Sensor fault detection using the Mahalanobis distance

A method is described by which a localized sensor abnormality can be detected using the Mahalanobis distance. The Mahalanobis distance is approximately the weighted distance from the hyperplane formed by the principal components to the particular observation. Qualitatively, the principal components correspond to the physical laws that govern the behavior of the systems and constraints placed on the system. If there are more sensors than principal components, there are redundant measurements. This redundancy can be used to detect abnormalities that are due either to sensor failure or a localized change in the system being measured. The method compares the distribution of the Mahalanobis distance during normal operation with the distribution during the current operation. A likelihood ratio test is then used to determine if a sensor has gone bad or if operations in the reactor are different from normal. The sensor whose value is not normal is identified by comparing Mahalanobis distances computed with one sensor masked. When the abnormal sensor is masked, the Mahalanobis distance for this subset of sensors will be within prespecified bounds. The method is demonstrated on 20 subassembly output thermocouples in the core of Experimental Breeder Reactor II.