1 / 33

Sensorweb Architecture and Dynamic Sensor Tasking in Mobile Sensor Networks

Sensorweb Architecture and Dynamic Sensor Tasking in Mobile Sensor Networks. Sanjoy K. Mitter, Massachusetts Institute of Technology Joint work with: Maurice Chu (currently at PARC, Palo Alto, CA) and Peter Jones (Lincoln Laboratory).

justis
Download Presentation

Sensorweb Architecture and Dynamic Sensor Tasking in Mobile Sensor Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Sensorweb Architecture and Dynamic Sensor Tasking in Mobile Sensor Networks Sanjoy K. Mitter, Massachusetts Institute of Technology Joint work with: Maurice Chu (currently at PARC, Palo Alto, CA) and Peter Jones (Lincoln Laboratory) SensorWeb MURI Review Meeting, Dec 2, 2005 Phone 617-253-2160, 617-258-8364 (Fax) Email mitter@mit.eduWeb http://web.mit.edu/mitter/www/ MURI Review: SensorWebData Fusion in Large Arrays of Microsensors Dec 2, 2005

  2. Sensorweb Architecture … • Extension of Maurice Chu’s Ph.D Thesis • Current research at PARC (Palo Alto, CA) • S.M. Thesis of Peter Jones (currently at Lincoln Laboratory, MIT)

  3. Distributed Attention for Sensor Networksand the Beginnings of a Conceptual Framework for Designing Resource Aware Distributed Algorithms Maurice Chu PARC

  4. Research Vision • Goal Move from specialized information processing systems engineered for specific domains toward general-purpose systems embedded in unstructured, dynamic environments • Framework for managing the complexity in designing information processing algorithms for data interpretation. • Complexity can come from • limited resources of the system like processing power, communication bandwidth, sensing capabilities, energy, link/node failure characteristics • application requirements – latency, scalability, robustness to link/node failures, reliability, scalability • inherent algorithmic complexity to extract information • Challenges • Information architecture • Mapping to physical system (e.g., distributed implementations) • Efficient representations of information • Interpretation algorithms (estimation, detection, inference)

  5. Wireless communication device processor Sensors Sensor Networks • Dense ad hoc network of heterogeneous sensor nodes (equipped with sensing, processing, communication) • Enormous potential for extracting all kinds of information from data sensed about the environment • Applications: surveillance, intelligent transportation grids, factory monitoring, battlefield situational awareness Limited sensing coverage, processing power, and communication bandwidth. Need collaborative in-network processing capabilities.

  6. PARC IDSQ Tracker in Action SensIT Experiment (video), 29 Palms, MCAGCC, November 2001 Tracking result (right) from post-processing acoustic amplitude data from 21 Sensoria wireless nodes (yellow dots). For more info: www.parc.com/ecc

  7. A few sensors can relatively easily find and track a vehicle in the desert. But how do we find and monitor anything amongst all these distractors and clutter? Moving to Complex Environments

  8. Large Scale Video Surveillance • GoalEnable video surveillance of large complex environments to monitor and detect multiple potential threats and surprises. • Difficulties • Unstructured environmentDistractors, clutter, occlusions • Information Overload • Human operators overwhelmed • Resource limitedInsufficient sensing, processing, and communication resources to monitor all phenomena over extended periods Solution: Distributed Attention Inspired by biological focusing mechanism

  9. Challenges • Information ProcessingHow do we efficiently represent and monitor known dynamic phenomena? • Abnormal BehaviorHow do we learn what is abnormal behavior, detect them, and react to them? • Peripheral AwarenessHow do we maintain awareness of newly emerging unboserved phenomena? How do we model the “emergence” of phenomena? • Resource AllocationHow do we share limited sensing resources among multiple competing tasks? How do we evaluate task priorities and what kinds of negotiations must occur to allocate resources optimally? • Distributed ImplementationHow do we implement algorithms in a distributed fashion under bandwidth constraints, energy considerations, processing time, latency constraints, etc.? What is the information exchanged and how does it flow through the network? • AdaptationHow can the system adapt to changes in the environment and failures and additions to the network?

  10. Outline • Problem Statement and Challenges • Distributed Attention ArchitectureConceptual view • Testbed Implementation • Future Work

  11. Outline • Problem Statement and Challenges • Distributed Attention ArchitectureConceptual view • Testbed Implementation • Future Work

  12. Layered Information Processing Architecture Adaptation Peripheral Awareness tasks tasks Resource Allocation controls Sensing observations Distributed Attention Architecture • Layered Information Processing Architecturedata interpretation unit (detection, estimation, inference) • Peripheral Awareness Moduleenables attention to the unobserved emergence of abnormal phenomena • Resource Allocationallocates tasks to resources • Adaptation - evolve system behavior according to dynamic system characteristics and environment

  13. Cognitive behavior recognition groups of tracks attacker identities Attentive adjust tracking priority tracking anomalous flow known track position Pre-attentive ignore anomalous flow detect flow to landmark optical flow sensor observation Knowledge level Signal level Layered Information Processing ArchitectureConcept • Global view of the transformation from data to information • Multi-layered filtering approach • Layers loosely ordered from continuous signal-level representations to discrete symbolic representations • Organize distinct information processing tasks into separate layers (modularity) • Two-way layer interactions • Bottom-up triggering • Top-down priming • Information flow considerations for distributed implementation • Lower layers – little cross node communications, high bit-rate local data • Higher layers – cross node communications, low bit-rate global data

  14. Cognitive behavior recognition groups of tracks attacker identities Attentive adjust tracking priority tracking anomalous flow Knowledge level known track position Pre-attentive ignore anomalous flow detect flow to landmark optical flow sensor observation Signal level Node 4 Node 3 Node 1 Node 2 Cognitive Cognitive Cognitive Cognitive Information flow through sensor network behavior recognition behavior recognition behavior recognition behavior recognition groups of tracks groups of tracks groups of tracks groups of tracks attacker identities attacker identities attacker identities attacker identities Attentive Attentive Attentive Attentive adjust tracking priority adjust tracking priority adjust tracking priority adjust tracking priority tracking tracking tracking tracking anomalous flow anomalous flow anomalous flow anomalous flow known track position known track position known track position known track position Pre-attentive Pre-attentive Pre-attentive Pre-attentive ignore anomalous flow ignore anomalous flow ignore anomalous flow ignore anomalous flow optical flow optical flow optical flow optical flow detect flow to landmark detect flow to landmark detect flow to landmark detect flow to landmark sensor observation sensor observation sensor observation sensor observation Layered Information Processing ArchitectureConcept • Transitioning to a distributed system architecture • Vertical cuts of layers • Communications across vertical boundaries within layers

  15. Distributed Attention ArchitectureConceptual View Layered Information Processing Architecture Adaptation Peripheral Awareness tasks tasks Resource Allocation controls Sensing observations

  16. Bernoulli process generates suspicion samples Propagation model simulates motion of suspicion samples Peripheral Awareness ModuleConcept • Models the emergence and propagation of abnormal behavior for intelligent focusing of attention on unobserved, emerging events. • Ex. Monte Carlo simulation of a stochastic process(flow of suspicion samples) • Mechanics • Sensed areas clear suspicion samples. • Detection of abnormal behavior handled by the information architecture • Effect • Potential emerging targets compete for resources with known targets

  17. LSP Progress Review Presentation Peter JonesMaster’s Student June 16, 2005

  18. Completed Thesis • Dynamic Sensor Tasking in Heterogeneous, Mobile Sensor Networks • Goal of time-optimal detection/discrimination in sensor networks • Extension of previous work in using conditional entropy/mutual information for sensor tasking • Methods applicable to multi-modal sensors • Coordination protocol developed for exploiting sensor inter-dependencies • Accepted May 5, 2005 in fulfillment of the requirements for EECS Master’s Degree

  19. Background

  20. Information Driven Sensor Query • Intended to limit communication (and power usage) in sensor networks • Compromises quality of inference (detection, tracking, etc.) for computational and communication simplicity • Primarily applied to static networks of power-limited, homogeneous sensors • Basic principle: choose a new “leader” in the neighborhood of the current leader to maximize the expected information gain of the next sensory action

  21. Multi-Modal Sensor Management • Tsitsiklis, Popp, Bailey (MIT/Alphatech) • Considered two discrete modes (HRR v. GMTI) • Optimized for footprint location (continuous variable) • Kreucher, Kastella, Hero (Veridian/UMich) • Use of information measure (Renyi entropy) • Considers only discrete modes • Similar to IDSQ (choose mode to maximize expected entropic change)

  22. Contributions

  23. Minimum Time Formulation • Optimization approach to multi-sensor scheduling • Definitions of objectives, constraints and actions • Objective: to finish in minimum (expected) time • Constraint: solve inference problem within a user-specified level of uncertainty (entropy) • Actions: deploy or query schedules for one or more sensors with a chosen set of sensor parameters • Optimization Equation

  24. Maximum Rate of Information Acquisition • Dynamic analysis leads to dynamic programming solution method • Allows for information feedback (sensor measurements) in decision process • Provably optimal action for stationary and decomposable underlying distribution when entropy is large compared to

  25. Coordinated Scheduling • Coordination Protocol • Iterative algorithm with bounded Pareto sub-optimality • Uses information theoretic utilities • Market-based negotiation • Axomiatic bargaining principles enforce “fairness” Coordination Algorithm • Each sensor chooses “ideal” jointaction from set of possible joint actions, Sk • Ratio of “ideal” utilities leads to mixed operating point • Set of dominating joint actions identified, Sk+1 • If Sk+1empty, end; else repeat process

  26. Results

  27. Simulation Setup • Indeterminate number of targets • Discrete number of possible locations • Time for measurements increases linearly with measurement area radius

  28. MIAR results

  29. MIAR results II • Multi-modal Simulation • Constants a,b now a function of mode • Results of 50 monte-carlo simulations

  30. Coordination Experiment Setup Cannonical Problem Definition: Two sensors, three locations, one common between the two. If both sensors attempt to measure the common location, neither receives a good measurement (mutual jamming).

  31. Coordination Results

  32. Advanced Coordination Experiment • Set of 5 heterogeneous sensors, each with different detection/discrimination characteristics • 25 possible target locations, 3 different target types • Entropic threshold set low enough that there were no missed detections or incorrect classifications in any of the 150 trials

  33. Summary • Information driven sensing helpful in groups of (possibly interacting) multi-modal sensors • Extension of entropy-based measures to time-optimal scheduling • Viable coordination protocol • Verification of entropy-based utilities and coordination via simulation

More Related