Research On The Key Issues And Applications Of Anomaly Detection Based On Support Vector Machines

In past years, with the development of sensor, communication and information processing technologies, it has been possible to monitor the industrial production process and device status information in real time. Herein, it is particularly important by the use of the monitoring data to recognize the anomaly in time, so that we could avoid the accident and reduce unnecessary casualties or economic losses. Therefore, the research on efficiently data-driven anomaly detection methods has significant theoretical and practical values.In this paper, for the demands of the coal floor water inrush prediction and conveyor belt tearing fault detection, the systematic analysis on the basic theory and structure of support vector machine(SVM) algorithm and the performance optimization of anomaly detection based on SVM are carried out. First of all, due to limited abnormal samples and the class-imbalance problem in anomaly detection, both SVDD(support vector data description) and SVM are utilized to detect the anomalies. The SVDD model is built only by using the normal data. Then, the major studies are conducted in following three aspects in terms of the enhancement of detection robustness and generalization, the improvement of detection efficiency and the optimization of detection model with online learning.(1) The coal floor water inrush dataset and the conveyor belt fault dataset are established to train and test the detection models. For the prediction of water inrush from coal floor, the main influence factors of water inrush are analyzed based on the current domestic and international research on the mechanism of inrush, and then the benchmark dataset of water inrush is established according to the current published reference. Meanwhile, to realize the water inrush prediction of a coal mine in Shanxi province, a measured dataset of water inrush is collected and labeled through the analysis of the hydrogeological data and exploration data from the coal mine. For the conveyor belt tearing fault detection, the conveyor belt monitoring images are acquired with the machine vision method, and then the median filtering is carried out. The gray histogram and gray level co-occurrence matrix are extracted as feature dataset based on the characteristic of the conveyor belt tearing images.(2) The detection accuracy and generalization performance of SVDD is highly sensitive to the noise disturbance. In this paper, two types of new SVDD methods, named R-SVDD and εNR-SVDD, are presented, which are constructed by introducing cutoff distance-based local density of each data sample and the ε-insensitive loss function with negative samples. It has demonstrated that the proposed methods can improve the robustness and generalization of SVDD for data with noise though the theoretical analysis and the extensive experiments on the UCI datasets. The experimental results have shown that the proposed εNR-SVDD is superior to other existing anomaly detection methods in terms of the detection rate and the false alarm rate. Meanwhile, the proposed R-SVDD can also achieve a better anomaly detection performance with only normal data. Finally, the proposed methods are used to detect the coal floor water inrush and the image-based conveyor belt faults.(3) Because the feature selection can improve the efficiency of the detection model, this paper presents an enhanced artificial bee colony-based support vector machine(EABC-SVM) approach to realize the effective feature selection. To improve the optimized capability of original ABC(artificial bee colony) approach, the EABC algorithm establishes two enhanced strategies including the Cat chaotic mapping initialization and current optimum based search equations. Several UCI datasets have been used to evaluate the performance of EABC-SVM and the experimental results show that this approach has better classification accuracy and convergence performance than the ABC-SVM, and other ABC variants-based SVM. Furthermore, the EABC-SVM is used to detect the coal floor water inrush and the image-based conveyor belt faults, which can achieve significant detection accuracy and reduce the amount of features.(4) For the anomaly detection cases, the monitoring data stream often changes dynamically. It is very important to identify some new critical information from the dynamic data streams. However, the traditional SVM model, which is based on the batch implementation, cannot effectively utilize these new data so that the detection accuracy would decrease. In order to effectively solve the problem, a new generalized KKT condition is defined with the sample distribution characteristics, and then the GKKT-ISVM algorithm is proposed based on the generalized KKT condition to select the optimal update dataset. Several UCI datasets have been used to evaluate the performance of GKKT-ISVM algorithm, and the experimental results show that the proposed algorithm can utilize the previous training results effectively and update the SVM model combined with the new samples quickly. Meanwhile, the proposed GKKT-ISVM algorithm is used to detect the coal floor water inrush and the image-based conveyor belt faults, which can improve the detection accuracy and meet the requirements of anomaly detection applications with the dynamic data streams.