Fault Detection And Classification Methods Of Analysis And Support Vector Machine Based On Robust Independent Component

In industrial process, in order to reduce product rejection rates and to satisfy increasingly safety and environmental regulations, there has been a large push to produce higher quality products. As product life cycles are getting shorter and international competition is getting acuter, so how to improve the quality and field of the product is the urgent problem to be solved. While methods of fault diagnosis based on data driven neither need modeling nor need precise priori knowledge, it deals with the running data of process including online data and offline data of the controlled system, mining implicit information and acquiring the running state of the process so as to complete fault detection and diagnosis.Independent component analysis (ICA) based on higher order statistics is one of the multivariate statistical process control methods which belong to data driven fault detection and diagnosis methods. Independent component analysis can extract process information and essentially describe process characteristics. It has been widely used in speech processing, biomedical signal processing, digital image feature extraction, telecommunications and econometrics field, etc.The main research work is organized in the following manner:1. According to the non-Gaussian characteristics of industrial process, ICA is deeply investigated including mathematical model, target function, estimation algorithm and engineering application. Three contaminated signals which are linearly mixed are given as the input of ICA model. According to non-Gaussian measurement principle based negentropy, FastICA is used to separate source signals. On this basis, statistics of I2and SPE are established. After that, kernel density estimation is utilized to determine the control limits. Whether the fault occurs in signal experiment or not should be judged by the value of statistics is beyond the control limit or not. The signal simulation and TE model simulation results demonstrate the effectiveness of the proposed method although there are some false alarms.2. As FastICA is sensitive to the selection of initial value, Robust independent component analysis (RobustICA) based on kurtosis which performs exact line search optimization of the kurtosis contrast function is put forward in this dissertation. RobustICA can deal with real and complex sources and escape local extrema, reaching very fast convergence. Three contaminated signals which are linearly mixed are given as the input of ICA model. Taking noise jamming into consideration, the basic principle of wavelet packet analysis is investigated. First, the noise of signals is wiped off using wavelet packet analysis. Then according to the non-Gaussian measurement principle based on kurtosis, RobustICA is utilized to separate source signals. On the basis of the first two steps, statistics of I2,Ie2and SPE are established and kernel density estimation is utilized to determine the control limits. The simulation results illustrate the superiority of the proposed method in comparison to FastICA. The above method is applied in TE model, reflecting the trend of the process and enhancing the monitoring performance in comparison to FastICA.3. Because of the excellent performance of solving limited samples, nonlinear and high dimension problems, the basic principle of support vector machines (SVMs) based on statistical learning theory and classification method are studied. The data of wine in UCI database is used to verify the effectiveness of classification method based on SVMs. In order to identify the fault types of TE model, one fault detection and classification method which combines of RobustICA and SVMs is put forward. The simulation results show the proposed method can effectively identify the types of faults.4. According to the nonlinear characteristics of industrial process, Robust kernel independent component analysis is put forward which introduces the kernel methods on the basis of RobustICA method, namely RobustKICA method. It nonlinearly maps the data from original space to feature space, RobustICA is used to analysis data and statistics of I2,Ie2and SPE are used to establish two combined indicesDa2andDb2, both of which are weighted sums of the three statistics in the feature space. The simulation results illustrate the proposed method can effectively and timely detect faults in comparison to PCA, FastICA and RobustICA.