Research On Industrial Process Fault Diagnosis Based On Combined KPCA And Improved ELM

With the development of science and technology, the complexity of modern industrial process is higher and higher. The whole system may collapse if any small part of the system becomes abnormal. Therefore, how to take effective measures to ensure the security and reliability of the system and monitor the industrial process have become important problems. For the reasons above, fault diagnosis of industrial process becomes more and more important.This thesis uses Tennessee Eastman(TE) Process as the background and studies problems in industrial process fault diagnosis based on data-driven methods. As is known, fault feature extraction and fault classification are two important parts in fault diagnosis field. Therefore, this thesis improves the existing algorithms from two aspects of fault feature extraction and fault classification. In the end, the new algorithms are applied in fault diagnosis of Tennessee Eastman Process. The main work of this thesis includes the following aspects.(1) This thesis uses statistical method to extract fault feature. Firstly, feature extraction method based on principal component analysis(PCA) is illustrated.According to the disadvantages of PCA when extracting information of nonlinear data,kernel principal component analysis(KPCA) method is introduced. Furthermore, when mapping features to high dimensions, traditional KPCA with a single kernel function has its limitation. Therefore, an improved KPCA based on kernel function combination is proposed. The improved KPCA method combines Gaussian radial basis kernel function and polynomial kernel function as the new kernel function. As a result, the improved KPCA method has the advantage of both global kernel function and local kernel function. That is to say, the improved method has a stronger learning and generalization ability. Then, the thesis applies the above three statistical methods to fault detection of TE Process and evaluates the feature extraction capability of the three methods by fault detection results. Experiment results show that the improved KPCA method is effective and superior when being using for feature extraction.(2) According to the disadvantages of traditional classification methods in computing speed, extreme learning machine(ELM) is introduced. However, the input weight W and the hidden layer bias b are selected randomly, which may make the output matrix called H of ELM not full column rank. This disadvantage will reduce theclassification accuracy and computing speed of ELM. To avoid the disadvantage, an improved ELM algorithm is proposed. The improved ELM calculates the input weight W and the hidden layer bias b according to some certain rules. As a result, the output matrix H is ensured full column rank theoretically. Then traditional ELM and improved ELM are applied to UCI data sets. The experiment results show that the improved ELM has better classification effect and stability.(3) According to the combined KPCA and the improved ELM, a fault diagnosis model is established. In this model, combined KPCA is used to extract features of original data and the feature information is obtained. Then the improved ELM is used to classify the faults. To compare the fault diagnosis effect of different methods,algorithm of combined KPCA with support vector machine, algorithm of combined KPCA with ELM, algorithm of combined KPCA with improved ELM are applied to TE process for fault diagnosis. Experiment results show that algorithm of combined KPCA with improved ELM performs best.