Sparse Attention Coding-Timestep Shrinking Spiking Neural Network for Gearbox Fault Diagnosis

In recent years, the application of deep learning in intelligent fault diagnosis has seen rapid growth and demonstrated remarkable success. However, most deep learning models rely on traditional artificial neural network (ANN), which lacks a certain interpretability. In contrast, spiking neural network (SNN), a biologically inspired neural network, has been taken more and more attentions since it was proposed. In view of the frequent failure of gearboxes in engineering, a sparse attention coding-timestep shrinking SNN (SAC-TSSNN) is proposed in this paper for intelligent fault diagnosis of gearboxes. The proposed SAC-TSSNN can achieve excellent diagnostic performance because of the specialized spiking coding, training strategies and spiking neurons. Specifically, in terms of spiking coding, a sparse attention coding (SAC) method is designed by incorporating sparse convolution into gated attention coding, which makes the coding information physically interpretable because it naturally fits with the sparsity character of fault signals. To speed up the model training, a timestep shrinking SNN (TSSNN) is established, which integrates timestep shrinkage layers to the SNN model to accelerate the training speed while retaining effective information. In addition, the parallel spiking neuron (PSN) is employed to construct the network model, which improves the processing efficiency of the network. A series of experiments are conducted on two gearbox datasets. The results demonstrate that the proposed SAC-TSSNN model can substantially improve the network performance as compared to the state-of-the-art SNN models with respect to diagnostic accuracy, model stability, training speed, convergence and coding sparsity.