Adaptive Decentralized Control of Underwater Sensor Networks for - PowerPoint PPT Presentation

Adaptive Decentralized Control of Underwater Sensor Networks for Modeling Underwater Phenomena Carrick Detweiler ‡† , Marek Doniec † , Mingshun Jiang ⋆ , Mac Schwager † , Robert Chen ⋆ , Daniela Rus † † Massachusetts Institute of Technology ‡ University of Nebraska-Lincoln ⋆ University of Massachusetts–Boston SenSys 2010, Zurich, Switzerland SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 1

Motivation: Underwater Sensing BP oil spill – riser pipe Image from reuters.com SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 2

Motivation: Underwater Sensing BP oil spill – extent is unknown SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 3

Motivation: Underwater Sensing Boston Harbor sewer pipe output Image courtesy Mingshun Jiang, UMass Boston SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 4

System Approach Depth adjustment enables: Many inexpensive sensors Easy deployment Networked for real-time feedback Easy recovery Collaborate with robot GPS or radio on surface Adjust depth for sensing using Optimizing position for: decentralized depth control algorithm Sensing (*this talk*) Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 5

Related Work Water column profilers Glenn et al. . The leo-15 costal cabled observatory – phase II for the next evolutionary decade of oceanography. Scientific Submarine Cable, 2006. Howe and McGinnis. Sensor networks for cabled ocean observatories. International Symposium on Underwater Technology, 2004. Joeris. A horizontal sampler for collection of water samples near the bottom. Limnology and Oceanography, 1964. Coverage and sensor placement Bullo, Cort´ es, and Mort´ ınez. Distributed Control of Robotic Networks. Applied Mathematics Series. Princeton University Press, 2009. Guestrin, Krause, and Singh. Near-optimal sensor placements in gaussian processes. International Conference on Machine Learning, 2005. Ko, Lee, and Queyranne. An exact algorithm for maximum entropy sampling. Operations Research, 1995. Schwager, Rus, Slotine. Decentralized, adaptive coverage control for networked robots. IJRR, 2009. Simulated underwater depth adjustment algorithms Akyildiz, Pompili, and Melodia. State-of-the-art in protocol research for underwater acoustic sensor networks. WUWNet, 2006. Cayirci, Tezcan, Dogan, and Coskun. Wireless sensor networks for underwater survelliance systems. Ad Hoc Networks, 2006. Related areas Drifting floats AUVs adjusting relative positions SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 6

Outline Motivation and Overview 1 Related Work 2 Approach 3 Decentralized Sensing Optimization Algorithm 4 Simulation Results AquaNode Underwater Sensor Network Experimental Results Future Work 5 Conclusions 6 SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 7

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Communication SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 8

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Improving Sensing SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 9

Contributions Decentralized depth control algorithm Optmizes depths for sensing Based on covariance measurements Provable convergence Low processing and communication Tested in simulation Implemented and tested on AquaNodes SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 10

Decentralized Depth Adjustment for Improved Sensing Measurement of water column properties Temperature, salinity, pH, dissolved O 2 , etc. Images Capture time-varying properties Constraints Power Minimize motion Minimize communication Acoustic communication bandwidth 11 bytes per packet Transmit just position, depth, and sensor reading SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 11

Covariance Measure q 1 from constrained path Changes at q 1 correlated to changes at p 1 Highest correlation when p 1 close to q 1 : Min ( Dist ( q 1 , p 1 )) More generally use covariance: Max ( Cov ( q 1 , p 1 )) Allows different sensing functions SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 12

Covariance Measure q 1 from constrained path Changes at q 1 correlated to changes at p 1 Highest correlation when p 1 close to q 1 : Min ( Dist ( q 1 , p 1 )) More generally use covariance: Max ( Cov ( q 1 , p 1 )) Allows different sensing functions SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 12

Covariance Measure q 1 from constrained path Changes at q 1 correlated to changes at p 1 Highest correlation when p 1 close to q 1 : Min ( Dist ( q 1 , p 1 )) More generally use covariance: Max ( Cov ( q 1 , p 1 )) Allows different sensing functions SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 12

Covariance Model Assume Gaussian Different variance along surface Better models with more knowledge SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 13

Covariance Model Assume Gaussian Different variance along surface Better models with more knowledge SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 13

Multiple Points Problem Idea: maximize sum over whole region: N � � Cov ( q , p i ) Q i =1 Problem: left and right are same: . 5 + . 5 + . 5 = 1 . 5 . 25 + . 25 + . 5 + . 5 = 1 . 5 SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 14

Multiple Points Solution Solution: invert sum of covariance and minimize � N � − 1 � � Cov ( q , p i ) Q i =1 Yields: . 5+ . 5 + 1 1 1 . 5+ . 25 = 2 2 1 . 5 = 3 . 5+ . 25 + 3 SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 15

Algorithm Approach Objective function: � N � − 1 � � H ( p 1 , ..., p N ) = Cov ( q , p i ) dq Q i =1 Decentralized gradient controller: p i = − k ∂ H ˙ ∂ z i � g ( q , p 1 , ..., p N ) 2 f ( p i , q )( z i − z q ) ∂ H ∂ z i = dq σ 2 Q d � N � − 1 � g ( q , p 1 , ..., p N ) = f ( p i , q ) i =1 � � ( xi − xq )2+( yi − yq )2 + ( zi − zq )2 − 2 σ 2 2 σ 2 f ( p i , q ) = Cov ( p i , q ) = Ae s d SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 16

Algorithm Convergence Each node moves according to: p i = − k ∂ H ˙ ∂ z i Theorem: decentralized controller converges to local minimum Proof: convergence proof using Lyapunov criteria H must be differentiable; ∂ H ∂ z i must be locally Lipschitz; H must have a lower bound; H must be radially unbounded or the trajectories of the system must be bounded. Verified in simulation, pool, and river experiments SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 17

Simulation Results: Versus Matlab’s fminsearch Total Search Time Objective Value 250 6000 fminsearch fminsearch 5500 distributed controller distributed controller 200 Obj Val (smaller better) 5000 4500 150 Minutes 4000 3500 100 3000 2500 50 2000 0 1500 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Number of Nodes Number of Nodes fminsearch: Matlab’s nonlinear unconstrained solver Much faster runtime Typically lower objective value SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 18

Posterior Variance 00 min 10 sec 01 min 40 sec 03 min 20 sec 05 min 00 sec 06 min 40 sec 0.83 0.82 posterior error 0.81 0.8 0.79 0.78 0 50 100 150 200 250 300 350 400 time [sec] Posterior variance Variance given sensor positions, assuming Gaussian process q | P = Cov ( q , q ) − Σ q , P · Σ − 1 σ 2 P , P · Σ P , q Requires matrix inversion ( O ( n 2 ) memory for n sensors) Decentralized depth control algorithm Tends to reduce posterior error Constant memory requirements SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 19

Simulation Results: Data Reconstruction Top row: Original data Bottom row: Depth adjustment algorithm SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 20

Outline Motivation and Overview 1 Related Work 2 Approach 3 Decentralized Sensing Optimization Algorithm 4 Simulation Results AquaNode Underwater Sensor Network Experimental Results Future Work 5 Conclusions 6 SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 21

Underwater Sensor Network: AquaNodes Multi-purpose underwater sensor network Acoustic, optical, and radio communication Easy to use and deploy Dynamic depth adjustment SenSys 2010–Carrick Detweiler (carrick@cse.unl.edu) 22