Abstract:

In order to address problems of low accuracy, slow speed and few damaged samples in diagnosis of train wheelset tread damage, a diagnosis model based on EPSA-ResNet was proposed. Firstly, model parameters were obtained by pre-training ImageNet data set through transfer learning, and transferred to wheelset tread damage characteristics for fine-tuning, so that shared model structure and parameters were obtained. Secondly, 3×3 convolution in ResNet-50 residual blocks was replaced by a Pyramid Split Attention (PSA) Block before a new EPSA-ResNet was gained, which integrated multi-level features of spatial and channel attention, and feature re-calibration was carried out adaptively. Finally, damage diagnosis results were obtained by Softmax classifier. The results show that this method can identify damage state of train wheelset tread and local subtle differences between surrounding environment, with an diagnosis accuracy up to 99.79%, which is better than other deep neural network models.

Key words: efficient pyramid split attention-residual network (EPSA-ResNet), wheelset tread, damage diagnosis, transfer learning, residual network