Fault detection and classification through multivariate statistical techniques

A technique for on-line process fault detection and diagnosis based on statistical analysis of process data is presented in this paper. A principal component model is developed from the normal operating data and the process operation is monitored form the plots of squared prediction errors and scores. Through multivariate statistical data analysis, the features of various faults can be discovered and used in fault diagnosis. In the technique presented here, principal component analysis is performed for the data corresponding to each fault to extract the fault direction in the measurement space. Fault diagnosis is performed by comparing the direction of the current online measurements with those of various faults. The fault whose direction is very aligned with the current data direction is a plausible fault and is taken as the diagnosis result. Fault diagnosibility is investigated. The technique is very easy to implement and can be used to complement current fault diagnosis techniques. Applications of the proposed technique to the on-line fault diagnosis of a continuous stirred tank reactor (CSTR) system demonstrate that the technique is very effective.