Research on multi-path quadratic convolutional neural network-based bearing fault diagnosis

In real-world complex situations, high levels of noise from the surroundings and other component resonances frequently distort collected vibration signals, giving the collected data non-linear features. This research presents a multi-path quadratic convolutional neural network (MPQCNN) for bearing fault diagnosis in response to the issue of the low generalisation performance of traditional deep learning-based bearing fault diagnosis methods and their limited diagnostic capabilities in noisy situations. The proposed MPQCNN combines an attention mechanism and a residual structure, utilising the potent feature representation capability of quadratic neurons to process the input in noisy situations. By using dilated convolutions with different dilation rates, the receptive field of the MPQCNN is expanded and the multi-scale features obtained are fused to enhance the fault diagnosis capability. Moreover, a dynamic balance adaptive threshold residual block is used to enhance the robustness of the model. To perform pertinent experiments, the MPQCNN uses bearing datasets from the Southeast University and Case Western Reserve University (CWRU). The results show that the suggested approach has strong noise immunity. The diagnostic accuracy of the MPQCNN for the CWRU and Southeast University bearing datasets can reach up to 100% when the signal-to-noise ratio (SNR) is 6.