Uncertainty propagation in intelligent sensor validation

A framework and methodology for performing sensor validation in a data-driven on-line expert system is presented. Two related methods have been developed: Algorithmic Sensor Validation (ASV) and Heuristic Sensor Validation (HSV). ASV focuses on the numeric values of sensors, uncertainties related to sensors and sensory values, and probability distributions for the states of sensors. HSV identifies whether or not faulty sensory values are attributed to sensor failure or process failure. If caused by sensor failure, HSV modifies the probability distribution associated with the sensor according to the information provided by ASV and heuristic knowledge. This system has been implemented as a diagnostic on-line expert system that monitors the heat rate degradation in a fossil fuel power plant.;The "validated" sensor values in the form of discrete probability distributions are the input to an influence diagram knowledge base derived from the logic (fault) trees that the specific plant uses to identify failures. The purpose of this knowledge base is to provide expert advice in reducing heat rate degradation in fossil power plants. The experts' experiential knowledge is reflected in the set of conditional and marginal probabilities that make up the influence diagrams. The distribution on sensor states are discretized and used as an input to the influence diagram. The output of the influence diagram is a mapping from symptoms to the likely failure mode.;Measuring device error, environmental noise, and flaws or limitations in the data acquisition and processing systems are some of the sources of uncertainty in sensory information. In developing a data-driven expert system, as with most on-line diagnostic expert systems, sensor uncertainty must be propagated throughout the reasoning schemes. This highlights the important role that sensor validation must play in distinguishing sensor and signal uncertainty, and possible sensor failure, from process variations and failures. One method for quantifying sensor confidence is to calculate the similarity of the data stream to a known reference distribution of a normal sensor of the same class for a specific level of gross generation. The adjustment of sensor probability distribution is made based on the sensor confidence measure and uncertainties propagated through the resulting parameter values. The Bayesian method is used in updating the reference distribution from a sampled distribution. (Abstract shortened with permission of author.)...