Abstract
Fault diagnosis of sensors is important as well as that in system dynamics. As a diagnosis method for sensors, generalized-likelihood-ratio (GLR) approach is proposed, which models the anomaly function appearing in the anomalous sensor by a step function. But, by the method, the discrimination of the anomalous type, such as gain or dead-zone anomalies, is impossible, and also, more important is that the step hypothesis is not always effective, because the anomaly function does not always have a meaningful bias over a round time.
From this point of view, the paper proposes another GLR method which estimates the anomaly function by using the system information about the state. By taking this approach, the number of the parameters to be estimated can be decreased compared to estimating the anomaly function by a linear combination of orthogonal series. This decrease in the number of parameters and the more accurate modelling of the anomaly function can make the fault diagnosis of sensors more easy. Actually, numerical examples show that the proposed method is superior to the conventional ones in detection rate, detection speed, the estimation of the anomaly occurrence time, and the discrimination of anomalous sensor and the type.
