Accuracy-Aware Aquatic Accuracy-Aware Aquatic Diffusion Process Diffusion Process Profiling Using Robotic Profiling Using Robotic Sensor Networks Sensor Networks Yu Wang, Rui Tan, Guoliang Xing, Jianxun Wang, Xiaobo Tan Michigan State University
Dec 27, 2015
Accuracy-Aware Aquatic Accuracy-Aware Aquatic Diffusion Process Diffusion Process
Profiling Using Robotic Profiling Using Robotic Sensor NetworksSensor Networks
Yu Wang, Rui Tan, Guoliang Xing, Jianxun Wang, Xiaobo Tan
Michigan State University
• Diffusion profiling• source location, concentration, diffusion speed• high accuracy, short delay
• Physical uncertainties– temporal evolution, sensor biases, environmental noises
04/19/2012 IPSN'12, Beijing, China 2
Harmful Diffusion ProcessesHarmful Diffusion Processes
Unocal oil spillSanta Barbara, CA, 1969http://en.wikipedia.org
BP oil spill,Gulf of Mexico, 2010
http://en.wikipedia.org
Chemicals/Waste Water PollutionUK, 2009, Reuters
04/19/2012IPSN'12, Beijing, China 3
Traditional ApproachesTraditional Approaches
• Manual sampling – labor intensive– coarse spatiotemporal
granularity
• Fixed buoyed sensors– expensive, limited coverage, poor adaptability
• Mobile sensing via AUVs and sea gliders– expensive (>$50K), bulky, heavy
04/19/2012 IPSN'12, Beijing, China 4
Aquatic Sensing via Robotic Aquatic Sensing via Robotic FishFish
• On-board sensing, control, and wireless comm.
• Low manufacturing cost: ~$200-$500
• Limited power supply and sensing capability
Smart Microsystems Lab, MSU
04/19/2012 IPSN'12, Beijing, China 5
Problem StatementProblem Statement
diffusion source
robotic sensors
•Maximize profiling accuracy w/ limited power supply
•Collaborative sensing: source location, concentration, speed•Scheduling sensor movement to increase profiling accuracy
04/19/2012 IPSN'12, Beijing, China 6
RoadmapRoadmap
• Motivation
• Background
• Profiling and Accuracy Modeling
• Movement Scheduling
• Trace Collection & Evaluation
yc
uxc
uz
cD
y
cD
x
cD
tc
yxzyx
2
2
2
2
2
2
04/19/2012 IPSN'12, Beijing, China 7
Diffusion Process ModelDiffusion Process Model
• Concentration at position (x,y,z) and time instance t
• Diffusion and water speed• Diffusion profile (source loc, α, β)
)exp(),( 2dtdc
04/19/2012 IPSN'12, Beijing, China 8
Sensor Measurement ModelSensor Measurement Model
• Sensor measurement• Actual concentration
– distance to diffusion source– elapsed time
• Sensor bias• Random noise,
nbcz
04/19/2012 IPSN'12, Beijing, China 9
Collaborative Diffusion Collaborative Diffusion Profiling Profiling
• Each sensor samples periodically• Samples from different sensors are fused
via Maximum Likelihood Estimation (MLE)
• How to model the accuracy of profiling? • How does the accuracy metric guide the
movement of sensors?
'4)''(
'
)(
21
22121
21
0
yLyLxLyLyLxL
xLxL
AxCRB
04/19/2012 IPSN'12, Beijing, China 10
Cramér-Rao Bound (CRB)Cramér-Rao Bound (CRB)
• Lower bound of estimate variance• Highly non-linear expression
e.g.
2121 ,,, yLyLxLxL
row vectors of all sensor coordinates
N
iiyCRBxCRB
A100 )(
1)(
1
04/19/2012 IPSN'12, Beijing, China 11
A New Accuracy MetricA New Accuracy Metric
• Sum of contributions of individual sensors
fixed in each profiling iteration node i's contribution tooverall profiling accuracy
),,( minddf ii distance b/w source
and sensor i min distance
to source
diffusion parameter
04/19/2012 IPSN'12, Beijing, China 12
Sensor Movement SchedulingSensor Movement Scheduling
Objective: find movement schedule for each sensor, s.t. profiling accuracy ω is maximized
Constraint:
• Movement Schedule: {orientation, # of steps}
MmN
ii
1
number of steps for sensor i
• Assign orientation– Find di
* that maximizes – If di > di
*, toward estimated source, otherwise
away from
• Allocate moving steps
– Maximize Σω(Δi), Δi – # of steps of sensor i
– Decomposition → dynamic programming
04/19/2012 IPSN'12, Beijing, China 13
Radial Scheduling AlgorithmRadial Scheduling Algorithm
di*
),,( minddf ii
diffusion source
robotic sensors
Putting All TogetherPutting All Together
1
2
3
• Collaborative profiling• Sampling• TX samples to node 2• Profiling via MLE estimation Estimated source location
• Movement scheduling• Orientation determination • DP-based step allocation
04/19/2012 IPSN'12, Beijing, China 15
Evaluation MethodologyEvaluation Methodology
• Trace collection– Rhodamine-B diffusion model– On-water Zigbee communication– GPS localization, robotic fish movement
• Trace-driven simulation– Profiling accuracy, scalability etc.
• Implementation on TelosB motes– Computation complexity
04/19/2012 IPSN'12, Beijing, China 16
Rhodamine-B DiffusionRhodamine-B Diffusion
discharge Rhodamine-B in saline water periodically capture diffusion with a camera expansion of contour → diffusion evolution
grayscale
model verification
04/19/2012 IPSN'12, Beijing, China 17
On-water ZigBee On-water ZigBee CommunicationCommunication
• PRR measurement using ZigBee radios on Lake Lansing
• 50% drop of comm. range compared to on land
04/19/2012 IPSN'12, Beijing, China 18
GPS and Movement ErrorsGPS and Movement Errors
• GPS localization errors– groundtruth vs. GPS measurement– average error is 2.29 m
• Robotic fish movement– 3m×1m water tank– tail beating frequency: 0.9 Hz,
amplitude: 23o
expected speed: 2.5 m/min
Linx GPS module
04/19/2012 IPSN'12, Beijing, China 19
Trace-driven SimulationsTrace-driven Simulations
• Profiling accuracy vs. elapsed time
profiling accuracy improves as time elapses
< SNR-based scheduling >orientation: gradient-ascent of
SNR# of steps: proportion to SNR
04/19/2012 IPSN'12, Beijing, China 20
Time ComplexityTime Complexity
• Implemented MLE estimation and scheduling algorithm on TeobsB motes
04/19/2012 IPSN'12, Beijing, China 21
ConclusionsConclusions
• Collaborative diffusion profiling using robotic fish– New accuracy profiling metric– Movement scheduling algorithm
• Evaluation in trace-driven simulation & real implementation
– High accuracy & low overhead
04/19/2012 IPSN'12, Beijing, China 23
Trace-driven SimulationsTrace-driven Simulations
• Profiling accuracy vs. number of sensors
profiling accuracy improves as more sensors are deployed