Deep Residual Shrinkage Networks with Self-Adaptive Slope Thresholding for Fault Diagnosis

In recent years, vibration signals have been applied in mechanical device fault diagnosis, however, vibration signals are submerged by a large number of background noises in practice, which reduces the fault diagnosis accuracy. In this paper, we present a combination unit of self-adaptive slope and soft thresholding in the Deep Residual Shrinkage Networks (DRSNs), the new unit enables the DRSNs effectively learn the useful information out of the threshold region rather than completely reserving them. Furthermore, we use the attention mechanism to automatically infer the adaptive slope. Many experimental results demonstrate that the improved DRSNs can obtain more superior performances compared with the original DRSNs under background noise.