Research On Fault Detection Method Based On Canonical Correlation Analysis

Increasingly complex industrial processes pose new challenges to system security and stability.Fault detection technology plays an important role in reducing accident risk and improving system reliability.The fault detection method based on canonical correlation analysis(CCA)has better performance for objects with clear input-output relationship and has been widely studied in recent years.In this paper,the fault detection methods of complex industrial processes based on canonical correlation analysis are studied.The adaptive CCA method and kernel principal component analysis-CCA method are proposed for time-varying objects and nonlinear objects respectively.For the time-varying process,the canonical correlation analysis modeling method is improved and combined with adaptive factors to realize the real-time updating of the model.The size of adaptive factor is determined by data stability to improve model updating speed.On this basis,the fault detection logic combined with Mahalanobis distance classifier is used to realize fault detection of time-varying objects according to the flow of detection-update-detection.Simulation results in superheated steam model show that the proposed method is effective in detecting objects with variable working conditions and slow time-varying objects.In order to solve the problem that kernel matrix is generally used to replace the input data and the interpretability is poor,kernel principal component analysis is used to linearize and reduce the dimension of the nonlinear input data,and score matrix is used to replace the input data.On this basis,the kernel equivalent substitution principle of Gaussian kernel function is used to replace the score matrix with the original input matrix,and the canonical correlation regression model of the original input data and output data is directly established,which has strong explanatory ability.On the basis of the established regression model,the original input data is decomposed into two orthogonal subspaces which are completely correlated and completely independent with quality.Statistics are constructed in two subspaces to achieve quality-dependent fault detection.Finally,the simulation comparison between the proposed method and the existing nonlinear quality-related fault detection methods in the Tennessee-Eastman process shows that the proposed method is very sensitive to whether the fault affects the quality variable,and has better detection effect in the quality-related fault.