WSAFormer-DFFN: A model for rotating machinery fault diagnosis using 1D window-based multi-head self-attention and deep feature fusion network

Abstract

Fault diagnosis is of great importance for rotating machinery maintenance. Deep learning is an intelligent diagnosis technology that attracts more attention at present. The ability to learn fault features of pure CNN and RNN models is limited due to inherent structural defects. Therefore, we build a fault diagnosis model (WSAFormer-DFFN) combining CNN and self-attention structure to enhance learning ability. Meanwhile, propose 1D window-based multi-head self-attention (1D W-MSA) for vibration signal and its representation to introduce local inductive bias, which improves the robustness of the model. A series of experiments on three rotating machinery datasets show that the proposed model has better recognition accuracy and less performance attenuation in interference conditions. On the bearing dataset of Case Western Reserve University, the accuracy rate increases by 21.29%–37.12% compared with several classic models when the signal-to-noise ratio is −6 dB.