Collaboration and Pattern Recognition in Distributed Sensor Networks: A statistical mechanics based approach
Open Access
- Author:
- Srivastav, Abhishek
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 02, 2009
- Committee Members:
- Asok Ray, Dissertation Advisor/Co-Advisor
Asok Ray, Committee Chair/Co-Chair
Alok Sinha, Committee Member
Shashi Phoha, Committee Member
Mark Levi, Committee Member - Keywords:
- Symbolic dynamics
Statistical mechanics
Information gain
Markov random fields
Sensor networks
Anomaly detection
Probabilisitc finite state automata - Abstract:
- Recent advances in the technologies of microcomputers and wireless communications have enabled usage of inexpensive and miniaturized sensor nodes that can be densely deployed in both benign and harsh environments as a sensor network for various applications. Distributed sensor networks can be used for both military (e.g., target tracking and surveillance) and civilian (e.g., weather, habitat and pollution monitoring, and structural health monitoring) applications. This dissertation addresses the issues of collaboration and pattern recognition in sensor networks. At the node level, the task of individual sensor node is to extract and identify primary temporal patterns contained in the time series data. An approach based on symbolic dynamic filtering is proposed for extraction of local patterns that are represented as probabilistic finite state automata (PFSA). A new measure of information gain is proposed based on the concepts derived from statistical thermodynamics, symbolic dynamics and information theory. The efficacy of this measure is demonstrated for anomaly detection in complex systems with known and unknown structures. Primary patterns, detected at the node level, may contain only partial information of the corresponding event and may not be sufficient for event identification and tracking. Also, sensor nodes in a network are often resource-constrained. Thus, at a higher level, collaboration among nodes is necessary for pattern identification, task allocation and division of labor. A mathematical framework - interacting Probabilistic Finite State Automata (i-PFSA) is proposed for modeling and analyzing collaborative principles in sensor networks. The sensor network is modeled as a Markov Random Field (MRF) and each node is represented as a discrete-event system with a finite number of states. Using a mean-field theoretic approach, interactions are modeled in terms of their node dynamics or the state occupation probability vectors. Each node is represented as an i-PFSA, that interacts with the dynamics of its neighboring nodes and takes events from the environments as inputs. This is a generic formulation for task assignment and collaboration and the proposed methodology is applied to address collaboration issues in sensor networks in simulated and laboratory settings. The proposed research takes a multidisciplinary approach for both pattern recognition and collaboration in sensor networks. A decentralized organization and control is envisioned for enhanced robustness and scalability of sensor networks. Moreover, a bottom-up organizational principle, found in many natural systems, allows for adaptability to localized or isolated disturbances in the operational environment without causing an overall change of the network dynamics.