Adaptive sparse principal component analysis for enhanced process monitoring and fault isolation

Principal component analysis (PCA) has been widely applied for process monitoring and fault isolation. However, PCA lacks physical interpretation of principal components (PCs) since each PC is a linear combination of all variables, which makes the fault detection difficult. Moreover, since the PCA model is time invariant while all real world processes are time varying and subject to disturbances. This mismatch may cause a false alarm or missed detection. Due to these motivations, we propose an adaptive sparse PCA (ASPCA) for enhanced process monitoring and fault isolation. which obtains sparse loadings by imposing a sparsity constraint on PCA. ASPCA with sparse loadings improves the interpretation and then facilitates the isolation of faulty variables. Meanwhile, ASPCA enhances model adaptability by updating the loadings with the sparsity constraint modified with changes in operating conditions. Next, a process monitoring and fault isolation strategy is presented based on ASPCA. Qusi-T2 and squared prediction error monitoring statistics are defined in the PC and residual subspaces, respectively. Nonzero variables in dominant PCs with most contributions to the fault are preferentially reconstructed. Case studies of TE process and waveform system demonstrate that the ASPCA method performs better in process monitoring and fault isolation compared to the PCA method.