Adaptive Multi-Channel Residual Shrinkage Networks for the Diagnosis of Multi-Fault Gearbox

Abstract

Intelligent fault diagnosis is a hot research topic in machinery and equipment health monitoring. However, most intelligent fault diagnosis models have good performance in single fault mode, but poor performance in multiple fault modes. In real industrial scenarios, the interference of noise also makes it difficult for intelligent diagnostic models to extract fault features. To solve these problems, an adaptive multi-channel residual shrinkage network (AMC-RSN) is proposed in this paper. First, a channel attention mechanism module is constructed in the residual block and a soft thresholding function is introduced for noise reduction. Then, an adaptive multi-channel network is constructed to fuse the feature information of each channel in order to extract as many features as possible. Finally, the Meta-ACON activation function is used before the fully connected layer to decide whether to activate the neurons by the model outputs. The method was implemented in gearbox fault diagnosis, and the experimental results show that AMC-RSN has better diagnostic results than other networks under various faults and strong noises.

Keywords

• deep learning
• multi-fault
• fault diagnosis
• residual network
• adaptive weights
• gearbox