Bidirectional sparse principal component analysis for process monitoring

Sparse principal component analysis (SPCA) was proposed to strengthen the interpretability of traditional PCA. However, SPCA remains “semi‐sparse” due to the lack of sparsity in its regression coefficient matrix, which can reduce the interpretability and accuracy of process monitoring results. To address this issue, we propose bidirectional SPCA (BSPCA). To solve the BSPCA modelling challenges, we investigate the matrix structure features and propose an improved particle swarm optimization (IPSO) strategy. Testing on the Tennessee Eastman process demonstrates that BSPCA significantly improves sparsity by eliminating less important relationships between variables. In addition, BSPCA is able to pinpoint specific faults in the relevant process variables, thus highlighting its diagnostic effectiveness.