Center for Embedded Network Sensing

In this poster, we consider two major themes in the design of sensor networks: data integrity, and sampling strategies. For the data integrity problem, we propose a signature-based fault detection system for identifying both intermittent faults and persistent faults. Data-dependent features using temporal, spatial, and spatio-temporal information that are useful for detecting faults are identified. These features are combined into signatures that characterize each of the different fault types. We also discuss the problem of simultaneous parameter estimation and fault detection. In this case, parameters must be estimated from a distribution that is truncated in various ways as a result of the fault detection algorithm, which can lead to biased estimates. We propose several methods to account for the bias in parameter estimates. For the sampling problem, we describe two on-going projects. The first one deals with situations where sampling as you move (using sampling paths) is more effective than contemplating sampling points. A PAR sensor riding on a NIMS 3D node is one such situation. This configuration is especially well-suited for sampling phenomena that exhibit latent geometric structure, such as light fields in forest understories. We will consider the case where the phenomena can be approximated by a piecewise-constant field and suggest a novel estimation approach when we have sample paths as observations. The second project considers the problem of finding a sampling strategy to optimize the selection of the correct regression model from a set of competing regression models. The solution is driven by minimizing the probability of error in the selection and consists of a sequential algorithm that directs the collection of measurements. We develop an adaptive sampling algorithm to sample the field with a set of static sensors and one mobile sensor. The algorithm aims to jointly minimize the probability of error in the selection and the mobility cost. The algorithm presented provides a significant improvement in the probability of error in the selection of the correct model over the random collection of measurements.