Non-Gaussian Process Performance Monitoring And Fault Diagnosis For Industrial Units

Performance monitoring and fault diagnosis is the key elements of CIPS(Computor Intergrated Process System) in process industry. In order to improve the product quanity and ecomomic benefit, the process condition should be closely monitored and fault information should be timely detected. In this paper, the structure of an industrial monitor system is divided into three stages of fault detection, fault diagnosis and fault isolating. An online process monitoring approach merging the three parts in a real-time condition is proposed. In view of the measurement variables of chemical process usually show the characteristic of nonlinear and non-Gaussian behaviors. Some improvements of kenel principal componet analysis (KPCA) have been made at different degree, and also intergrated the support vector data description (SVDD). Some new monitoring algorithms are proposed.Then relevant vector machine (RVM) is be used in fault classifier to perform the second-step indentification and system alarm. Finally, the source of fault are been researched by fault isolating technology. The main contribution of this thesis is as follows:(1) A novel modeling method was proposed by integrating the improved KPCA with SVDD. a) The Mexican hat wavelet function was introduced to construct the kernel function by utilizing the advantage of extracting the subtle feature of nonlinear non-stationary signal. The nonlinear mapping and anti-noise capability of kernel function was enhanced. b)Then the cluster analysis was used in the kernel feature space. The data that represented the characteristic center in every cluster were chosen, which can decrease the computational complexity and improve the results of real-time monitor. Furthermore, the SVDD was adopted to describe the feature space with dimension-reduction, and a new monitor index was constructed by SVDD to describe the non-Gaussian information. The method was applied in Tennessee-Eastman benchmark process and a solvent dehydration rectification process, which is very effectively.(2) Through the process can be detected by the KPCA, it can not know that the process belong to the category of fault. In view of this problem and few fault data, an integrated strategies for fault diagnosis is presented, which extract feature vector combines by kpca and then combined RVM to build the diagnosis model based KPCA-RVM. On the other hand, a new multi-class classifier with simplified structure is presented. The next stage is to find the reason of fault. Contribution plot is a fault isolating method based by KPCA, it consider the revelance of space and change the one-variable stasticticts technology. This methods quantitate each process variables which is related contribution. According to the total contribution CONT value, it can distinguish fault source and judge the reason of losing control. The Tennessee-Eastman process simulation result is effective.