Industrial Systems Anomaly Detection Based On Deep Support Vector Data Description

With the increasing scale and complexity of industrial processes,anomaly detection of industrial systems plays an increasingly important role in ensuring process safety and improving product quality.In recent years,due to the wide application of computer technology,a large number of operating data have been stored.Therefore,data-driven anomaly detection method of industrial system has become one fascinating topic.However,because of the complexity of industrial processes,traditional anomaly detection methods are difficult to achieve ideal detection results.In view of the large scale and high dimension characteristics of complex industrial process data,this paper investigates an anomaly detection method for industrial systems based on deep support vector data description,and verifiy the method with the test data.Firstly,an anomaly detection method based on ensemble deep support vector data description(EDe SVDD)is proposed to deal with the problem of poor detection of single De SVDD.By constructing deep SVDD sub-models with different network structures and initialization parameters with Bayesian integration strategies,a comprehensive monitoring model is constructed,which eliminates the uncertainty of process monitoring.The simulation results of the MNIST,Fashion-MNIST,Tennessee Eastman(TE)process,and WM-811 K data set have shown great improvement compared with the detection effect of deep SVDD.Then,aiming at the problem that the deep support vector data description fails to effectively retain the original data structure,a deep data structure preserving support vector data description(DSPSVDD)method is proposed.In this paper,the decoder structure and selfsupervised strategy are introduced to further retain the structural information of the original data.The simulation results of MNIST,Fashion-MNIST and WM-811 K data sets have shown effective improvement with the above problems,indicating that the proposed method can effectively retain the original data information.Finally,aiming at the problem that the above methods fail to make full use of the prior information,an anomaly detection method based on semi-supervised deep data structure preserving support vector data description(Semi-DSPSVDD)model is proposed.Firstly,we establish the main model with normal samples,and then get the abnormal discrimination submodel with priori abnormal mode.Finally,the weighted Bayesian strategy is applied to obtain an overall monitoring statistic,so as to realize the fusion of prior information.The results of the WM-811 K data set verify the effectiveness of the method,indicating that the strategy can effectively take advantage of the prior information.