Small Sample Fault Diagnosis Method of Point Machine Based on Improved Relation Network

The point machine is a pivotal component of the turnout system and is essential for safe conversion. Focusing on the issues of low fault diagnosis and sluggish convergence time brought on by a lack of real-world fault data, a small sample fault diagnosis approach based on an improved relation network (IRN) is proposed in the research. Firstly, a strip pooling convolutional network is proposed as a feature extractor of the relation network to deeply mine the fault feature information and strengthen the feature extraction ability of point machines vibration data. Secondly, a combined optimizer is proposed. The Sophia optimizer is used to optimize the feature extractor, and the metric module is optimized using the Adam optimizer, which accelerates the convergence speed of the model. Finally, the proposed IRN method is verified by the small sample vibration data of the point machine. The experimental results demonstrate that in contrast to the other three meta-learning methods, the accuracy of the IRN method in the 1-shot task is 95.33%, and it has good small sample diagnostic performance.