Improved diagnosis of sensor faults using multivariate statistics

This paper analyses a variable reconstruction technique for identifying a faulty sensor. The reconstruction is associated with the application of principal component analysis (PCA) and attempts to remove "fault information" from the sensor reading. It is shown that the reconstruction (i) affects the geometry of the PCA decomposition (ii) leads to changes in the covariance matrix of the sensor readings and (iii) alters the determination of PCA based monitoring statistics in terms of their confidence limits. These changes must be incorporated into the monitoring scheme, as false alarms may otherwise be encountered. Consequently, an improved reconstruction based fault diagnosis is proposed here.