A novel semi-supervised prototype network with two-stream wavelet scattering convolutional encoder for TBM main bearing few-shot fault diagnosis

Accurately sensing the main bearing state and diagnosing fault types is crucial to ensure the safe operation of the main drive system of tunnel boring machines. Currently, the research on large-scale bearing fault diagnosis in industrial scenarios is severely limited by the quality and quantity of monitoring data. Conventional external vibration monitoring devices are difficult to adapt to complex and harsh working conditions of excavation equipment, and constantly changing low-speed and heavy-load operating conditions make similar labeled samples very scarce. To tackle this concern, we propose a semi-supervised prototype network with the two-stream wavelet scattering convolutional encoder (TWSCE-SSPN) based on roller state signals. By fusing radial and axial features of rollers using the two-stream structure and employing wavelet scattering transform and attention mechanism in the convolutional feature encoder, the model exhibits excellent feature mapping capabilities. Following the semi-supervised meta-learning paradigm, the proposed model uses the prototype generated by unlabeled sample features to modify the initial prototype generated by labeled sample features to augment the accuracy of classification in few-shot learning. The integrated sensing roller main bearing testbed was set up and fault datasets were established to verify the few-shot classification and anti-noise ability of the algorithm. Experimental results show that TWSCE-SSPN achieved 98.17 % accuracy at 1 shot, which is at least 18.17 % higher than existing methods. Furthermore, even under a signal-to-noise ratio of 0 dB, the few-shot recognition accuracy can remain 91.83 %. This verifies the superiority of the model in diagnosing main bearing faults under few-shot and strong noise conditions.