Anamoly Detection For Railway Vehicle Transmission Systems And Its Application To Rolling Bearings

There always are a large number of samples obtained during condition monitoring on healthy railway vehicles. Contrarily, samples from fault components of railway vehicle are too difficult to acquire. Therefore, anomaly detection can not be performed in a traditional way without fault samples. It is an urgent and tough problem that improving the security level of railway vehicle by detecting anomaly performance while lacking of fault samples. Feature extraction and detection modeling are the two core techniques in anomaly detection for railway vehicle transmission systems. The sensitive features which are able to efficiently characterize equipment operating state can be discovered by feature extraction. On the other hand, a reliable detection method can only be invented with valid detection modeling method based on sensitive features.For assisting in anomaly detection of railway vehicle transmission components and focusing on the above core techniques, this thesis developed research on anomaly detection method based on local mean decomposition (LMD) and support vector domain description (SVDD). Main work and innovations of the thesis are as follows.(1) To solve the over-sifting and under-sifting problems that emerge in sifting procedure, adaptive sifting stopping criterion for LMD has been proposed. The adaptive sifting stopping criterion, which is based on product functions (PFs) definition, takes local estimation error into account, and constructs the objective function of sifting times. Optimal sifting times can be computed by minimizing this objective function. Experiment results indicated that the adaptive sifting stopping criterion restrains over-sifting and under-sifting, finally helps to improve the accuracy and efficiency of LMD, because it avoids subjective threshold settings, which is common in many sifting criteria.(2) To solve the false component interfusion problem in PFs, a recursive ranking method for false component recognition has been proposed. In the recursive ranking method, comprehensive K-L divergence is computed to quantify the similarity between each PF and original signal and its components, and false component can be finally recognized by this index. Experiment results indicated that the recursive ranking method evaluates the similarity between PFs and components of original signal in a more comprehensive way, instead of comparing each PF only with original signal, and higher accuracy of false component recognition can be achieved.(3) To solve the over-fitting and under-fitting problems that emerge in SVDD modeling, a parameter selection method based on kernel matrix statistics has been proposed. Under the analysis of relationship that underlies fitting extent versus dispersion of kernel matrix elements, an objective function for kernel parameter based on kernel matrix statistics has been constructed, which is able to return the optimal parameter value for the Gaussian kernel SVDD. The kernel statistics based parameter selection method bypasses the training stage, and thus calculation efficiency is improved greatly. Experiment results indicated that over-fitting and under-fitting problems are restrained to a large extent by this method.(4) For rolling bearing, experiment has been designed with references to practical railway vehicle working condition. With the experimental data as input, LMD equipped with adaptive sifting criterion and recursive ranking false component recognition method have been performed to realize feature extraction. Then, SVDD detection model was constructed with parameter selected by the kernel matrix statistics method. Experiment results showed that the anomaly state, which contains inner race fault, elements fault and outer race fault of rolling bearing can be detected accurately by the anomaly detection method proposed by this thesis.