Online diagnosis for bogie bearing unbalance impact faults of railway trains with wavelet phase space reconstruction mechanism

The reliability of bearings directly affects the operational status of mechanical equipment, and there is a serious imbalance in its shock vibration characteristics. This paper constructed wavelet phase space in the offline diagnosis stage to locate the part of the fault signal that was concentrated in energy to reconstruct and generate pseudo-fault samples, and guides the category balancing weights in the two-phase feature learning network to reduce the loss contribution of the easy-to-classify samples, while the adaptive threshold is adjusted to balance the local noise with the global statistics, so that the model focuses on the difficult samples more. Finally, small batch incremental learning was applied to online diagnosis to improve the efficiency of industrial diagnosis. Thirty-six sets of experiments under multiple unbalance cases show that the reconstructed signals are of high quality, and the G-mean value can reach 95.04%, which provides a new idea for the online diagnosis.