Research on bearing fault classification method based on one-dimensional dense neural network embedded with MWRN

Abstract Processing noisy data in mechanical systems is a crucial issue. Currently, the construction of discrete wavelet networks is limited to creating a few wavelet network layers, overlooking the connection with the overall network. The selection of mother wavelet functions and decomposition levels makes the length of post-transformation features unknown, hindering the construction of deep networks. Moreover, some studies ignore the inherent nature of the wavelet network as a network structure.To address these problems, this paper begins with the discrete wavelet formula, reveals the operating mechanism of the one-dimensional discrete wavelet network, and discovers it as a specific type of convolutional neural network. Then, it establishes a formula for the wavelet network feature length and proposes a multi-scale discrete wavelet residual network block (MWRN). Next, using the one-dimensional dense attention neural network (DenseNet) as the backbone and improving SKAttention as the transition block, a deep discrete wavelet dense network structure (DDenseNet) is constructed.Experiments show that when trained on public datasets, DDenseNet has higher accuracy and stability than networks like WMA-CNN. Its classification accuracy exceeds that of popular models at SNR of −4, −2, 0, and 4. In the noisy bearing fault experiment of rolling mills, when the mother wavelet is db36 and the decomposition level is 1, the accuracy on the noise-free dataset exceeds 97%. When the mother wavelet is coif3 and the decomposition level is 1, the accuracy on datasets with SNR of −4, −2, and 4 is about 92%, demonstrating excellent performance.