A Robust and Sparse Process Fault Detection Method Based on RSPCA

As a method widely used in fault detection, principal component analysis (PCA) still has challenges in applicability due to its sensitivity to outliers and its difficulty in principal components (PCs) interpretation. In this paper, a robust sparse PCA (RSPCA) model is proposed to improve the robustness and interpretability of the PCA-based fault detection methods. Specifically, better robustness is achieved through capturing the maximal L1-norm variance while the sparse non-zero loadings are given for the PCs to achieve improved interpretability. A fault detection method is developed based on the RSPCA model, which uses the Genetic Algorithm (GA) to determine the number of non-zero loadings in each PC via the variance-sparsity tradeoff and defines T-2 and SPE statistics based on the selected PCs. In addition, an outlier removing strategy is proposed based on the SPE statistic, which can achieve further improvement to the robustness of the proposed method. The effectiveness and efficacy of the proposed method is evaluated by applying it to both two numerical simulation examples and the Tennessee Eastman (TE) process.