Abstract
A multiblock multiscale multiway principal component analysis (MSPCA) modeling approach is presented for multivariate statistical process performance monitoring of batch processes. Process measurements, representing the cumulative effect of many underlying process phenomena, are decomposed into scales using wavelet transformations. The decomposed process measurements are then arranged into blocks of scales and approximations. MSPCA is then used to build a model that can be used for fault detection and identification of incipient sensor and process malfunctions. The proposed method was evaluated on a benchmark fed-batch penicillin fermentation process for the detection and diagnosis of subtle sensor and process faults, and the results were compared with those obtained using conventional multiway PCA (MPCA). This study contributes to the discussions of the advantages and benefits from using multiscale-based multivariate statistical process monitoring schemes over conventional multivariate statistical process control approaches.