Manifold-Preserving Sparse Graph and Deviation Information Based Fisher Discriminant Analysis for Industrial Fault Classification Considering Label-Noise and Unobserved Faults

Fault classification is one of the most important topics in the area of industrial process monitoring. Existing industrial fault classification models based on Fisher Discriminant Analysis (FDA) and its variants still fail to cope with the issues of label-noise and unobserved faults simultaneously. To fill this important gap, a novel Manifold-preserving sparse graph and Deviation information based FDA (MDFDA) industrial fault classification model is originally presented in this paper. Firstly, the manifold-preserving sparse graph technique is utilized to filter training samples with label-noise; by keeping high connectivity between the obtained sub-graph and the original graph, not only the underlying manifold structure of training samples is preserved after being filtered, but also some mislabeled training samples with evident statistical characteristic deviations are filtered; in this way, the information quality of training samples is greatly improved. Secondly, based on the deviation information obtained from the filtered training samples, a deviation threshold is defined for each observed fault; if the deviation information of a testing sample is larger than the deviation threshold of each observed fault, it will be recognized as the testing samples from the unobserved faults. Experiments based on the data from the benchmark Tennessee Eastman process and a real industrial air separation unit demonstrate the effectiveness and superiority of the proposed MDFDA classification model.