Multimode process fault detection method using local neighborhood standardization based multi-block principal component analysis

Traditional PCA assumes the unimodal distribution of process data and may not perform well in monitoring the process with multiple operation conditions. In order to detect the faults in multimode process, this paper proposes a novel process monitoring method using local neighborhood standardization strategy based on multi-block principal component analysis (LNS-MBPCA). Firstly, the proposed method applies local neighborhood standardization (LNS) strategy to transform the multimode data into unimodal data. Then, all the monitored process variables are partitioned into three sub-blocks based on their correlation with global principal components. By utilizing the variable division results, global PCA model is divided into three PCA sub-models and the local monitoring statistics are constructed. Lastly, Bayesian inference is applied to integrate the monitoring results of each sub-block and two probability-based monitoring statistics are developed to indicate process status. Simulation results on a simulated continuous stirred tank reactor (CSTR) system show that the proposed method has better process monitoring performance.