Inception module and deep residual shrinkage network-based arc fault detection method for pantograph-catenary systems

When a train is running at a high speed, due to the of-line condition between the pantograph and the catenary, a pantograph catenary arc (PACA) can seriously endanger the normal operation of the train, and the current collection quality of pantograph catenary system. To achieve stable current collection, PACAs are studied in this paper. Under the condition of the same contact current, five groups of experiments with different contact pressures and sliding speeds were carried out, and current signals under the normal state and the fault state were obtained. The collected current signal is converted into a picture through the Markov transition field (MTF) to construct a picture data set. To fully mine the feature information in image samples, a PACA recognition algorithm based on Inception_DRSN is proposed. First, the features of the input image sample are extracted through an inception module with multi-scale parallel convolution operation. Then convolution results of different sizes are concatenated and imported into the deep residual shrinkage network (DRSN) with an attention mechanism (AM) and a soft thresholding (ST) to perform deeper extraction and to highlight the current features under the PACA. Finally, the recognition results are classified by the global pooling (GAP) and full connection (FC) layer. Experimental results show that the average accuracy of this method is 98.50%, which is better than that of the other comparison methods. In addition, this method has robustness and superiority in the field of PACA recognition.