Fault Diagnosis Methods On Chemical Process

Fault diagnosis is an important part of industrial process monitoring.The goal of fault diagnosis is to detect abnormal behavior and find out root causes of faults.Since fault diagnosis can be viewed as a classification problem,existing classification techniques can be applied to it.In order to find out causes of a fault,a feasible way is to perform feature selection.This thesis focuses on feature selection methods in the “feature selection + classifier”model for the chemical process data.The “feature selection + classifier” model not only reduces the training time of classifiers through dimensionality reduction,but also improves the diagnostic performance of classifiers.The contributions of this thesis are summarized as follows.(1)For nonlinear data,the Gaussian kernel support vector machine recursive feature elimination(GKSVM-RFE)is introduced into fault diagnosis,and the diagnosis model based on GKSVM-RFE and support vector machine(SVM)is proposed.The key issue of this model is the choice of the Gaussian kernel parameter.In this thesis,we compare three kernel parameter selection methods to find the most suitable kernel parameter.The proposed fault diagnosis model can improve the performance of traditional diagnostic techniques.(2)Based on the fact that the data in the same class has similar distribution,we can assume that different distributions of the features in the different classes may contain usefully discriminative information.This thesis measures the difference between the two feature distributions by introducing the Kullback-Leibler(KL)divergence,and proposes a fault diagnosis model based on the KL divergence feature selection method and SVM.The proposed model can achieve good diagnostic performance.(3)A novel sequential backward feature selection method based on dissimilarity is proposed to detect the difference of features between the normal and fault data.The method can determine the importance of features by computing the dissimilarity of different feature subsets and find out the relevant features.The proposed fault diagnosis model has good performance on chemical fault diagnosis.The above fault diagnosis models combine feature selection with SVM to improve the performance of fault diagnosis.Experimental results on the Tennessee Eastman process dataset show that the proposed three fault diagnosis models are superior to the previous fault diagnosis techniques,and have better performance.In particular,the above feature selection methods can further identify fault-related features.