AMCW-DFFNSA: An interpretable deep feature fusion network for noise-robust machinery fault diagnosis

Deep learning is applicable in mechanical fault diagnosis, ensuring the secure operation of mechanical systems. However, the lack of interpretability and noise robustness in deep learning methods has been a common challenge faced by academia and industry. An interpretable deep feature fusion network, referred to as AMCW-DFFNSA, is proposed to address these issues. First, the network integrates the discrete wavelet transform, which extends the feature learning space to the domain of wavelets, allowing for better learning of distinguishable fault features in the wavelet domain. Second, a cosine-enhanced channel attention mechanism is proposed to learn and highlight valuable interpretable features while filtering out irrelevant information. Then, a deep feature fusion network based on self-attention (DFFNSA) is proposed, which explores a closer relationship between self-attention and convolution to achieve a more profound fusion of global and local features. Extensive experiments have been meticulously conducted on two test bench datasets and a real wind turbine gearbox dataset. The results demonstrate that the proposed method is superior to the comparative fault diagnosis models regarding interpretability and noise robustness.