Multi D-norm Driven Algorithm Unfolding Network: A Rolling Bearing Fault Feature Extraction Method

The deconvolution method (DM) is an effective tool for enhancing the impulsive features of rolling bearings. Deep network-based deconvolution methods transform complex numerical computations into network optimization, improving the performance of DMs. However, current deep deconvolution methods do not consider the possibility of utilizing nonlinear operations to further enhance the performance. Inspired by convolutional sparse learning and algorithm unfolding, a bearing fault feature extraction method, Multi D-norm driven algorithm unfolding network (MDN-AUN) is proposed. First, a convolutional sparse coding model is established with the fault impulse as the target sparse code. Then the Iterative Soft Thresholding Algorithm (ISTA) for solving the model is unfolded from the iterative direction and the algorithm unfolding network is generated. Finally, the multi D-norm, which can evaluate the periodicity and impulsiveness of the signal, is chosen as the objective function of the sparse code to train the neural network. MDN-AUN is compared with several deconvolution methods through simulation and real vibration signals. The results show that MDN-AUN has the superior ability to enhance the impulsive features, and the fault diagnosis performance is significantly better than the compared methods.