kNN Fault Detection Based on Multi-block Information Extraction and Mahalanobis Distance

In this paper, we propose a fault monitoring method based on multi-block information extraction and the Mahalanobis distance to address the problem that the traditional k-nearest neighbor fault detection algorithm only considers neighboring sample observation information. By mining accumulated and change-rate information in the original data, we improve the monitoring performance of the traditional k-nearest neighbor fault detection algorithm on faults caused by a small offset and pulse oscillation. Combined with observation data, we construct three types of information sub-blocks. Based on the Mahalanobis distance and Bayesian fusion strategy, new statistics are established for monitoring. The proposed method is numerically simulated and applied to monitor the Tennessee-Eastman and blast furnace processes, the results of which verify the effectiveness and monitoring performance of the proposed method.