Fault detection and operation mode identification based on pattern classification with variable selection

A novel method is proposed for fault detection and operation mode identification in processes with multimode operations. The proposed method employs the support vector machine as a classification tool together with an entropy-based variable selection method to deal with normal data clusters corresponding to multiple operational modes and abnormal data corresponding to faults. The use of the classification method in fault detection and operation mode identification allows us to build decision boundaries among the data clusters without the assumption of normal distribution. In addition, selection of variables by minimizing the total entropy of training data ensures superior generalization performance in classification. The performance of the proposed method is compared with that of the traditional PCA-based fault detection method using test data where class information is already known. While the existing method has produced considerable error rate (1.9% by T-2 chart and 24.2% by Q chart) in detecting faults, the proposed method has shown no error in this example in either fault detection or operation mode identification. Despite these outstanding results, it should be noted that the performance of the proposed method depends critically on the quality of the training data.