Principle component analysis (PCA) based chi-square test is more sensitive to subtle gross errors and has greater power to correctly detect gross errors than classical chi-square test. However, classical principal component test (PCT) is non-robust and can be very sensitive to one or more outliers. In this paper, a Huber function liked robust weight factor was added in the collective chi-square test to eliminate the influence of gross errors on the PCT. Meanwhile, robust chi-square test was applied to modified simultaneous estimation of gross error (MSEGE) strategy to detect and identify multiple gross errors. Simulation results show that the proposed robust test can reduce the possibility of type II errors effectively. Adding robust chi-square test into MSEGE does not obviously improve the power of multiple gross error identification, the proposed approach considers the influence of outliers on hypothesis statistic test and is more reasonable.
A novel nonlinear combination process monitoring method was proposed based on techniques with memory effect (multivariate exponentially weighted moving average (MEWMA)) and kernel independent component analysis (KICA). The method was developed for dealing with nonlinear issues and detecting small or moderate drifts in one or more process variables with autocorrelation. MEWMA charts use additional information from the past history of the process for keeping the memory effect of the process behavior trend. KICA is a recently developed statistical technique for revealing hidden, nonlinear statistically independent factors that underlie sets of measurements and it is a two-phase algorithm: whitened kernel principal component analysis (KPCA) plus independent component analysis (ICA). The application to the fluid catalytic cracking unit (FCCU) simulated process indicates that the proposed combined method based on MEWMA and KICA can effectively capture the nonlinear relationship and detect small drifts in process variables. Its performance significantly outperforms monitoring method based on ICA, MEWMA-ICA and KICA, especially for long-term performance deterioration.
Multivariate statistical process control methods have been widely used in biochemical industries. Batch process is usually monitored by the method of multi-way principal component analysis (MPCA). In this article, a new batch process monitoring and fault diagnosis method based on feature extract in Fisher subspace is proposed.The feature vector and the feature direction are extracted by projecting the high-dimension process data onto the low-dimension Fisher space. The similarity of feature vector between the current and the reference batch is calculated for on-line process monitoring and the contribution plot of weights in feature direction is calculated for fault diagnosis. The approach overcomes the need for estimating or tilling in the unknown portion of the process variables trajectories from the current time to the end of the batch. Simulation results on the benchmark model of penicillin fermentation process can demonstrate that in comparison to the MPCA method, the proposed method is more accurate and efficient for process monitoring and fault diagnosis.