CFs-focused intelligent diagnosis scheme via alternative kernels networks with soft squeeze-and-excitation attention for fast-precise fault detection under slow & sharp speed variations

The evolution of deep learning-based intelligent fault diagnosis is mainly confronted with challenges on the analysis of complex non-stationary signals and the design of strong robust models. The causal relationship between the two promotes the development of better models, among them, convolution frameworks (CFs) are regarded as one of the most versatile structures. In standard CFs, the receptive fields of the kernels in each layer are designed to share the same scale, which easily leads to model performance degradation in non-stationary data analysis. Consequently, we propose a dynamically selective mechanism in CFs that allows every kernel to adaptively adjust its receptive field by multi-scale information, which is named as alternative kernels networks (AkNets). Combining with specially designed squeeze-and-excitation (SE) attention, the AkNets utilize information-guided soft attention to fuse multiple branches with different kernel scales, which generates different effective receptive fields of kernels in fusion layer. Five bearing vibration data collected under slow & sharp speed variations verify the effectiveness of proposed approach. The results indicate that the AkNets greatly improve the efficiency on the premise of high recognition accuracy. Moreover, the extended application of the AkNets’ unit can assist various state-of-art CFs-based models improve the recognition accuracy by 3%–12%.