GMA-DRSNs: A novel fault diagnosis method with global multi-attention deep residual shrinkage networks

The large environmental noise interference has a negative impact on the fault diagnosis of vibration signals. To solve the problems, we present novel global multi-attention deep residual shrinkage networks (GMA-DRSNs), by using attention mechanism. In this paper, the self-adaptive Leaky Thresholding shrinkage function is firstly proposed to substitute the original soft thresholding function in the deep residual shrinkage networks (DRSNs), where all the inner parameters of the approach are automatically inferred based on the attention sub-networks. Secondly, a novel activation function is further presented based on the above improvement, in order to realize the corresponding adaptive nonlinear transformation of each signal. Various experimental results show that our work can achieve better performance compared with the previous works. Finally, we systematically analyze the threshold's tendency, and surprisingly find the same consistency with the receptive field of convolutional neural networks, which is the first geometry explanation work about DRSNs' structure.