Abstract:
Multivariate statistical monitoring method uses normal operation data to select features. However, in practical processes, different faults will affect different features, and these features may change over time and the action of the control system. When faults occur and change over time, in order to achieve better fault detection abilities, it is necessary to gather effective fault-sensitive features. This paper presents a two-layer adaptive ensemble residual principal component analysis model (AERPCA), whose sub-models consist of different features and prominently present one or several related faults. First, the principal component analysis (PCA) features are calculated according to normal data, and different features are used to construct the linear sub-models and the corresponding residual spaces. Considering that the nonlinear and effective features of the residual space are more dispersed, kernel PCA (KPCA) is used for extracting and forming different KPCA sub-models in the same area by the second feature selection. Then, Bayesian method is used to form an integrated KPCA sub-model and complete the division and integration of each residual space. Finally, after obtaining multiple PCA sub-models in the main space and the integrated KPCA sub-models in the residual space, the sliding window is used to determine the model with the best monitoring effect at the current time. Tennessee Eastman process is used to verify the effectiveness of AERPCA.