Fault Monitoring And Diagnosis In Industrial Process Based On MSPCA-KECA

With the continuous improvement of industrial automation, intelligent instruments and distributed control systems have been widely used in practical industrial process. Moreover, massive data has been collected and saved, which reflects the working conditions. It has been a research focus in fault detection and diagnosis technology that historical data is used for modeling to extract the statistic laws of normal or fault conditions for fault detection and diagnosis of current industrial process. However, these data are nonlinear, uncertain of data distribution and multi-scale, because of automatic systems of industrial process are large and complex. So, most traditional multivariate statistical monitoring algorisms are applied with several ideal assumptions of process data, which has some limitations. Aiming at some practical problems in process industry, traditional multivariate statistical monitoring technology is improved, and a new fault detection and diagnosis algorism based on MSPCA-KECA is proposed in this paper, which solves the problems like multi-scale, nonlinear and uncertainty of data distribution. The main contents of this thesis are summarized as follows:1) A fault monitoring method based on Kernel Entropy Component Analysis(KECA) is presented to solve the nonlinear and uncertainty problems of the data distribution. Data is analyzed directly with no need to assume the data distribution. The introduction of Kernel function solves the nonlinear problem of data. The number of principal components selected by the KECA algorism is much less than the KPCA algorism, which can effectively reduce computational complexity. This is achieved by selections onto eigenvalue and eigenvector based on the value of Renyi Entropy. Research shows that KECA reveals angular structure of the input space data set, based on which A new statistic——Cauchy-Schwarz divergence measure is proposed. The statistic and structure of features extracted using KECA complement each other. The simulation results show that it has a better performance in fault monitoring compared with the traditional method and statistics.2) Aiming at the multi-scale of process data, a data preprocess algorism based on improved Multi-scale Principal Component Analysis(MSPCA) is proposed. When fault occurs at different time, different location or with various levels, it appears different in the changes of process variables, which is the multi-scale feature in data. Data is decomposed to several scales with wavelet analysis method. Scales which may contain fault-related information are selected by using PCA algorism. And then data dimension reduction and feature extraction are achieved by scales reconstructed with a second application of PCA. The simulation results are shown that this method can extract multi-scale information effectively and highlight the changes in variables.3) A new process monitoring strategy combining MSPCA and KECA is proposed. On the one hand, the characteristic information of the process is more prominent after data preprocessing, which is conducive to the fault monitoring and diagnosis. On the other hand, KECA is used as single classifier, which is to say that each KECA classifier is dedicated to a specific fault, with the advantages of simple model, fast computing speed, high fault recognition rate, convenient update and so on. The simulation results are shown that this algorithm is of great practical value by applying to Tennessee Eastman(TE) process and ASHRAE 1043-RP chiller process. It has also been found that the algorithm can effectively solve the misdiagnose problem when different levels of the same type for Chillers faults occurs.