Application of Sensor Fault Classification Algorithms to a Benson Steam Boiler

Abstract

This chapter describes a computer-based fault supervision method for the detection, localisation, estimation, and classification of soft sensor faults like biases, drifts, or scale-factor deviations. The algorithms are based upon the analytical redundancy between different measurements, given in form of a mathematical model of the process to be controlled. On the assumption that this model is exact, a basic instrument fault diagnosis scheme is designed which uses statistical hypothesis tests and recursive least-squares algorithms to detect, estimate, and to classify sensor faults. However, the most significant problem of such a model-based fault detection scheme is the separation of the interesting sensor faults from unavoidable modelling errors and from process parameter variations. To overcome this problem, we use state-augmented and hypothesis-conditioned Kalman filters to track time-varying process parameters, and de-correlation filters to suppress dynamic modelling errors. It is shown that the detectability and separability of different sensor faults, described by some figures of merit, depend highly on the number of measurements to be supervised and the number of tracked process parameters. For the application to the supervision of temperature measurements of a steam boiler used in power stations, some results from off-line processing of real measurement data are given. However, the algorithms, implemented on a personnel computer, are fast enough to allow on-line processing of the measurements.