A Robust Semi-Supervised Fisher Discriminant Analysis for Industrial Fault Classification

Owing to the advanced sensing technologies, data-based modeling has become a popular choice in the area of industrial process monitoring. For the data-based semi-supervised industrial fault classification problem, the samples from unobserved faults may severely degrade the performance of the classification model. Specifically speaking, on the one hand, the offline training samples from unobserved faults will act as outliers and seriously hinder the offline model training; on the other hand, the online testing samples from unobserved faults will be inevitably misclassified into observed fault categories in the online model usage. Despite the importance of these two issues, there are still no related works that can address them simultaneously. To this end, a robust semi-supervised Fisher discriminant analysis (FDA) model is proposed in this article. First, before the model training, based on the deviation information of training samples for each observed fault, a sample recognizing technique is designed to preprocess the training samples, with the purpose of recognizing the training samples from the unobserved fault. Second, to fully employ the supervised and unsupervised information hidden in preprocessed training samples, a regularized between-class scatter matrix and a within-class scatter matrix are constructed, and a semi-supervised FDA (SFDA) classification method is developed. Third, during the online model usage, the sample recognizing technique is also exploited to recognize the testing samples from the unobserved faults. Experiments on the benchmark Tennessee Eastman (TE) process and a real industrial air separation unit (ASU) demonstrate the effectiveness and the superiority of the proposed model.