Manifold-Preserving Sparse Graph-Based Ensemble FDA for Industrial Label-Noise Fault Classification

Abstract

For the fault classification of chemical industries, the typical Fisher discriminant analysis (FDA) model requires that all the training samples should be correctly labeled. Actually, training samples tend to be polluted by mislabeled samples for some unavoidable reasons, and thus the performance of the FDA may be severely affected. However, in the chemical process monitoring community, there is little attention to fault classification with label-noise in training samples. To settle this open issue, a manifold-preserving sparse graph (MPSG)-based ensemble FDA model is initially proposed in this paper. First, under the condition of maintaining the underlying manifold structure of training samples, an MPSG is utilized to filter some mislabeled samples. Second, to improve model robustness to the left mislabeled samples, Bagging based on FDA is used to construct several subclassifiers, which are combined to form a robust classifier. Experiments on the benchmark Tennessee Eastman process and a real industrial air separation unit demonstrate the effectiveness and the superiority of the proposed model.