Networked Computing in Wireless Sensor Networks for Structural Health Monitoring

TL;DR

This paper develops distributed algorithms for efficient in-network computation of SVD in wireless sensor networks used for structural health monitoring, reducing energy use and delay.

Contribution

It formulates the distributed SVD computation as a clustering problem, proves NP-hardness, and proposes exact and approximate algorithms including a distributed version.

Findings

Approximate algorithm achieves near-optimal energy-delay trade-offs.

Distributed algorithm performs well in simulations and experiments.

Lower bounds established for the problem.

Abstract

This paper studies the problem of distributed computation over a network of wireless sensors. While this problem applies to many emerging applications, to keep our discussion concrete we will focus on sensor networks used for structural health monitoring. Within this context, the heaviest computation is to determine the singular value decomposition (SVD) to extract mode shapes (eigenvectors) of a structure. Compared to collecting raw vibration data and performing SVD at a central location, computing SVD within the network can result in significantly lower energy consumption and delay. Using recent results on decomposing SVD, a well-known centralized operation, into components, we seek to determine a near-optimal communication structure that enables the distribution of this computation and the reassembly of the final results, with the objective of minimizing energy consumption subject to a computational delay constraint. We show that this reduces to a generalized clustering problem; a cluster forms a unit on which a component of the overall computation is performed. We establish that this problem is NP-hard. By relaxing the delay constraint, we derive a lower bound to this problem. We then propose an integer linear program (ILP) to solve the constrained problem exactly as well as an approximate algorithm with a proven approximation ratio. We further present a distributed version of the approximate algorithm. We present both simulation and experimentation results to demonstrate the effectiveness of these algorithms.