data. The sonar range data is
fairlycompactwhenpointeddirectlyatthewall(θ = 0◦),buttherangemeasurements
become more widely dispersed as the angle increases. Figure 3.6 clearly delineates the
limit for the sensor angle; beyond θ = 25◦, the the sensor has difficulty detecting
critical
the wall, and the range readings vary widely from about two meters to six meters.
These finding fully support the data in Table 3.3 and Table 3.4.
Sonar Characterization
Edgar Mine Army Tunnel
6
Mine Wall
Sonar Data
5
4
3
2
1
0
0 1 2 3 4 5 6
X (meters)
Figure 3.6: Sonar Characterization in the Edgar Mine (Overhead View)
The sensor orientations of Sonars 5, 6, 11, and 12 in Figure 3.5 are 45◦. This is the
minimum angle to avoid interference among sensors. Admittedly, the original sonar
characterization took place in a laboratory with numerous acoustically reflectively
surfaces, so the sonars were originally thought to have over a 45◦ critical angle. The
near–wall critical angle supported this claim, but, as indicated in Figure 3.6, the far
37
)sretem(
Y |
Colorado School of Mines | wall critical angle was found to be far lower (25◦). Discovered long after the data was
collected, and requiring a significant hardware change, the rest of this thesis assumes
a 45◦ critical angle. This issue is addressed in Section 5.4.
The data in the mine was taken across a relatively straight portion of the mine
adit, representing the worst–case scenario; typically, the mine has irregularities that
give the ultrasonic sensors a stronger signal. In fact, the sensor directed toward
one such irregularity at a 40◦ angle yielded a 0.076m standard deviation from 2.8m
away. However, the sensor lost the signal altogether when rotated by 5◦ in either
direction. In certain situations, the mine environment can improve the critical angle
and the confidence (standard deviation) of the range measurements, but in general,
the ultrasonics are no more reliable in a mine than they are indoors.
Another characteristic of the ultrasonic sensors is evident in the mean; for certain
scenarios, µ is the same as µ , despite being rotated by 25◦ or 45◦. Due to the
c 0
wide beam width, the location of the object (point on the wall) reflecting an echo
is uncertain, causing the range measurement to repeat through sensor rotation. For
example, anultrasonicsensorwitha10◦ beamwidthmaydetectthesameobject(and
therefore report the same distance) when rotated from 0◦ all the way to 20◦, creating
a detection pattern in the shape of an arc. In other words, neither the distance nor
the angle are necessarily accurate.
The LV–MaxSonar(cid:13)R–EZ3TMdatasheet [21] describes the beam width—or rather,
the detection pattern—of the sonar sensors in more detail. Figure 3.7 is the detection
pattern of an EZ3 sensor for a 3.5(cid:48)(cid:48) diameter rod for 5V (black line) and 3.3V (red
dots) on a 12(cid:48)(cid:48) grid. The detection pattern is highly dependent on the reflecting
object, but a 3.5” diameter rod is a fair approximation for a mine wall; the round
surface disperses acoustic waves in several directions and the rod is large enough to
be detectable at distances expected in an underground mine.
For the 3.5(cid:48)(cid:48) rod, the detection pattern is about 2ft wide, or ±0.3m from the
38 |
Colorado School of Mines | center. Given that the shape and reflectivity of a detected object are typically un-
known, the standard deviation of the EZ3 sonar sensors can be over–estimated as
σ = 0.3m. Similar to Crowley’s [6] method for approximating the sonar error,
EZ3
σ is extended in both the axial and perpendicular directions for simplification;
EZ3
the maximum axial standard deviation (Edgar Mine Location at 25◦ in Table 3.4)
and the perpendicular standard deviation (sensor beam width, Figure 3.7) are both
near 0.3 m. Thus, the wall–profiling sensors have a 68% confidence interval (±1σ )
EZ3
that the actual object is within ±0.3m of the detected position. At 3m away, this
corresponds to about ±5◦.
Figure 3.7: Sensor Detection Pattern [21]
3.2.5 Mobile Radio Components
Because the radio is a standalone system, requiring only a power connection to
charge the built–in battery, it was the last hardware component added to the AMR,
so it is briefly described here. Without the mobile radio, the AMR would simply be
’Autonomous.’ The radio workhorse is a Rajant radio, mounted on top of the AMR’s
electronics housing, shown in Figure 3.8. Each Rajant radio automatically connects
to nearby Rajant radios to form a wireless mesh network, robust to the dynamic
nature of the caravan scenario depicted in the introduction to this paper. The nature
of these radios is beyond the scope of this paper; for more information, see the thesis
by Meehan [22].
39 |
Colorado School of Mines | Figure 3.8: Rajant Radio and Mounting Location on the AMR.
This concludes the vehicle platform hardware section, which covered actuation,
sensors, andsensorcharacterization. Atthispoint, thegolfcartcanonlybecontrolled
manually. The remaining sections cover switching between manual, remote, and
computer–controlled modes; adding the computer and software, which includes the
computer architecture, the microprocessor platform, and the microprocessor code;
power distribution; and engineering safety controls.
3.2.6 AMR Operation Modes
Transitioning from manual control to computer control requires a switch—in this
case, a software-monitored mechanical mode switch. The AMR has three user-
switchable modes of operation: Manual, Remote, and Autonomous. Manual mode,
as the name implies, is intended for manual, human operation. An operator ensures
that the steering motor is disconnected, and the AMR is ready for manual ’golf cart’
driving mode. To return to the other two modes, a human operator must rotate the
mode switch and re-connect the servo motor to the steering column.
The remaining two modes utilize the on–board computers, which are described in
the following section. Remote mode is intended for driving the AMR using a remote
40 |
Colorado School of Mines | by the RoboteQ AX3500 control board command set and configuration, summarized
in Section A.4.
The communication packet structure between the robotic controller and the au-
tonomous controller is also in the Appendix (Section A.5 on Page 97 and Section A.6
on Page 99). These packet structures standardize the communication protocol for
error handling, sensor data, and vehicle status.
3.2.8 Microprocessor Platform
Originally, the microprocessor was purchased based on slightly different interface
requirements than depicted above: one RS232 port, one RS485 port, 15 analog to dig-
ital converter channels, four digital inputs, one digital output, one input capture pin,
one PWM channel, and one Universal Serial Bus (USB) port. The FLEX platform—
the product of two Italian companies, Evidence Srl and Embedded Solutions Srl—was
one of the only platforms to meet the criteria, particularly because the RS485 and
USB combination is rare among microprocessor development platforms. The FLEX
full board (Figure 3.11), when combined with the multi-bus daughterboard, is fully
customizable with various Controller Area Network (CAN), Universal Asynchronous
Receiver/Transmitter (UART—RS232 and RS485), Serial Peripheral Interface (SPI),
and ethernet communication modules.
Figure 3.11: FLEX Platform Base Board [7]
43 |
Colorado School of Mines | Interrupts are used for important time-dependant events, such as data collection,
RS232 communication, and the emergency stopping procedure. All other tasks, such
as the watchdog timer, command processing (servo set point calculations), and data
processing (conversion to engineering units), are handled using polling.
The code, in its entirety, is in Section A.7 of the Appendix, along with companion
flow charts in Section A.8 to help decipher the important parts of the code. For
information on the code in the autonomous controller, see Hulbert [13].
This concludes the software section, which covered the computer architecture, the
microprocessor platform, and the microprocessor code. The following sections discuss
how the computer and the remaining hardware are tied together.
3.2.10 Power Distribution and Wiring
With the sensors, actuators, drive motor, computers and mobile radio equipment,
powerdistributionbecomesanimportantissue. Topowerthewheelsandtheon-board
computer, the golf cart relies on six 6V lead acid batteries, wired in series, for a total
of 36V. Due to their high capacity, these main batteries also feed the brake actuator
and the steering servo (after passing through an isolated 24V industrial regulator).
Because the main battery bank tends to be electrically noisy while the golf cart motor
is turning, all other components added to the stock golf cart configuration receive
power from an isolated, standalone 12V lead acid battery with voltage regulation for
a 5V and a 3.3V tap. Table 3.5 lists the electronics on the AMR, along with their
power requirements.
Power distribution to all 14 ultrasonic sensors, four IR sensors, four Hall Effect
sensors, and one string potentiometer requires copious amounts of cabling. Thus, two
power distribution boards were designed to reduce the number of cables and clutter
in the main electronics housing. The sensor connector locations and the pinouts for
both boards are located in Section A.9.
45 |
Colorado School of Mines | Table 3.5: AMR On Board Electronics Power Requirements
Component Input Voltage Power Source
Golf Cart Computer 36V 36V Main Battery Bank
Golf Cart Drive Motor 36V 36V Main Battery Bank
Brake Actuator 24V 36V Main Battery Bank/ 24V
Regulator
Steering Servo Motor 24V 36V Main Battery Bank/ 24V
Regulator
AX3500 Servo Controller 24V 36V Main Battery Bank/ 24V
Regulator
FLEX Microprocessor Board 12 V 12V Auxiliary Battery
(robotic controller)
Rajant Radio 12 V 12V Auxiliary Battery
Ethernet Router 12V 12V Auxiliary Battery
Ultrasonic Sensors 3.3V 12V Aux. Battery/ 3.3V Reg-
ulator
Infrared Sensors 3.3 V Aux. Battery/ 3.3V Regulator
String Potentiometer 3.3V Aux. Battery/ 3.3V Regulator
Hall Effect Sensors 5V Aux. Battery/ 5V Regulator
Throttle Control Circuitry 5V Aux. Battery/ 5V Regulator
FWD/REV/Thottle Relays 5V Aux. Battery/ 5V Regulator
With up to 15 feet of cable routing analog signals from the sensor to the robotic
control, electromagnetic noise from the golf cart motor became an issue. In an at-
tempt to attenuate the noise, the control battery, all power supplies, voltage regu-
lating circuits, sensors, and cable shielding were grounded to a single point in the
electronics housing, as shown in Figure 3.13. Despite the star grounding, some noise
is still present, both in the signal and the power supply of the sensors. At the recom-
mendation of the manufacturer, a low-pass filter was implemented on each ultrasonic
sensor’s voltage input. The low-pass filter is a first-order Resistor-Capacitor (RC)
filter with R = 100Ω and C = 100µF, giving a cutoff frequency of 16 Hz.
Still having problems with noisy sensor readings, additional RC filters were in-
stalled on all affected sensor signals, including all analog sensor lines and the 4 Hall
Effect sensor lines; with R = 100Ω and C = 1µF, each line has a filter with a cut-
off frequency of 1.6 kHz. This cutoff is sufficiently high to leave the normal sensor
46 |
Colorado School of Mines | Figure 3.13: AMR Star Ground
readings unaffected, yet low enough to block the noise from the golf cart motor.
With the wiring and power distribution completed, the AMR is fully capable of
teleoperation. The last section outlines the safety controls developed as the teleoper-
ation capabilities were tested.
3.2.11 Engineering Safety Controls
SafetywasparamountthroughouttheAMR’sdesignandconstruction; autonomous
vehicles can be a serious danger, especially during preliminary testing, resulting in
damaged toolboxes, equipment, property, or injured persons. Thus, numerous safety
measures were implemented during development.
First, there are two emergency off (EMO) switches located at the front and rear
of the vehicle. The robotic controller detects these switches in a high-priority inter-
rupt service routine and immediately shuts off throttle and engages the brake when
the EMO switch is depressed. The autonomous controller can also send a software
emergency stop, which behaves similarly to the physical EMO switches.
Second, when changing to remote mode (see Section 3.2.6), the robotic controller
engagesthebrakeandblocksthrottleandsteeringcommandsuntilitdetectsaspecific
startup procedure from the operator. To unlock the robotic controller, the operator
must, while the brake is engaged on the transmitter, toggle the reverse switch on the
47 |
Colorado School of Mines | CHAPTER 4
LOCALIZATION ALGORITHM DEVELOPMENT
This chapter covers the equations and algorithms involved in localization. To-
gether, the Ackermann vehicle dynamics, the matching algorithm, and the extended
Kalman filter form the overall localization algorithm. This localization algorithm is
based on the scanning laser sensor research efforts in order to determine if these al-
gorithms can cope with the inherent problems of ultrasonic sensors. Where problems
arise, new or modified algorithms are proposed.
Below is an overview of the localization algorithm. This chapter elaborates on
each step.
1. Obtain an environment map of the mine wall. The assumption is that an
accurate a–priori map is not provided, so the environment map is built from an
initial pass of the mine using dead reckoning data.
2. Run the localization algorithm. The localization algorithm is based on the
ICP–EKF algorithm developed by Madhavan et al. [19].
(a) Initialize the vehicle.
(b) Perform the EKF prediction step, which uses Ackermann vehicle dynamics
to predict the vehicle pose.
(c) Perform the EKF measurement update, which corrects the dead reckoning
drift using the sonar range readings in an iterative matching algorithm.
(d) Iterate the algorithm while the vehicle is moving.
The list above is re–visited in more detail at the end of the chapter.
4.1 Ackermann Vehicle Dynamics
Similar to most automobiles, the AMR has Ackermann steering; Ackermann ve-
hicle dynamics mathematically describe a means of localization via dead reckoning:
51 |
Colorado School of Mines | V = S ·cos(θ), (4.1)
x
V = S ·sin(θ), (4.2)
y
dθ 1
= ·S ·tan(α) [3], (4.3)
dt L
where V and V are the horizontal and vertical components of the vehicle speed, S;
x y
θ is the heading of the vehicle, defined as positive in the counter-clockwise direction;
α is the steering angle, also defined as positve when turning counter-clockwise; and
L is the vehicle wheelbase. V and V can be integrated to give position, (X,Y).
x y
Figure 4.1 is a graphical illustration of these variables.
Figure 4.1: Ackermann Vehicle Dynamics Variables
Ackermann steering dynamics estimate the entire vehicle as a point at location (X,Y)
oriented in direction, θ, analogous to a ‘virtual’ wheel located directly between the
front wheels. The location of (X,Y) on the vehicle is arbitrarily defined to coincide
with the ‘virtual’ wheel.
Equations (4.1) through (4.3) are integrated in the discrete time domain. At each
time iteration (k, or every 250 ms), Equations (4.4), (4.5), and (4.6) compute X, Y,
and θ, respectively:
52 |
Colorado School of Mines | X = X +(d −d )·cos(θ ), (4.4)
k k−1 k k−1 k−1
Y = Y +(d −d )·sin(θ ), (4.5)
k k−1 k k−1 k−1
1
θ = θ + ·(d −d )·tan(α ), (4.6)
k k−1 k k−1 k
L
where d is the total distance traveled at time iteration number k. The expression
k
d −d , or the distance traveled during one time iteration, occurs frequently, so will
k k−1
henceforth be referred to as dd .
k
4.2 Map Building
Combined with the vehicle path defined by Ackermann dynamics, readings from
the 14 ultrasonic range sensors are used to create a map, or environment model, of the
mine walls. To do this, consider the illustration in Figure 4.2. The AMR starts at an
arbitrary location, typically at the beginning of the mine. The universal coordinate
system, {U}, is defined as shown with the origin centered at the midpoint of the front
axle (refer to Section 4.1 on Page 51).
Figure 4.2: Wall Mapping Scenario: AMR Starting point
53 |
Colorado School of Mines | TheAMRdrivesanarbitrarypath,withanendinglocationasshowninFigure4.3.
Frame {U} remains fixed relative to the earth, while the vehicle frame, {V} travels
with the vehicle. The distance and direction that the AMR traveled is represented
by the vector UP , with orientation, θ. Note that θ is defined positively in the
VORG
counter-clockwisedirection; anegativeangleisasignconventiontodenoteaclockwise
direction. Theultrasonicsensor’scoordinateframe, {S}isfixedrelativetothevehicle,
though translated and rotated relative to Frame {V}.
Figure 4.3: Wall Mapping Scenario AMR Finishing Point
To illustrate the coordinate system transformation equations, see Figure 4.4. This
particular illustration uses sonar 8 as an example, but the process is identical for the
remaining range sensors. At time k, Sonar 8 detects an object some distance away.
Since the orientation of Sonar 8 is known, the range measurement can be converted
into a vector, denoted SP. The same point on the wall can be represented with
vectors VP, which is relative to the vehicle’s coordinate frame ({V}), and VP, which
is relative to the universal coordinate frame ({U}). Equation (4.7) relates SP and
VP, and by extension, Equation (4.8) relates all three vectors:
(cid:20) (cid:21) (cid:20) (cid:21)
VP SP
k,m = VT · k,m , (4.7)
1 S m 1
54 |
Colorado School of Mines | Figure 4.4: Wall Mapping Scenario: Sonar 8 Range Measurement
(cid:20) (cid:21) (cid:20) (cid:21) (cid:20) (cid:21)
UP VP SP
k,m = UT · k,m = UT · VT · k,m , (4.8)
1 V k 1 V k S m 1
where k is the time iteration index and m is the ultrasonic sensor number index. Note
thatthe‘1’isaplaceholderforpropermatrixdimensionagreement. Thehomogeneous
transformation matrix, UT, in Equation (4.8) consists of 2 components: a rotation
V
matrix and a translation vector, as shown in Equation (4.9) [5]:
cosθ −sinθ 0 X
k k k
(cid:20) (cid:21)
UT = U VR k UP VORG,k = sinθ k cosθ k 0 Y k , (4.9)
V k 0 0 0 1 0 0 1 Z k
0 0 0 1
where θ is the angle relative to the initial orientation, as depicted in Figure 4.3, and
X and Y are the horizontal and vertical displacements of the vehicle, respectively.
The third dimension, Z, can be ignored for this particular application, but is included
for completeness. Similarly, VT is broken down in Equation (4.10):
S
cosβ −sinβ 0 x
m m m
(cid:20) (cid:21)
VT = V SR m VP SORG,m = sinβ m cosβ m 0 y m , (4.10)
S m 0 0 0 1 0 0 1 z m
0 0 0 1
55 |
Colorado School of Mines | where β is the angle of the ultrasonic sensor relative to the vehicle’s coordinate frame
({V}) and (x,y) refers to the location of the sonar sensor on the AMR (see Fig-
ure 4.5). Again, the variable z is not necessary for a ground vehicle, but is included
for completeness. Each ultrasonic sensor has unique values for x, y, and β; a table of
these values can be found in Section A.11 of the Appendix.
Figure 4.5: VP and β for Ultrasonic Sensor 8
SORG
The overall idea of the map building equations is to build one large matrix (called
‘Map’) that contains the coordinates of every wall measurement in the environment,
using one sensor at a time; for each time iteration k, the sonar sensor’s readings in
Frame {U} give the (x,y) coordinates of one point on the map. Repeating the process
for the remaining sensors gives 13 more sets of (x,y) coordinates. Conglomerating
the 14 points from each time iteration (every 250 ms) during the entire experiment
gives the environment map of the wall. For an example of a resulting environment
map, see Figure 5.11 on Page 79. Additionally, a step–by–step outline of the map
building procedure is given at the end of the chapter. The next section attempts to
find unique features in the environment map (the large matrix called ‘Map’).
56 |
Colorado School of Mines | 4.3 Matching Algorithm
For this project, the purpose of a matching algorithm is to use exteroceptive
sensor data to identify unique features on a map produced from Section 4.2 and give a
correspondingvehicleposeestimateateachtimeiteration,k. Theposeestimatesfrom
the matching algorithmserveas themeasurements forthe measurement updatein the
Extended Kalman Filter algorithm (Section 4.4). Based on previous research efforts
[4] [1] [19], two different types of matching algorithms are tested: an Iterative Closest
Point (ICP) algorithm and an Iterative Closest Line Segment (ICLS) algorithm.
The fundamental difference between the ICP algorithm and the ICLS algorithm
is how the data association problem is addressed. The ICP algorithm, as its name
implies, simply associates each measured point with the closest point in the environ-
ment map. The ICLS algorithm addresses the space in–between points on the map,
associating each measured point with the closest point on the nearest line segment in
the environment map. Figure 4.6 illustrates the difference between the ICP and the
ICLS association method.
Figure 4.6: ICP Versus ICLS Algorithm Data Association
IfPointJ isasensor’smeasurementofthewall, theICPalgorithmwouldassociate
point E with the measured point, and the ICLS algorithm would associate point M
with the measured point. The coordinates of point M are found using Equation
57 |
Colorado School of Mines | (4.11), given by,
EC·EJ
P = P + , (4.11)
M E
||EC||
whereP isavectorcontainingthecoordinatesofPointM, P isavectorcontaining
M E
the coordinates of Point E, EC is a vector starting at Point E and ending at Point C,
EJ is a vector starting at point E and ending at point J, (EC·EJ) is the dot product
ofthevectorsECandEJ, and||EC||isthelengthofvectorEC. Ingeneral, theICLS
data association is closer to the original measurements; however, the ICLS algorithm
is computationally more expensive. Repeating the association process illustrated in
Figure 4.6 for each of the 10 wall–profiling sensors builds the data association matrix,
UP (a [10 x 2] matrix).
wall
After data association is complete, the remaining steps are the same for both ICP
and ICLS algorithms. Let q be the matching algorithm iteration number, and A be
a [10 x 3] matrix as shown in Equation (4.12):
UPT 1
q,2
UPT 1
q,4
UPT 1
q,5
UPT 1
q,6
UPT 1
A = q,8 . (4.12)
q
UPT 1
q,10
UPT 1
q,11
UPT 1
q,12
UPT 1
q,13
UPT 1
q,14
The matrix A represents the readings from the 10 wall–profiling sensors, expressed
58 |
Colorado School of Mines | in the universal coordinate frame, along with a column of ones, which serve as a
placeholder. Note that the numeric subscripts correspond to the sensor number (Fig-
ure 3.5) and that only the wall–profiling sensors are used in this analysis (the soft
bumper sensors, i.e. Sonars 1, 3, 7, and 9, are skipped in the numerical ordering
in matrix A). The vector UPT in Equation (4.12) is the transpose of the vector
q,m
computed in Equation (4.8). UP is a function of the vehicle pose, so it is updated
q,m
at every matching algorithm iteration, q, with the latest values for X , Y ,
match match
and θ . Note that the sensor range measurements, SP in Equation (4.8), are
match m
constant during the ICP or ICLS iterations.
Additionally, let UP represent the associated points on the wall, corresponding
wall
to either the ICP or the ICLS algorithm. Equation (4.13) calculates the rotation
matrix (R) and translation vector (B) required to align the ultrasonic sensor readings
with UP :
wall
(cid:20) (cid:21)
R
q = A−1 ·(UP ). (4.13)
B q wall,q
q
Note that the rotation matrix and translation vector, R and B, are similar to UR and
V
UP from Section 4.2, respectively, but they are not identical.
VORG
The matrix A is not square, so inverting it requires a pseudo inverse, such as
singular value decomposition. With the rotation matrix and translation vector (R
and B) in hand, there are numerous ways to back out the vehicle displacement and
translation errors, (X ,Y ,θ ). Realizing that R is actually the transpose of
drift drift drift
UR (from Equation (4.9) on Page 55) makes computing θ straightforward:
V drift
θ = sin−1(R (1, 2)), (4.14)
drift,q q
whereR(1, 2)istheelementofR locatedonthefirstrowandthesecondcolumn. The
rotation matrix is with respect to the universal frame, {U}, so the rotation affects
the translation vector. Equation (4.15) isolates X and Y from B:
drift drift
59 |
Colorado School of Mines | [X ,Y ] = B R−1 = B RT. (4.15)
drift,q drift,q q q q q
Then,Equation(4.16)updatesthematchingalgorithmvehiclepose,X ,Y
match match
and θ , by adding the computed vehicle drift:
match
X X X
match,q+1 match,q drift,q
Y match,q+1 = Y match,q + Y drift,q . (4.16)
θ θ θ
match,q+1 match,q drift,q
On the first iteration, the vehicle pose is initialized to the Kalman filter a–priori es-
timate (Equation (4.18) on Page 62). The index, q, is incremented, and the process
repeats,startingwithupdatingthedataassociation,untilconvergenceoruntilamaxi-
mumnumberofiterationsisreached. Onthelastiterationinq, X , Y , and
match,q match,q
θ become the measurement in the Kalman filter ([X ,Y ,θ ], in
matchq match,k match,k match,k
Equation (4.28) on Page 64) for the time iteration, k.
After convergence, the last step of the matching algorithm is to compute the
measurement uncertainty. There are many factors that go into the measurement un-
certainty, including the confidence in the map building process, the ultrasonic sensor
readings, and dead reckoning drift. One method for addressing the measurement er-
ror is to rely on dead reckoning pose for short distances. Adapted from Kolter et al.,
Equation (4.17) computes the standard deviation of the matching results [15]:
σ (X −X )
X match,k match,qfinal match,0
σ Y match,k = (Y match,qfinal −Y match,0) , (4.17)
σ (θ −θ )
θ match,qfinal match,0
match,k
where σ , σ , and σ are the standard deviations of X , Y , and
X Y θ match match
match match match
θ , respectively; qfinal is the value of q at the final ICP (or ICLS) iteration;
match
and X , Y , and θ form the initial vehicle pose, defined to be the
match,0 match,0 match,0
Kalman filter a–priori update. The standard deviation in Equation (4.17) reflects
the idea that the vehicle is unlikely to drift far during one time iteration, so the
60 |
Colorado School of Mines | farther away the matching algorithm pose deviates from the dead reckoning pose, the
higher the uncertainty. With the variance in vehicle pose defined in this manner, the
Kalman filter is unlikely to change the dead reckoning pose dramatically during one
time iteration; however, small corrections can add up significantly over the course of
testing.
Withthevehicledynamicsequations, atechniqueformapbuilding, andtheability
tolocalizerelativetothemap(i.e. thematchingalgorithm)inhand, alloftheindivid-
ual ‘ingredients’ are prepared for creating the main entr´ee; the next section develops
an Extended Kalman Filter, which is a method for blending the three ‘ingredients’
together.
4.4 Extended Kalman Filter (EKF)
An EKF provides a means of fusing exteroceptive information from a matching
algorithm with proprioceptive information from dead reckoning data. Combining the
two complementary localization techniques eliminates the problems of using either
algorithm individually. The EKF breaks the process into two steps: a prediction step
and a measurement update. During the prediction step, Ackermann dynamics pro-
vide a pose estimate based strictly on odometry and steering angle sensor readings.
The measurement update then uses ultrasonic range measurements and a matching
algorithm to correct for drift. The process repeats at each time iteration. The follow-
ing EKF equations are a modified version of the ICP–EKF algorithm from Madhavan
et al. [19].
4.4.1 EKF Prediction Step
Beginning with the prediction step, Equations (4.4) through (4.6) update the
vehicle pose. Additionally, two states are added to keep track of vehicle slip (d )
s
and skid (α ) defined as the error in vehicle odometers and the error in in vehicle
s
61 |
Colorado School of Mines | cos(θ ) 0 0 0
k−1
sin(θ ) 0 0 0
k−1
W k = L1 tan(α k +α s,k−1) L1 cosd 2d (k α+ k+d s α, sk ,− k1 −1) 0 0 , (4.26)
0 0 1 0
0 0 0 1
σ2 0 0 0
dd
k
0 σ2 0 0
Q k = 0 0α k σ2 0 . (4.27)
dd
s,k
0 0 0 σ2
α
s,k
4.4.2 EKF Measurement Update
The second step of the EKF is the measurement update. At each time iteration,
the matching algorithm uses range readings to measure the wall’s profile and return
a prediction of the vehicle’s location. The results from the matching algorithm serve
as the measurement that fine–tunes the vehicle pose from the prediction step. The
measurement update begins with computing the innovation, y˜ :
k
X
match,k
y˜ k = Y match,k −Hxˆ− k, (4.28)
θ
match,k
where H is the measurement sensitivity matrix, defined by Equation (4.29); and
X , Y , and θ are the output of the matching algorithm (Section 4.3 on
match match match
Page 57). The matrix H is defined as follows:
1 0 0 0 0
H = 0 1 0 0 0 . (4.29)
0 0 1 0 0
¯
The measurement update continues with computing the Kalman gain, K:
K¯ = P−HT(HP−HT +R )−1, (4.30)
k k k average,k
64 |
Colorado School of Mines | 1. Build an environment map:
(a) Define the origin of the universal coordinate frame as the vehicle’s starting
point.
(b) Start the vehicle
(c) At time k, record the vehicle’s sensor data: the odometers on each wheel,
the steering angle sensor, and the 14 range measurements from the sonar
sensors.
(d) Using the odometry and steering angle values, apply the Ackermann vehi-
cle dynamics equations (Equations (4.1) through (4.3)) to determine the
vehicle’s location ((X ,Y )) and orientation (θ ).
vehicle vehicle vehicle
(e) Using the sonar range data and the vehicle pose (location and orientation),
apply Equation (4.8) to find UP for each sensor. This equation trans-
k,m
lates the sonar range readings to a vector (a set of coordinates) referenced
in the universal coordinate frame. This step produces 10 measurements of
the wall.
(f) Add the 10 sets of coordinates from the previous step to the environment
map.
(g) Repeat Step 1(c) through 1(f) until map completion (the end of the ex-
periment).
2. Applythelocalizationalgorithmtoanysubsequentpathwithintheenvironment
map:
(a) Initialize the vehicle’s location relative to the environment map.
(b) Begin moving.
(c) At time k, record the vehicle’s sensor data.
(d) PerformtheEKFpredictionstep. Thisstepusestheodometryandsteering
angle data.
66 |
Colorado School of Mines | i. Update the states with Equation (4.18). This step produces an esti-
mate of the vehicle pose using dead reckoning.
ii. Calculate P−, the a–priori estimate of the covariance matrix, using
k
Equation (4.24).
(e) Perform the EKF measurement update. Using the sonar range measure-
ments in a matching algorithm, this step attempts to correct the vehicle’s
dead reckoning drift using the surrounding wall’s texture.
i. UsethesonarrangedataineithertheICPorICLSmatchingalgorithm
to correct the dead reckoning drift. The matching algorithm serves as
the Kalman filter ‘measurement.’ This step has several parts:
A. Record the vehicle pose at the start of the matching algorithm,
(X ,Y ,θ ).
0 0 0
B. Initializethematchingalgorithm’sposeestimate,(X ,Y ,θ ),
match match match
with the initial pose estimate.
C. For each sonar range measurement, find the corresponding closest
point on the environment map of the wall (or the closest point on
the closest line segment for the ICLS algorithm). This produces
UP , a [10x2] data association vector.
wall
D. For each sonar sensor, translate the sensor range measurements to
the universal coordinate frame with Equation (4.8), using the lat-
est matching algorithm vehicle pose, (X ,Y ,θ ). The
match match match
translated vectors form matrix A in Equation (4.12).
E. UsingthedataassociationvectorandmatrixA, findtheestimated
dead reckoning drift, (X ,Y ,θ ), using Equations (4.13),
drift drift drift
(4.14), and (4.15).
F. Use the estimated dead reckoning drift to update the matching
algorithm vehicle pose estimate (Equation (4.16))
67 |
Colorado School of Mines | CHAPTER 5
RESULTS
This chapter covers the results of the localization algorithm. Initially, data is
fabricatedtotestthealgorithm. Afterthealgorithmisproventoworkusingsimulated
data, itisappliedtorealdatacollectedattheEdgarMine. Thelocalizationalgorithm
and data post–processing is executed in MathWorks(cid:13)R MATLAB software.
The preliminary testing is performed exclusively in software, beginning with sim-
ulated sensor data. In this chapter, the first section describes the simulator that
produces the sensor data. The ICP–EKF localization algorithm is initially tested in
an ideal situation using the simulated data (Section 5.2). Afterwards, the localization
algorithm is tested in Section 5.3 using data collected from one section of the Army
Tunnel at the Edgar Mine. For this test, dead reckoning drift is added to the original
data to determine whether the localization algorithm can compensate. Due to sensor
calibration issues, a ‘multi–map’ method was created to get the localization algorithm
to work properly. The last test was the ‘double–pass’ test using sensor data from two
full passes of the Edgar Mine’s Army Tunnel (Section 5.4); the localization algorithm
uses the second pass to localize against the first. Because the sensor data was too
noisy for the localization algorithm, the experiment was repeated using simulated
sensor data.
5.1 Simulator
The simulator provides a means of testing the localization algorithms without the
complexity of actual data. The first step is to create an arbitrary vehicle path with
simlulated odometer and steering angle vectors. The wall is created by sending a
vector of sonar data through the wall mapping equations (Equation (4.8) on Page 55)
for Sonar 2 and Sonar 4, while running the simulated odometer and steering angle
69 |
Colorado School of Mines | values through the Ackerman dynamics equations (Equations (4.4) through (4.6) on
Page 53). The vector of sonar data begins with a series of line segments, and then
noise is added and the vector is passed through a low–pass moving average filter. The
output is a discrete set of horizontal and vertical displacement values, X and Y ,
sim sim
that form the vehicle path and a discrete set of points that form a map of the wall.
Figure 5.1 is the resulting map.
Map Simulator
10
Mine Wall
Vehicle Path
5
0
−5
−5 0 5 10
X (meters)
Figure 5.1: Map Building and Vehicle Path Simulation
After the wall is created, for each point on the vehicle path, [X ,Y ], a
sim,k sim,k
simulated set of sonar sensor readings is created. Figure 5.2 is an example of the
simulated sonar data collection process. The sonar projection is a line segment be-
tween a range reading of 0 meters and a range reading of 6 meters for each sonar
sensor (simulating a sensor beam). The distance from the sonar sensor to the nearest
intersection point becomes the sonar range reading.
Localization algorithm testing initially occurs using the simulated data. The orig-
inal simulated vehicle path is the ‘ground truth.’ A small amount of noise or drift
is added to the odometer data, the steering angle data, and the sonar sensor data.
The localization algorithm from the previous chapter receives the noisy sensor data
and attempts to predict the original ‘ground truth.’ For comparison, the same noisy
70
)sretem(
Y |
Colorado School of Mines | data runs through the Ackermann vehicle dynamics equations to produce the dead
reckoning predicted path. The closer a path is to the ground truth, the better the
performance.
Figure 5.2: Simulated Range Data Collection
5.2 Localization Algorithm Tests Using Simulated Data
A set of data for the 10 wall–profiling sensors, the odometer, the string poten-
tiometer, and the mine wall is fabricated in order to test the localization algorithm.
Furthermore, each sensor has artificial noise added: the odometer has a 30% drift,
the steering measurements from the string potentiometer have a random gaussian
noise with a 10◦ standard deviation, and the ultrasonic sensors’ range readings are
given a gaussian noise with a 15.2 cm standard deviation. Additionally, at each time
step, the orientation of each ultrasonic sensor, β , is given a gaussian noise with a 5◦
m
standard deviation, corresponding to the beam width of the ultrasonic sensors. These
noise values were initial estimates of data confidence and drift in a mine. Figure 5.3
is the resulting graph of the simulated data test using the ICP algorithm.
The ‘ground truth’ line in Figure 5.3 is the vehicle path before adding the noise.
The dead reckoning path calculates the vehicle path using Ackermann vehicle dy-
71 |
Colorado School of Mines | ICP Localization Algorithm
10
Mine Wall
8 Ground Truth
EKF Predicted Path
Dead Reckoning
6 Predicted Path
4
2
0
−2
−4
−6
−6 −4 −2 0 2 4 6 8 10
X (meters)
Figure 5.3: ICP Localization Algorithm Test Using Simulated Data
namics with the noisy odometer and string potentiometer data (measured steering
angle or heading), and the EKF predicted path uses the same noisy data, with the
added benefit of fusing the noisy dead reckoning data with the simulated noisy ul-
trasonic range data. Even for a worst-case scenario, the EKF predicted path is able
to eliminate the majority of the dead reckoning drift (‘noise’). The ICLS algorithm
performed in a similar manner—Figure 5.4 shows the results of the same test and
data using the ICLS algorithm.
The ‘ground truth’ line in Figure 5.3 is the vehicle path before adding the noise.
The dead reckoning path calculates the vehicle path using Ackermann vehicle dy-
namics with the noisy odometer and string potentiometer data (measured steering
angle or heading), and the EKF predicted path uses the same noisy data, with the
added benefit of fusing the noisy dead reckoning data with the simulated noisy ul-
trasonic range data. Even for a worst-case scenario, the EKF predicted path is able
to eliminate the majority of the dead reckoning drift (‘noise’). The ICLS algorithm
performed in a similar manner—Figure 5.4 shows the results of the same test and
data using the ICLS algorithm.
The ICLS algorithm yields identical results as the ICP algorithm, primarily be-
72
)sretem(
Y |
Colorado School of Mines | cause the map of the mine wall has small spacing between points. The only noticeable
difference between the ICLS algorithm and the ICP algorithm is that the tests take
68.3 and 4.2 seconds to run, respectively. In order to run in realtime, the ICLS code
would require optimization.
One method for increasing the processing speed of both matching algorithms is
to simply restrict the map available to the matching algorithm; map points over 8
meters away from the vehicle are an unlikely target as the maximum range reading
for these ultrasonic sensors is only 6 meters. Restricting the map access allows for
very large environment maps without a significant increase in processing time; data
storage becomes the primary limiting factor of map size.
ICLS Localization Algorithm
10
Mine Wall
8 Ground Truth
EKF Predicted Path
Dead Reckoning
6 Predicted Path
4
2
0
−2
−4
−6
−6 −4 −2 0 2 4 6 8 10
X (meters)
Figure 5.4: ICLS Localization Algorithm Test Using Simulated Data
Regardlessofthealgorithmused, thelocalizationalgorithmisveryrobusttonoisy
and drifting sensor data. In fact, both the EKF–ICP and the EKF–ICLS localization
algorithms managed to reject the dead reckoning drift when the noise on the steering
and odometry sensors were increased by 50%.
73
)sretem(
Y |
Colorado School of Mines | 5.3 Localization Algorithm Test Using Collected Data
With the promising results using simulated data, the next step is to test the
localization algorithms on actual data. The first step is to build a map—Figure 5.5
uses Equation (4.8) to build the environment map of a portion of the Edgar Mine
Army Tunnel for each sonar sensor from data collected on March 11, 2010. Figure 5.5
highlights several issues with the ultrasonic sensors in a mine: the range readings are
very noisy, some sensors (Sonars 5 and 8) appear to be incorrectly calibrated, and
the diagonal sensors (Sonar 5, 6, 11, and 12) have large spikes in their range readings
that appear to be more than noise.
Map Building Sonar 2
S onar 4
Sonar 5
12
Sonar 6
10 Sonar 8
Sonar 10
8 Sonar 11
Sonar 12
6
Sonar 13
4 Sonar 14
2
0
−2
−4
−6
−5 0 5 10 15 20
X (meters)
Figure 5.5: Map Building Using Ultrasonic Sensor Range Data
First, regarding the noise issue, when the AMR was stationary, the standard de-
viation of some sensors were over one meter, though the average standard deviation
came out to a more–reasonable 0.25 meters. Regarding the calibration, these partic-
ular sensors seemed to drift from day to day, so they required frequent calibration.
Regarding the large spikes in the diagonal sensors, the ultrasonic sensors exhibited
similar behavior, though more severe, when ranging a flat wall at a steep angle of
incidence (see Section 3.2.4 on 35).
TheenvironmentmapinFigure5.5isverycluttered. Thus, foraestheticpurposes,
74
)sretem(
Y |
Colorado School of Mines | an averaged mine wall was created. Figure 5.6 compares the environment map (the
raw data) to the averaged wall; the averaged wall is the result of vertically averaging
the mine wall measurements from the environment map. The averaged mine wall is
not the environment map, but rather, a cleaned–up representation of the environment
map,soitisusedintheremainingfiguresinthissection. Thenoiseintheenvironment
mapfarexceedsthewall’stexture; thetextureseenintheaveragedwallisanartificial
by–product of the averaging technique.
Averaged Map Building
Environment Map
Averaged Wall
10
5
0
−5
−5 0 5 10 15 20
X (meters)
Figure 5.6: The Averaged Wall
The drifting calibration for individual sensors were a persistent problem in all
data collected. To alleviate the issue, a wall is constructed for each wall–profiling
sensor (instead of building one composite wall using all sensors), making a total of 10
maps; thus, each line in Figure 5.5 represents one map. The matching algorithm only
associates a sensor’s measurements with points from its own map. Other techniques
include averaging the points together to build one composite, but this technique tends
to dull the unique mine features into one continuous line, making localization nearly
impossible. Figure 5.7 are the results of the localization algorithm tests using the
75
)sretem(
Y |
Colorado School of Mines | multi–map method.
ICP Localization Algorithm
12
Mine Wall
10 Ground Truth
EKF Predicted Path
8 Dead Reckoning
Predicted Path
6
4
2
0
−2
−4
−6
−10 −5 0 5 10 15 20 25
X (meters)
Figure 5.7: ICP–EKF Algorithm Test Using Collected Data
To reiterate, the mine wall in Figure 5.7 is the result of vertically averaging the
mine wall measurements from the environment map in Figure 5.5, and the matching
algorithm uses the environment map (now comprised of 10 maps, one for each wall–
profiling sonar sensor). For the purpose of testing, the ground truth is the vehicle
path predicted from the original dead reckoning data. A 20% drift was added to the
original odometry data, and a Gaussian noise with a 2◦ standard deviation was added
to the steering angle. The noisy sensor data was used to form the dead reckoning
predicted path in Figure 5.7. Notice that the EKF predicted path performed little
correction in the odometry data (the paths are about the same length); with no
unique features, the localization algorithm is unable to determine an exact position.
However, the localization algorithm was able to correct vehicle orientation, so the
EKF predicted path was, for the most part, aligned with the ground truth path.
Furthermore, as Bakumbu predicted [1], the localization algorithm easily degenerates
with the relatively flat walls of this particular section of the Army Tunnel, as shown
in Figure 5.8.
Degeneration occurs when the matching algorithm causes the pose estimate to
76
)sretem(
Y |
Colorado School of Mines | ICP Localization Algorithm
12
Mine Wall
10 Ground Truth
EKF Predicted Path
8 Dead Reckoning
Predicted Path
6
4
2
0
−2
−4
−6
−10 −5 0 5 10 15 20 25
X (meters)
Figure 5.8: ICP–EKF Algorithm Degeneration
get off–track. Once off–track, the range readings have little correlation with the
environment, so the matching algorithm is unable to converge, or converges on the
wrong wall segment, causing the pose estimate to become even more off-track. This
cascading effect results an unstable localization algorithm, which causes the vehicle
pose oscillation depicted in the EKF predicted path of Figure 5.8. As Bakumbu
suggests, matching algorithm results are ignored in these portions of the mine [1].
Repeating the same test for the ICLS algorithm gave the results in Figure 5.9.
Similar to the ICP localization algorithm, the ICLS localization algorithm was able
to correct the vehicle orientation, but not the odometer readings.
This section provides evidence that the ICP–EKF algorithm can work given per-
fectlyconsistentrangemeasurements; iftherangemeasurementsdonotchangeduring
subsequent passes, the localization algorithm demonstrates a solid ability to elimi-
nate some of the dead reckoning drift. The following section tests the localization
algorithm in a more realistic scenario; the experiment involves two separate passes
of the Edgar mine to determine whether the localization algorithm can use the data
collected during the second pass to localize against the first.
77
)sretem(
Y |
Colorado School of Mines | ICLS Localization Algorithm
14
Mine Wall
12 Ground Truth
EKF Predicted Path
10 Dead Reckoning
Predicted Path
8
6
4
2
0
−2
−4
−6
−5 0 5 10 15 20
X (meters)
Figure 5.9: ICLS–EKF Algorithm Test Using Collected Data
5.4 Double Pass Testing
The final localization algorithm test uses data from two separate passes of the
Edgar Mine Army Tunnel. Using dead reckoning, data from the first pass builds
the environment map. Similar to the technique used by Ma¨kela¨ [20], the localization
algorithmreceivessensordatafromthesecondpasstolocalizerelativetothefirstdata
set’s environment map. The purpose of this test is to determine if ultrasonic range
data can localize relative to a map created by ultrasonic range data. This would
prove useful if the AMR would need to return to a previously traversed location.
Figure 5.10 is the map built from the first pass, before and after noise correction.
Environment Map, Pass #1 Sonar 2 Environment Map, Pass #1 Sonar 2
Raw Data Sonar 4 Noise Corrected Sonar 4
15
S So on na ar
r
5
6 15
S So on na ar
r
5
6
10 S So on na ar r 8 10 10 S So on na ar r 8 10
Sonar 11 5 Sonar 11
5 Sonar 12 Sonar 12
0 S So on na ar r 1 13 4 0 S So on na ar r 1 13 4
−5 −5
−10 −10
−15 −15
−20 −20
−25 −25
−30 −30
0 10 20 30 40 50 0 10 20 30 40 50
X (meters) X (meters)
(a) Before Noise Correction (b) After Noise Correction
Figure5.10: EnvironmentMapoftheEdgarMineArmyTunnelUsingCollectedData
78
)sretem(
Y
)sretem(
Y
)sretem(
Y |
Colorado School of Mines | The environment map is built using dead reckoning data; thus, dead reckoning
drift is inherent. Ma¨kela¨ et al. corrected for the drift using a linear scaling factor
on the odometry data relative to a known passage distance [20]. However, since an
emergency rescue vehicle is not intended to be a surveying instrument, an absolutely
accurate map is unnecessary, so drift correction, or some other means of ground truth
(such as a total station), for the environment map was not implemented.
A second complete pass of the Army Tunnel forms the test data set. The starting
position is identical for both passes. Figure 5.11 is an overlay of the two data sets,
created from dead reckoning and ultrasonic range data. The test data set has a
noticeabledriftcomparedtotheenvironmentmap,whichexpectedfordeadreckoning.
Even with the noise correction for both data sets, the mine wall thickness (a measure
of confidence) is on the order of one meter, which is relatively large.
Edgar Mine Environment Map Overlay
5
0
−5
−10
−15
Environment Map (Pass #1)
Test Data (Pass # 2)
−20
−10 0 10 20 30 40 50 60
X (meters)
Figure 5.11: Edgar Mine Environment Map Overlay
Any attempt to localize using the map and test data set depicted in Figure 5.11
proved futile. Averaging techniques, the multi–map technique (each sonar sensor can
onlybeassociatedwithitsowndata), andthesinglemaptechnique(eachsonarsensor
can be associated with any data), all resulted in either no dead reckoning correction
or an unstable localization algorithm. The localization algorithm relies on consistent
79
)sretem(
Y |
Colorado School of Mines | range measurements, which the ultrasonic sensors are unable to provide; the uncer-
tainty and noise of the ultrasonic sensors washes out the unique features of the mine
wall, essential erasing its ‘fingerprints.’ The multi–map method employed in the pre-
vious section appeared to correct the vehicle orientation because the uncertainty in
the environment map was essentially ignore. Once factoring in the uncertainty of
the map–building process, the uncertainty in the matching algorithm far exceeds the
uncertainty in dead reckoning for any reasonable amount of distance traveled. Thus,
the matching algorithm is unable to correct the dead reckoning drift.
The inconsistent range readings caused the difficulties experienced when using
actual data. To test this theory, the simulator is revisited. Figure 5.12 is an overlay
of two simulated data sets. Similar to Figure 5.11, the test data set has a slight dead
reckoning drift and each sonar sensor has range and beam width noise to duplicate
sensor response in the mine. Unlike the actual range data, the simulated sonar data
has repeatable, Gaussian noise about the wall.
Simulator Environment Map Overlay
8
6
4
2
0
−2
−4
−6
−8 Environment Map (Pass #1)
Test Data (Pass # 2)
−10
−5 0 5 10 15 20 25
X (meters)
Figure 5.12: Simulator Environment Map Overlay
The two simulated sets of data get the same treatment as the actual data sets.
The first pass is used to create the environment map and the localization algorithm
is tested using the second set of data. Figure 5.13 are the results. With the simulated
80
)sretem(
Y |
Colorado School of Mines | data, the ICP localization algorithm has little trouble with localization on the second
pass; the AMR could return to any point on its map with a fair amount of accuracy.
Thus, ultrasonic range finders could theoretically pass the double–pass test.
ICP Localization Algorithm
Environment Map
8 Ground Truth
EKF Predicted Path
6 Dead Reckoning
Predicted Path
4
2
0
−2
−4
−6
−2 0 2 4 6 8 10 12 14 16
X (meters)
Figure 5.13: ICP Localization Algorithm Double Pass Test Using Simulated Data
The failure of the localization algorithm on the collected data and its success on
the simulated data is due to one fundamental difference: the simulated data set is
intentionally formulated with wall features that exceed the sonar range measurement
noise. The sonar range measurement noise has been a recurring issue throughout
the project; the confidence in the range measurements increases dramatically as the
sensor approaches the critical angle (Section 3.2.4 on Page 35) and the sensors were
unable to pick up wall texture in the collected data (Figure 5.11). Furthermore, the
four sensors mounted at an angle greater than the critical angle exacerbate the noise
issue, as their standard deviation far exceeds the assumed standard deviation (as seen
in Sonars 5 and 6 in Figure 5.10(a)). In other words, the sensor noise washes out
the texture in the Edgar mine wall. Thus, the ultrasonic sensors are incapable of the
resolution required for localization at the Edgar Mine.
81
)sretem(
Y |
Colorado School of Mines | CHAPTER 6
CONCLUSION
Theprimarygoalofthisthesiswastoproducealocalizationalgorithmusingultra-
sonic sensors, along with the documentation required to expand upon MineSENTRY
experiments. To demonstrate the capabilities of localization algorithm, the first step
was to build a map; though the ultrasonic sensor readings were noisy and inconsis-
tent, a map of the Edgar Mine Army Tunnel was successfully built (Figure 5.10).
Using the map–building algorithm, a simulator was developed to test the theoretical
capabilities of a matching algorithm (Section 5.1). Lastly, an ICP–EKF localization
algorithm and an ICLS–EKF localization algorithm were developed (Sections 4.3 and
4.4) and successfully tested on simulated data (Section 5.4). Though designed for
use with scanning laser sensors, computer simulations suggest that, using ultrasonic
sensors, the ICP–EKF (or ICLS–EKF) localization algorithm consistently improves
upon the dead reckoning predicted path, allowing accurate localization in the thick
smoke common in mine disaster scenarios.
The matching algorithms were also tested on range data collected in the Edgar
Mine (Section 5.4); however, the sensors’ noise exceeded the variation in the mine
wall, preventing the localization algorithm from working properly. This does not
imply that ultrasonic sensors are insufficient for localization, but rather suggests that
further testing is needed using alternative sensors.
6.1 Recommendations for Future Work
An immediate next step is toexamine alternatives tothe particularultrasonic sen-
sors used, and to test the alternatives in a mine environment. Though the ultrasonic
sensors’ wide detection pattern is advantageous in some situations (e.g. robustness
to vehicle tilting, object detection), a narrow beam width is required for wall profil-
83 |
Colorado School of Mines | ing because of the lower uncertainty. Alternatively, one research project developed
a sonar system that directly addressed the wide beam width problem inherent in
most ultrasonic range finders [16]. Inspired by successful usage of ultrasonic waves
for navigation in nature (e.g. bats and dolphins), Kreczmar developed a sonar system
with one transmitter and two receivers. Similar to the two ears on a bat, the extra
receiver aids in triangulating the source of a reflected acoustic wave, as shown in
Figure 6.1. The graph on the left (Figure 6.1a) illustrates the range measurements
from a single sonar as the sensor is rotated through the environment. The sensor
is unable to determine the direction of the reflecting acoustic wave, resulting in the
sweeping arcs. The graph on the right (Figure 6.1b) illustrates the performance of the
“Tri–auler” sonar system; the added receiver allows for better triangulation of objects
within the environment, eliminating the sweeping arcs. Implemented on the AMR,
the “Tri–auler” sonar system may reduce the sensor uncertainty and noise due to the
inherent wide beam width of ‘traditional’ ultrasonic range finders, thereby improving
the performance of the localization algorithm.
Figure 6.1: “Tri-auler” Sonar System [16]
Additionally, the sensors used suffered from a low critical angle; the diagonally–
oriented sensors had difficulty detecting the mine wall. Ultimately, the alternative
sensors need to be tested for their critical angle and re–oriented accordingly for future
iterations of the MineSENTRY project.
84 |
Colorado School of Mines | Another recommendation for future work involves converting the localization al-
gorithm into a full SLAM algorithm. The current navigation routine follows along the
main adit using a wall–following algorithm with logic to avoid the side drifts. This
works well if the main adit is relatively straight with no branches, or if a mine map is
readily available to anticipate the side drifts. However, the environment map is not
always available, so to be truly autonomous, a vehicle requires a SLAM algorithm.
A successful localization algorithm is a step towards a full SLAM algorithm; the ICP
(or ICLS) matching algorithm used in localization plays intricate role in the SLAM
algorithm of many research efforts. The main difference between a SLAM algorithm
and the navigation routine currently implemented is a SLAM algorithm’s ability to
correct dead reckoning drift on the first pass (the mapping stage). This is necessary
if the mine is ‘circular’ (i.e. paths overlap), as the robotic vehicle would need to
accurately recognize a previously–traversed path during its forward progress. The
advantages of a SLAM algorithm are clear; a full SLAM algorithm allows a robotic
vehicle to navigate mine passages without a–priori knowledge of the environment and
with no human intervention, while simultaneously correcting dead–reckoning drift.
The navigational abilities would be superior to the simplistic wall–following tech-
niques currently implemented.
6.2 Implications of Research
Even without a full SLAM algorithm, the AMR could still prove to be an in-
valuable tool for emergency rescue workers. Equipped with a simple wall-following
control algorithm and the wireless communication control algorithms developed in
work done by Meehan [22], the AMR can autonomously establish and maintain a
wireless mesh network for vital rescue operation communications. Further equipped
with the navigation routines developed in work done by Hulbert [13] and the ICP–
EKF algorithm developed in this thesis, the AMR has the added capability of re-
85 |
Colorado School of Mines | A.4 Robotic Controller to Servo Controller Communication
Description: The communication between the Robotic Controller and the motor
control board is determined by the motor control board (Roboteq AX3500) command
set and configuration. The robotic controller initiates all communications (except the
AX3500 Watchdog) and the AX3500 responds in one of two ways depending on the
type of command sent by the robotic controller. The AX3500 command set consists
of ASCII strings which correspond to either a request for an action or a request for
data. If a command is issued to request an action, the AX3500 will reply with a
plus (+) to acknowledge the request. If a command is issued that requests data, the
AX3500 will reply with the data to acknowledge the request. If any error occurs with
the communication, the AX3500 will reply with a minus (-) to indicate the command
should be repeated.
AllnumberseithersenttoorreceivedfromtheAX3500arerepresentedinhexadec-
imal format with ASCII characters, and all commands are followed by the carriage
return character,“\r.”
The data is transferred via standard RS-232 signal levels at a 9600 Baud, syn-
chronous data stream with 7 data bits, 1 start bit, 1 stop bit, and even parity. Refer
to Table A.2 for the packet structure specifics.
Table A.2: Robotic Controller to Servo Controller Communication Packets
Byte # Name Value Type Unit Description
0:5 BrakePosition “!Bnn\r”or“!bnn\r” String N/A Brake position relative to zero,
where ‘zero is about half of the full
desired travel. “nn” is a number in
hexadecimal between 00 and 7F (0-
127). In this case, !B7F\r causes a
full release of the brake and !b7F\r
causesthebraketobefullyengaged.
6:10 SteeringPosition “!Ann\r”or“!ann\r” String degrees Steeringpositionintermsofangular
positionofthewheels. nnisanum-
ber in hexadecimal between 00 and
7F(0-127). Inthiscase,!A7F!bnn\r
causes the steering wheel to turn
right at full speed and !a7F!bnn\r
causesthesteeringwheeltoturnleft
atfullspeed.
96 |
Colorado School of Mines | A.5 Robotic Controller to Autonomous Controller Communication
Description: The communication between the Autonomous Controller (AC) and
the robotic controller (VC for vehicle controller)) is determined by the robotic con-
troller mode set and status. The AC initiates all communications (except emergency
brakes) and the VC responds in one of two ways depending on the type of packet sent
by the AC. The AC communication set are ASCII strings which correspond to either
a request for an action or a request for data. If a command is issued to request an
action, the VC will reply with a sensory response and a corresponding packet-type to
acknowledge the request. If a query is issued that requests data, the VC will reply
with a sensory response to acknowledge the request. If any error occurs with the
communication, the VC will reply with...
All numbers either sent to or received from the AC are represented in hexadecimal
format with ASCII characters and all commands are followed by the carriage return
character “\r.”
The data is transfered via standard RS-232 signal levels at a 115200 Baud, syn-
chronous data stream with 8 data bits, 1 start bit, 1 stop bit, and no parity.
Table A.3: Robotic Controller to Autonomous Controller Header (common for all
packets)
Byte # Name Value Format Unit Description
0:7 StartSequence “{start:}” string N/A Thestartsequenceisusedincaseone
sideisnotinsyncwiththeother.
8 PacketType 0 uint8 N/A This field indicates that the packet is
asystempacket.
9 Length X uint8 N/A Length of the packet in bytes, exclud-
ing the start sequence, packet type,
andlengthfields.
97 |
Colorado School of Mines | ABSTRACT
The economic extraction of thin-seam coal deposits are often problematic due to several
significant limitations associated with conventional mining methods, operating practices, and
equipment. Mines with low-seam heights are endemic of operations that possess low labor
productivities, high operating costs, and relatively small production capacities. Furthermore, the
ability to implement new equipment and automation in order to efficiently exploit these thin-
seams is hampered by the limited cash-flow positions of most of these operations and the
inability to amortize their high costs over a sufficiently large resource base. Consequently, these
mines are usually small, labor intensive, and rely extensively on used and rebuilt equipment
modified to operate in these challenging work environments. It appears that the most prudent
way to extract these resources in a more economical way is through the development of
technology to remotely extract these resources from the surface. Displacing workers from the
underground work environment will eliminate the inherent hazards of mining these deposits and
reduce the direct costs associated with ventilation, support, and equipment.
Despite several potential benefits, there are a number of technical challenges that must be
overcome to advance the concept of in-situ borehole extraction of non-soluble resources to a
commercially viable stage. Paramount among these include the continued technical advancement
of drilling and excavation systems, the mechanisms used to crush, bail, and transport cuttings
from the borehole, the required instrumentation to effectively control and monitor the mining
process, and a technical understanding between cavity formation and stability for a given set of
operating characteristics and geomechanical rock properties. While each of these areas are
important to the overall success of the technology, understanding the structural dynamics of
iii |
Colorado School of Mines | these cavities is a key element in designing a mining system capable of sufficient resource
recovery to economically justify the capital investment. The unintended collapse of these cavities
could potentially result in the detrimental loss of mineral reserves, as well as surface subsidence,
the incidence of significant dilution, and the loss of equipment. In addition, adverse alterations in
cavity geometry caused by failure in the surrounding host rock will significantly hamper the
ability to recover and bail fragmented mineral from the borehole. Given these factors, borehole
placement, the excavation strategy, extraction ratios, production rate, and overall project
economics are all highly influenced by the stability of these subsurface excavations. Cavity
stability, in turn, is the product of a complex set of multi-dimensional variables that include in-
situ stresses, rock properties, cavity geometry, time, and the rate and manner of excavation. A
number of additional confounding issues associated with the proposed excavation methods (e.g.,
fluid pressurization of the cavity) may also adversely influence the stability of these cavities,
where their potential effects need to be quantified.
Several models were analyzed using Flac2D to perform a parametric study of
factors that could potentially impact the cavity stability during borehole mining process, whereas
information derived from the literature was useful in identifying several parameters that could
possibly affect cavity design and the mining process. The results and observations of these
studies (numerical modeling and literature search) led to a proposed protocol design for the
creation of stable cavities during borehole mining.
iv |
Colorado School of Mines | AKNOWLEDGEMENTS
First and foremost, praises and thanks be to God, the Almighty, for His showers of
blessings throughout my research work and the ability to complete my dissertation successfully.
I am extremely grateful to my parents for their love, prayers, caring, motivation, and
sacrifices for educating and preparing me for my future. I am very much thankful to my wife,
Zahra, and my daughter, Layla, for their love, understanding, prayers, and continued support to
complete this research work.
I would like to express my deep and sincere gratitude to my advisor, Dr. Hugh Miller for
giving me the opportunity to do research and provided invaluable guidance throughout this
research. His dynamism, vision, sincerity, and motivation have deeply inspired me. It was a
great privilege and honor to work and study under his guidance.
Finally, my thanks go to all the people who have supported me to complete research work
directly or indirectly. I would never have been able to finish my dissertation without the
assistance of my committee members and support from my family and wife.
xv |
Colorado School of Mines | CHAPTER 1
INTRODUCTION
1.1) Introduction
The economic extraction of thin-seam coal deposits with thicknesses of less than 1 meter
are often problematic due to several significant limitations associated with conventional mining
methods, operating practices, and equipment. Mines with low-seam heights are endemic of
operations that possess low labor productivities, high operating costs, and relatively small
production capacities. Furthermore, the ability to implement new equipment and automation in
order to efficiently exploit these thin-seams is hampered by the limited cash-flow positions of
most of these operations and the inability to amortize their high costs over a sufficiently large
resource base. Consequently, these mines are usually small, labor intensive, and rely extensively
on used and rebuilt equipment modified to operate in these challenging work environments.
Britton [2] has investigated the productivity of thin-seam mining, where he attributes the
inherent low productivity of these operations to the small tonnage produced per unit length of
linear advance and the challenges associated with mining in a confined work environment,
particularly with regards to material handling, unit operations at the face, and logistics. Another
pragmatic issue identified by Britton is ventilation, where obtaining the required airflow at
specific work areas can be difficult due to the obstruction of a large percentage of openings by
equipment. [2]
1 |
Colorado School of Mines | Fotta et al [3] did research to identify the types of injuries common to operating mines
that exploit thin-seam resources. Due to their research objectives, longwall operations and large
mines that employing more than 50 workers were specifically excluded from this study. Fotta
found a direct correlation between working height and worker safety principally attributed to
restricted employee mobility, reduced vision and poor posture, and limitations of using
protective canopies and other common engineering safety devices/controls as a consequence of
space constraints. [3]
Building upon Fotta’s research, Peters et al [4] studied significant potential hazards that
are substantially dependent upon coal seam height. The study focused on accidents that occurred
between 1990-1996, where 117 workers were fatally injured at small (non-longwall)
underground bituminous coal mines. The three types of incidents that were responsible for a
preponderance of these fatalities were associated with: 1) roof falls, 2) powered haulage, and 3)
machinery. Based on the same research, the six most prominent types of incidents responsible
for non-fatal days lost (NFDL) accidents were: 1) handling materials, 2) machinery, 3) powered
haulage, 4) slips and/or falls of person, 5) roof falls, and 6) non-power hand tools. [4]
Figures 1.1 and 1.2 illustrate the fatal and non-fatal day lost incident rates for
underground coal mines between 1998-2008. Using the standard industry convention, the
incident rate is calculated by the equation:
IR = (Number of Injuries x 200,000) / Number of Employee Man-Hours
Based on the number of employees, U.S. underground coal mines have been divided into two
distinct categories: small (employing less than 50 employees) and large (employing greater than
2 |
Colorado School of Mines | 50 employees). Using data obtained from the Mine Safety & Health Administration (MSHA)
Injury Experience Reports for Coal, Figure 1.1 shows, with the exception of 2007, the fatal
incident rates for small underground mines are higher as compared to larger operations. As
Figure 1.2 shows, except for years 2004 and 2008, the NFDL incident rates are higher for small
underground coal mines as well.
During 2008, an estimated 5.2 million tons of coal were produced in the U.S. from thin
seams deposits with operating heights less than 1.06 m (42 inches) [12]. This represents
approximately 14.5% of the total U.S. underground coal production. Of the 657 underground
coal mines operating in 2008, nearly 70% (n=454) of them have less than 50 employees. This is
consistent with previous research that indicates 94% of mines operating in seams of 1.06 m or
less employed fewer than 50 people [3].
<50 employees
0.12
>50 employees
e
t
a 0.1
r
t
n 0.08
e
d
i c 0.06
n
i
l 0.04
a
t
a
F 0.02
0
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Year
Figure 1.1. Fatal incident rate for small and medium/large underground coal mines for 1998-
2008. (After data – MSHA Injury Experience Reports for U.S. Coal Mining 1998-2008).
3 |
Colorado School of Mines | <50 employees
12
e >50 employees
t 10
a
r
t
n 8
e
d
i 6
c
n
i
4
L
D
F 2
N
0
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Year
Figure 1.2. Non-fatal day lost incident rate for small and medium/large underground coal
Mines for 1998-2008. (After data – MSHA Injury Experience Reports for U.S. Coal Mining
1998-2008).
Papas et al [5] conducted research that focused on injuries attributed to roof and rib falls
that occurred during the period 1995-1998 in all the U.S. underground coal mines. As expected,
this study concluded that mines operating in thin seams (< 1.09 m) tended to be smaller scale
operations that solely used room-and-pillar extraction methods. The research showed that small
mines in thin seams had a ground fall fatality rate that is 44% greater than the national average
(Figure 1.3), while small mines operating in thick seams had a ground fall fatality rate that is
53% lower than the national average. The consequence of this data has led to the conclusion that
small mines in themselves are not a contributing factor in the incidence of rock fall fatalities,
whereas seam height appears to be a dominant and influential factor. The study also indicates
that the potential root-causes of these fatalities largely stem from the lack of engineering controls
specifically suited for these small, low seam environments, such as protective cabs and canopies
on operating equipment. In these operating environments, workers are especially at risk should a
4 |
Colorado School of Mines | massive roof or rib failure occur. These results are consistent with the conclusions reached by
both Fotta and Peters. [5]
Despite these challenges, a number of companies are focusing on ways to exploit these
narrow, tabular deposits. In specific geographical locations whose economies are heavily
dependent on coal mining, the depletion of thick seams have placed greater pressure to develop
innovative technologies to economically recover coal resources from much thinner deposits.
Donovan et al [6] predicts this will include novel mining methods to exploit resources located in
seams less than 90 cm (36 in) in height. According to Holman et al [7], it is estimated that a
resource of over 5 billion tons of undeveloped, high quality coal in seams of between 14 to 28
inches in thickness exists in Virginia alone.
0.07
r
h
0
0 0.06
0
0
0
2 0.05
/s
lla
0.04
<50 workers
F
b
50-149 workers
iR 0.03
>149 workers
&
f 0.02
o
o
R
la
0.01
t
a
F 0
<43 43-60 >60 USA
Seam Height (inches)
Figure 1.3. Roof and rib fall fatality rates by mine size and seam height for room-and-pillar
mines, 1995-1998. [5]
An examination of world coal resources also shows that a substantial percentage is
contained in thin-seam deposits. Shiua [8] estimates that approximately 96% of China’s total
5 |
Colorado School of Mines | coal production is from underground mines, where more than half of these reserves are reside in
thin seams less than 0.7m. As such, these thin-seams are of strategic importance and represent a
significant source for supplying the country’s future energy needs. However, advancements in
safety and extraction methods are requisites for exploiting these resources. Coupled with
competing demands for surface land-use, it appears that the most prudent way to achieve these
objectives is through the development of technology to remotely extract these resources from the
surface. Displacing workers from the underground work environment will eliminate the inherent
hazards of mining these deposits and reduce the direct costs associated with ventilation, support,
and equipment.
A number of researchers have been working on the development of in-situ applications
for the extraction of coal and other economic minerals through boreholes. For example, Dr.
George Savanic of the U.S. Bureau of Mines conducted experiments using in-situ borehole
mining equipment on several different commodities including coal, gold, and industrial minerals,
beginning in the mid-1980s. His test results for the borehole mining of coal showed that it was
technical feasibility but not economically viable at the time. Through his work, several
advantages associated with this method were quantified, including improved worker safety and
the mitigation of adverse environmental impacts. [9]
Bairu et al [10] also studied a number of borehole hydraulic coal mining systems and
concluded that the method possessed the following advantages: lower cost of capital
construction, improved work intensity, and safer coal extraction. These conclusions were similar
6 |
Colorado School of Mines | to those reached by Wang and Miller [19] and Wiley [11]. Wang and Miller also identified
potential reductions in the cost of social impacts over conventional mining systems.
Wiley et al [11] addressed several advantages for borehole mining including: safety,
ability to work in remote area, minor environmental impacts, mobility, selectivity, low capital
cost, system simplicity, and ability to work in various operating conditions. According to his
work, borehole mining can compete with conventional methods in areas that possess tabular
structures, low ore concentration, complicated hydrogeological conditions, dangerous conditions,
and inaccessible locations.
1.2) Definition of Borehole Mining Method (BHM)
In the context of this research, Borehole Mining (BHM) is a remote operated method of
extracting (mining) mineral resources through boreholes by means of high pressure fluid jets.
This process can be conducted through a variety of operating configurations and drilling
platforms, including conventional surface locations, sites within existing mines (both open pit
and underground), and floating vessels/rigs.
A borehole is drilled from the surface to a desired depth, where the actual mining process
will take place. After the hole has been drilled, a column of casing is then inserted into the hole.
The purpose for using casing depends on the method in which the cuttings are bailed (removed)
from the hole. In conventional applications, the casing provides hole stability, minimizes
material loss, and reduces the potential for dilution caused by wall erosion. Since a cavity is
formed as part of the normal excavation process, a casing shoe is strategically positioned within
7 |
Colorado School of Mines | 1.3) Problem Statement
There are a number of technical challenges that must be overcome to advance the concept
of in-situ borehole extraction of coal and other non-soluble tabular resources to a commercially
viable stage. Paramount among these include the physical development of the drilling and
excavation equipment, the mechanisms used to crush, bail, and transport cuttings from the
borehole, the required instrumentation to effectively control and monitor the excavation process,
and a technical understanding between cavity formation and stability for a given set of operating
characteristics and geomechanical rock properties. While each of these areas is important to the
overall success of the technology, understanding the structural dynamics of these cavities is a
key element in designing a mining system capable of sufficient resource recovery to
economically justify the capital investment. The unintended collapse of these cavities could
potentially result in the detrimental loss of mineral reserves, as well as surface subsidence, the
incidence of significant dilution, and the loss of equipment. In addition, adverse alterations in
cavity geometry caused by failure in the surrounding strata will significantly hamper the ability
to recover and bail fragmented coal from the borehole.
Given these factors, borehole placement, the excavation strategy, extraction ratios,
production rate, and overall project economics are all highly influenced by the stability of these
subsurface excavations. Cavity stability, in turn, is believed to be the product of a complex set of
multi-dimensional variables that include in-situ stresses, rock properties, cavity geometry, time,
and the rate and manner of excavation. In applications where coal is being extracted through
these in-situ processes, a number of additional issues associated with the proposed excavation
methods may adversely impact the stability of these cavities, where their effects need to be
quantified. These factors include: fluid pressurization of the cavity, degassing of the coal
9 |
Colorado School of Mines | formation, and the introduction of gas and superheated fluid into the sub-surface cavity as part of
the excavation process.
This design process is further complicated by the inherent desire to improve resource
recovery and operating efficiencies by strategically orienting individual cavities relative to each
other and at times, excavating material through the interaction of multiple boreholes. This would
effectively allow for the remote extraction of panels and large block sections of resources
through a concerted operating strategy between different boreholes.
1.4) Objective
In an effort to address one of the prevailing technical challenges adversely impacting the
commercial utilized of borehole mining in coal application, the ultimate purpose of this research
is to establish a set of design protocols (guidelines) for estimating optimum cavity geometry and
orientation for several critical excavation and geological factors. In this context, optimal refers to
the maximum economic limits for extracting the resource in a defined volume while the
excavated cavity remains structurally stable for a specific period of time. In addition, the impact
of cavity failure on surface subsidence must also be quantifiable. The development of this design
methodology will contribute to a larger research effort which focuses on evaluating the economic
and technical development feasibility of in-situ borehole mining relative to thin seam deposits. In
the present stage, a parametric sensitivity study to investigate the effect of several parameters
like internal pressure, cavity size, etc. on the stability has been performed.
10 |
Colorado School of Mines | 1.5) Originality
Although a tremendous amount of research has been performed in evaluating the structural
integrity of underground openings, very little has been published on issues related to cavity
formation and stability as it pertains to the processes which comprise borehole mining. The
creation of these cavities in coal applications presents several unique technical challenges, where
a number of confounding parameters associated with borehole mining need to be assessed.
Addressing these parameters and developing a design methodology for estimating cavity stability
for a given set of geologic, geometric and operating characteristics will make this dissertation
unique and instrumental in advancing this technology. Paramount among these confounding
parameters are efforts to quantify the following questions:
1) The stability of the cavity during the extraction process is dynamic and influenced not
only by the geologic and geomechanical properties of the deposits but also by the mining
strategy, the mode and rate of extraction, the size and shape of the cavity (both
intermediate and final), and the physical interaction between adjacent and/or interrelated
excavations (cavities).
2) During the mining process, fluid pressure within the cavity will be controlled to facilitate
the collection/removal of cuttings and to assist in material excavation. The pressure range
may vary from 250 psi (17.2 atm) to nearly a vacuum (~ 0 atm). As such, questions exist
regarding the impact of these pressure fluctuations on the structural integrity of the
cavity.
3) One of the extraction techniques being considered utilizes superheated fluid and/or steam
as the primary fragmentation mechanism. As such, it is important to understand the
11 |
Colorado School of Mines | potential impacts that high temperature fluids and gases will have on the stability of the
cavity during excavation process.
4) As cuttings are removed from the cavity, methane gas and subsurface water will be
liberated from the host formation. Determining whether these phenomena influence the
structural stability of the cavity are issues that need to be quantified.
In this dissertation, only Item 1 has been performed. Items 2-4 will be studied in the
future works.
1.6) Research Limitations
Given the scope and complexity of this research topic, a number of limitations were
imposed. Paramount among these included assumptions regarding the research methodology
deemed necessary to simplify the analysis.
This dissertation is a part of a larger on-going research project to develop an integrated,
highly mechanized borehole excavation system. The focus of this dissertation seeks to address
one of the fundamental challenges that must be overcome to make this technology viable,
namely the development of a methodology to estimate the maximum geometry of a stable cavity
under specific geologic, geomechanical and operating conditions. As such, important factors
associated with the mode of excavation, the extraction (bailing) of cuttings, and the drilling
system are excluded from this dissertation. Although in-situ borehole mining can be applied in
different geologic settings and in the recovery of numerous mineral commodities, the scope of
this dissertation is limited to bituminous coal deposits within thin, flat seams.
12 |
Colorado School of Mines | In reality at least half of the cavity is filled with water and slurry during BHM. In this
dissertation, the cavity condition is assumed to be low or without fluid for the purpose of
numeric modelling. Since rock cutting is performed using jets with higher pressures and much
smaller flow rates (less than 100 gpm) than conventional hydraulic mining systems, water levels
in the cavity can be readily controlled during the mining cycle. In addition, the cavity will be
void of water nearly 50 percent of the time during the excavation process. Since in this
dissertation only thin seam deposits are considered, subsidence has been ignored.
It is important emphasize that the purpose of this dissertation is to establish an initial
preliminary design protocol based on simplified numerical models. As such, two-dimensional
software (Flac2D) was used and the input materials were assumed to be isotropic and
homogenous.
1.7) Dissertation Outline
Chapter 1 is the introduction to the dissertation and the research topic. This chapter also
includes the problem statement, thesis objectives, research limitations, a discussion regarding the
unique and original contribution of this work, and the organization of the dissertation.
Chapter 2 represents the results of the literature search. It includes an examination of
specific topic areas related to past research into borehole mining, cavity design and modeling
techniques, and the structural impact of several dynamic factors that may influence cavity
stability. Chapter 4 is also including explanation of some of the borehole mining operating
parameters that may impact cavity and borehole stability (e.g., strength deterioration).
13 |
Colorado School of Mines | CHAPTER 2
BOREHOLE MINING- LITERATURE REVIEW
2.1) Introduction
This chapter presents the results of a literature search on key elements relevant to the
dissertation objectives and has attempted to identify some of the confounding parameters
important to this discussion. In Chapter 3, these confounding parameters, as well as their
theoretical impact relative to this research will be discussed in detail. Meanwhile, the structure of
this chapter focuses on the following:
1) The definition of borehole and solution mining systems and alternative approaches
towards remotely extracting different minerals through one or more drill holes. Previous
work and studies into borehole mining tools, techniques, and operational parameters are
also presented,
2) Cavity measurements and stability assessment during borehole mining,
3) Fundamental geologic factors that are believed to be critical in the extraction of coal
using borehole mining applications, and
2.2) Borehole and Solution Mining
In general terms, in-situ mining is defined as the physical extraction of the valuable
components of a mineral resource through a borehole. Based on extraction techniques, in-situ
mining categories include:
a) In-situ Leaching
15 |
Colorado School of Mines | Solution Mining: Generally utilized in the extraction of salt and other soluble minerals. This
mining method forms a cavity as a consequence of the mineral constituencies within a designed
volume going into solution and being pumped from the excavation.
In-situ (leaching) Mining: Usually employed in the extraction of non-soluble or near-soluble
minerals. In most cases a cavity is not created but the localized area around the borehole
undergoes significant alteration. This alteration occurs as a consequence of a targeted mineral
being removed by a lixiviate or through some chemical process. These systems can utilize one or
more boreholes using a variety of differential leaching fluids, pressure gradients and operating
strategies.
b) Conventional Borehole Mining
This process is comprised of a waterjet nozzle assembly and a downhole slurry pumping
system. This system is designed to extract the target minerals through a single borehole drilled
from the surface. In this system, waterjets are strategically placed as part of a mining tool. The
function of these waterjets are to physically excavate the rock mass, place the cutting and
fragmented material in suspension, and then bail the resulting slurry to the surface. The slurry
will then be transported via pipeline to storage or a processing facility. A cavity will be created
as a consequence of this excavation process.
2.2.1) Solution Mining
The term solution mining usually refers to the extraction of impermeable salt and other
soluble deposits through one or multiple boreholes. While most impermeable deposits are mined
16 |
Colorado School of Mines | through the injection of hot water into the geologic formation, a wide range of other solutions
and chemicals can be used depending on the unique characteristics of the resource and
surrounding host rock. The creation of a void or cavity is the end-product of removing a mineral
laden slurry that contains dissolved and potentially dissolved solids from around the borehole. In
the case of salt, brine is pumped from the cavity to the surface; where the slurry will then be
processed to obtain marketable commodities.
Figure 2.1 shows an illustration of two alternative solution mining methods in extracting
salt from a confined resource. In both methods, a borehole is initially drilled to the final target
depth within the salt formation, where the casing is cemented and then perforated. Two leaching
strings are then inserted into an outer tubing. Heated fluid (i.e. hot water) is pumped through one
string while the excavated high-density brine is removed through the other.
In order to protect the roof of the cavern from being dissolved, a blanket medium
comprised of propane, butane, nitrogen, or air is pumped into the opening via the outer annulus.
Two different operational procedures are used to control the development of cavern shape: direct
circulation leaching and indirect (reverse) circulation leaching. In direct circulation, the fluid is
injected through the inner string and the salt laden brine is then pumped via the outer strings. In
reverse circulation, the fluid is injected through the outer string and the brine is pumped through
the inner annulus. By alternately applying both directions of circulation and shifting leaching
strings and blanket level, cavern shape and orientation can be remotely controlled in accordance
with the specific technical and commercial objectives of the operation and the geological
characteristics of the deposit.
17 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.