Self-learning guided residual shrinkage network for intelligent fault diagnosis of planetary gearbox

Abstract

The original vibration signals of the fault gear under different working conditions have a large distribution difference, and there will be insufficient feature extraction during fault diagnosis, which leads to the problem of low diagnostic accuracy. Therefore, a self-learning model based on residual shrinkage network (SLRSN) is proposed. The model constructs a deep residual shrinkage network as the main network for feature extraction of the original vibration signal to enhance the robustness of the model. Then self-believing loss and self-doubting loss are proposed to achieve self-confidence and suspicion of health status prediction. The first is self-confidence loss, which adopts sub-domain distribution adaptation to actively align learned cross-domain features. The second is self-doubt loss, which provides SLRSN with the ability to extricate from wrong experience. Finally, to mitigate the effects of negative transfer, a novel adaptative weight allocation mechanism is designed to recalibrate the weighting of each source domain sample. Through the experiment of two gearboxes, it is verified that the proposed SLRSN method has good diagnostic reliability under the condition of gear speed and load change.