Here's a question for UAI listeners: >Bob, >Have you seen any work that with sensor validation or reconstruction >with Bayesian Networks? >By sensor validation I mean if a some sensor has a shift or increase in >noise etc., then >detect a sensor failure and possibly reconstruct the signal? >Thanks, >Jeffrey DeCicco, Ph.D. >Solutions Engineer >Gensym Corporation It seems there are several methods for sensor validation. One tradition is to look at the time spectrum as you mention. This would use a dynamic (Bayes net) model. Another is to validate the sensor based on the values of other sensors. If the Bayes net is modeled to predict the desired measurement and the sensor value is modeled as dependent on that and possibly includes fault behavior, then the Bayes net will predict the desired measurement. This prediction will rely on the sensor when the probability of fault is low and will deviate from the sensor when the fault probability is high. The probability of fault can be inferred from the readings of other sensors or, from drift and other changes to the time spectrum. Here are some references I found: "A probabilistic model for sensor validation" Ibarguengoytia, Sucar & Valdera, Proceedings of UAI, 1996. "Any time probabilistic reasoning for sensor validation," Ibarguengoytia, Sucar & Valdera, Proceedings of UAI, 1998 "Sensor Validation using Dynamic Belief Networks", Nicholoson and Brady, Proceedings of UAI, 1992. Bob Welch Gensym Corporation
