Nonlinear Process Fault Detection And Diagnosis Method Based On Random Forest

The fault diagnosis of industrial process is of great significance to ensure the safe and stable operation of the process.The Random Forest(RF)method has been applied to the field of fault diagnosis.However,the method is mainly used to classify the process data,and the nonlinear,correlation and dynamic characteristics of the data are not considered.The strong coupling and dynamic characteristics of industrial process data have great influence on the accuracy of RF.In order to apply the RF method to the field of fault diagnosis,this paper studies the nonlinear fault detection and diagnosis method based on Random Forests.The main work of this paper is as follows:The traditional One-Class Random Forest(OCRF)method is influenced by the correlation between variables and dynamic characteristics,the outliers can't well represent the fault data,and the OCRF method with the original majority voting to detect fault,are unable to monitor the process.Aiming at the above problems,a process monitoring method based on Dynamic One-Class Random Forest(DOCRF)is proposed.This method uses a variety of random methods to divide the set of variables into multiple subsets.In each subset,the dynamic characteristics and correlations of the process data are extracted by the Canonical Variate Analysis(CVA)method,and the irrelevant typical features are obtained.The decision tree is trained to form the DOCRF model,and the calculation method of the monitoring statistic and the control threshold is given by finding similarity of the sample.This method can monitor the process well.In order to further diagnose of the fault and improve the diagnostic accuracy of the Random Forest method,a Canonical Rotation Forest(CRF)method is proposed in this paper.By applying the CVA method to the subset of the original data variable,the dynamic uncorrelated features of the data are extracted.The strength of the RF method is improved and the accuracy of diagnose is increased.Through the analysis of Tennessee Eastman(TE)process on the simulation results,the feasibility of CRF fault diagnosis method is proved,and its shortcomings are found at the same time.Aiming at the shortcomings of CRF method for the low accuracy of nonlinear,strong coupling fault,and the widespread of wide nonlinear relationship in modern industrial processes.We extend the CRF method to the nonlinear domain with the improved kernel CVA method,and the kernel CRF(KCRF)method is proposed.The original data is mapped to the high-dimensional feature space,and the principal component analysis(PCA)method is used to extract the kernel principal feature,which avoids the singularity problem that may be generated by using the kernel CVA directly.the CVA method is used to extract the dynamic correlation of data,and the irrelevant canonical variables are obtained;the KCRF is trained by the above variables,and it can be a good fault diagnosis process.Simulation of the TE process verifies the effectiveness of the fault diagnosis method based on KCRF.