Fault Detection And Diagnosis Based On KICA And SVM

With the development of industrial of process, FDD (Fault Detection and Diagnosis) has become a critical issue. It is welcomed to avoid the fault process to cut down the loss through detecting, diagnosing and dealing with the fault. Data-based approach has been concerned extensively in recent years, because of the difficulty in establishing the accurate mathematical model from the complex industrial systems and the large amount of data recorded in the production process. In this thesis, an algorithm of fault detection and diagnosis based on KICA (Kernel Independent Component Analysis)-SVM (Support Vector Machine) is discussed. The main contribution is as follows:Firstly, the existing methods of fault detection and diagnosis are reviewed. Theory and usage of algorithms of the traditional KICA-SVM have been studied and focused. With the application of KPCA (Kernel Principle Component Analysis), the data whitening and dimension reduction are implemented, by which the overall distribution of the data is obtained. Feature extraction is carried with ICA (Independent Component Analysis) to extract the hidden information of a non-Gaussian process, by which the dynamic characteristics of the distribution is obtained. With the KICs, kernel density estimation is proposed for the estimate of joint probability density of them. Confidence interval is obtained after the set of confidence limit with the density for fault detection. In view of the new term of SRM (Structural Risk Minimization) in the machine learning theory, SVM based method is applied to get classification of multiple faults for fault diagnosis. However, the traditional KICA-SVM designed from the whole process data can hardly get a best effect of fault detection and diagnosis in every segment of the process as the distributions of data in the transient and steady process are quite different.Secondly, a method of segmented KICA-SVM is proposed to solve the problem above in this thesis. The algorithm of fault detection and diagnosis is designed in the segments of the process through a segmented method according to the dynamic characteristics of the distribution shown in the KICs. The effectiveness of the method is verified in the simulation result of TEP (Tennessee Eastman Process) which achieves a better result than the traditional method. However it requires a lot of resourses and time in training and diagnosis of SVM for the large amount of data in the KICs.Thirdly, a method of KICA-K means cluster-SVM is proposed in this thesis to solve the problem above. K-means clustering is used for cluster analysis for the KICs to remove the similar data for the reduction of consuming of resources and time in training and diagnosis of SVM, while the effect of this method declines. Thus, a segmented KICA-K means cluster-SVM method is proposed in which segmented KICA-SVN and the the KICA-K means cluster-SVM are combined. The effect of the method is verified in the simulation result of TEP. With a better effect, the consuming of resources and time are reduced in improved method. Therefore, the improved method has a better performance than the traditional method. It also shows the potential for the appliction in industrial process.At last, the thesis is summarized and the open problems for further research are also discussed.