Few-shot fault diagnosis of rail transit switch machine based on sparse convolution capsule relation network

The switch machine is the foundational rail transit signal equipment, which has an essential impact on the safety and reliability of rail transit. Aiming at the challenge of inadequate diagnostic precision and flexibility caused by the scarcity of fault samples in the real scene of switch machines, a few-shot fault diagnosis method based on a sparse convolutional capsule relation network is proposed. First, a one-dimensional sparse convolution capsule network is proposed to approximate the sparse combination of different fault features, reduce the interference of the outside world to the original signal, and strengthen the feature extraction ability of the action current. Then, the relationship metric module is used to measure and classify these feature vectors. Finally, the proposed approach is demonstrated by different small sample experiments under two conversion states. The testing results demonstrate that compared with the other five approaches, the proposed approach achieves the highest diagnostic accuracy, reaching 96.98% and 96.05% in the two conversion states, which meets the practical application requirements of the rail transit switch machine.