Study on online outlier detection method based on principal component analysis and Bayesian classification

Outliers detection is an important part of the online model prediction. Due to the difficulty in determining the suitable control limits for traditional PCA method for the outlier detection, an online outlier detection method is presented based on principal component analysis and Bayesian theory. Firstly, principal component analysis (PCA) is used to calculate Q statistics with the training data collected in the normal process. Secondly, using the priori knowledge and the sample data which is updated by sliding window technology, the Q statistic is classified from the normal process and the disturbance process by the Bayesian classification method. If the current sample is from the disturbance process, it should be further determined that the value is caused by the case of the abnormal value or the process changes, which realizes the online outlier detection for the process data. The simulation using the data from the UCI machine learning repository shows that the proposed method has the lower misjudgment rate compared with the traditional PCA method, and it can effectively identify the abnormal values and process changes in the process data. The simulation result verifies the effectiveness of the proposed method.