Nonlinear Fault Detection And Identification Based On Kernel Forecastable Component Analysis

With the development of the modern industry,the industrial process is now becoming more and more complicated.Due to the highly nonlinear coupling of production process and complexity of the production data,the mathematical model and method based on qualitative knowledge are difficult to reach the actual application.Consequently,the data driven fault detection and identification method,especially the multivariate statistical method,for complex industrial process shows strong advantage in the handling large-scale industry data.The multivariate statistical method which can combine all kinds of data mining and machine learning methods to extract features from large number of historical data to obtain the process of hidden state information,as a result it can be widely utilized in the industry process monitoring.According to the characteristics of complex processes data and the shortcomings of the traditional multivariate statistical methods,the nuclear element(KFore CA)which is based on Entropy Spectral Density Minimization Theory is introduced into fault detection and identification area.The method considers the time sequence character of the data correlation and combines with the kernel processing method,which can extracts the dynamic time sequence characteristics and predictable data characteristics at the same time without the assumptions on the data statistical distribution.The KFore CA method can well deal with high dimensional nonlinear industry process data,comprehensively describe the trends and essential features of process data.In this thesis,the KFore CA method is studied as follows:1)In order to improve the deficiency of single kernel model in nonlinear process detection performance and overcome the problem of kernel parameters selection,a nonlinear fault detection method based on integrated kernel predictive element(KFore CA)is proposed through combining with the integrated learning theory.To obtain the different KFore CA sub models,firstly several kernel functions are selected.Then the Bayesian method is employed to convert the monitoring statistics of each sub model to the probability of fault.Finally,the results of each sub model are integrated to obtain the final result by the weighted combination strategy.The simulation results show that this method can effectively improve the detection effect and robustness of nonlinear fault detection.2)For improving the deficiency of the traditional fault diagnosis method in dealing with the sparse sample data,solving the problem of multi fault diagnosis and reducing the requirement of fault detection for data feature distribution,the support vector data description(SVDD)method is introduced in this section.The k NN nearest neighbor algorithm is used to make a more accurate description of the sample set,reflect the location relationship of the sample in the data,describe the contribution of different sample points to the spherical surface and solve the classification problem of imbalanced data.With utilization of the k NN method,the local density of sample points is calculated.To increase the weight of boundary sample points,reduce the weight of sample points in the class,the k NN is combined with the membership function of fuzzy theory to design a fuzzy support vector data description algorithm(Fuzzy Support Vector Data Description,FSVDD)based on local density.The FSVDD can improve the accuracy of the description of the target sample set,through setting the sample membership according to the weight of the sample.3)In order to avoid the influence of the accumulation of observation data on the fault identification model and improve the effect of fault identification,this section combines FSVDD with KFore CA to propose a fault identification model based on KFore CA-FSVDD.Firstly,KFore CA is employed to reduce the dimension of highdimensional data,and the nonlinear dynamic time series feature is extracted.Then the FSVDD model is established based on the extracted feature space,and both the single and multi fault identification are realized by the identification rules.This method not only retains the strong nonlinear feature extraction ability of KFore CA,but also combines the superior performance of FSVDD in the classification and fault identification,which can effectively improve the recognition model.4)In this section,the effectiveness of the proposed method in the field of GIS process monitoring and its potential applications are demonstrated.The EKFore CA based fault detection method and the KFore CA-FSVDD based fault identification method are applied in the field of Gas Insulated Switchgear(GIS)to detect and identify the GIS fault based on the closing coil current of the circuit breaker.