Few-shot fault diagnosis of turnout switch machine based on semi-supervised weighted prototypical network

Abstract

The turnout switch machine is a critical equipment of the signal system, which has a significant influence on the safety of train. However, it is difficult to obtain a mass of labeled fault data in real scenes, resulting in low diagnostic precision and poor generalization performance of the fault identification model for the switch machine. Given the above-mentioned issues, a semi-supervised weighted prototypical network (SSWPN) is proposed for fault classification of switch machines in the absence of labeled fault samples. Firstly, an effective dual-scale neural network (DSNN) is proposed to strengthen the ability of extracting switch machine fault features and expressing different scales. Furthermore, a new semi-supervised weighted prototype updating strategy is proposed, which utilizes unlabeled data to fine-tune the original prototype to enhance fault diagnosis performance. Finally, the few-shot fault data from the switch machine are validated under two conversion processes. The results demonstrate that the proposed SSWPN has good robustness and generalization. Compared with eight methods, SSWPN has obvious advantages in the actual scenario of data scarcity, which can present a theoretical basis for a few-shot fault diagnosis of the switch machine.