Fault diagnosis of ZDJ7 railway point machine based on improved DCNN and SVDD classification

Problems such as poor noise immunity and overfitting are prone to occur when convolutional neural network (CNN) is exploited in the fault diagnosis of ZDJ7 railway point machine. In addition, some fault features are unbalanced and have the features of multiple tags, which lead to low diagnosis accuracy. Therefore, an improved deep convolutional neural network (DCNN) and support vector data description (SVDD) classification is proposed. First, the depthwise separable convolution in the Xception structure is used to optimize the extraction of fault features. Second, the adaptive batch normalization processing (AdaBN) is performed to improve the noise immunity. Meanwhile, the global average pooling layer (GAP) is used instead of the fully connected layer to improve the generalization ability of the network. Aiming at the unbalanced features of the railway point machine sample, an improved quantity learning algorithm for hypersphere coordinate mapping based on SVDD is proposed. The classification is realized under unbalanced samples. The experiment shows that the accuracy based on the improved DCNN and SVDD is 96.59%. It has a good anti-noise performance under different convolution kernels and SNRs. When the sample distribution is unbalanced, the performance indexes obtained by the proposed model are the best.