LGMA-DRSN: a lightweight convex global multi-attention deep residual shrinkage network for fault diagnosis
Zhijin Zhang, Chunlei Zhang, Lei Chen, He Li, Ping Han
Abstract
Abstract Recently, the fault diagnosis domain has witnessed a surge in the popularity of the deep residual shrinkage network (DRSN) due to its robust denoising capabilities. In our previous research, an enhanced version of DRSN named global multi-attention DRSN (GMA-DRSN) is introduced to augment the feature extraction proficiency of DRSN specifically for noised vibration signals. However, the utilization of multiple attention structures in GMA-DRSN leads to an escalation in the computational complexity of the network, which may pose practical deployment challenges. To address this limitation, this paper proposes a lightweight variant of GMA-DRSN, referred to as lightweight convex global multi-attention deep residual shrinkage network (LGMA-DRSN), building upon our prior work. Firstly, the numerical variation regularity of the adaptive inferred slope parameters in the global parametric rectifier linear unit is analyzed, where we surprisingly find that a convex parameter combination always occurs in pairs. Based on this convex regularity, the sub-network structure of the adaptive inferred slope with attention mechanism is optimized, which greatly reduces the computational complexity compared to our previous work. Finally, the experimental outcomes demonstrate that LGMA-DRSN not only enhances diagnostic efficiency, but also ensures a high level of diagnostic accuracy in the presence of noise interference, when compared with our prior work.