A Novel Bearing Fault Diagnosis Method Based on Deep Incremental Learning

In the face of continuous dynamic changes in working conditions, the static deep learning bearing fault diagnosis model gradually loses its effectiveness. To this end, a novel bearing fault diagnosis method based on deep incremental learning is proposed in this paper. Firstly, a one-dimensional signal is converted into a two-dimensional image by Gramian angle field (GAF) to realize the extraction of global features of the input signal. The image is dimensionally transformed and input into a selectable classification network, which can improve the robustness and generalization ability of the model. Secondly, a parameter importance estimation method based on the change of loss value is designed. It can effectively alleviate the catastrophic forgetting problem in incremental learning caused by inaccurate estimation of the importance of the parameters of old working conditions. Then, the gradient clipping method using Huber regularization eliminates the gradient explosion induced by the quadratic terms in the regularization process. Finally, the validity of the method was verified with a dataset from the bearing data center of Case Western Reserve University. The results show that when dealing with incremental faults, the method can achieve diagnostic accuracies of 97.75% and 98.25% on different classification networks, respectively. And it can well overcome the stability-plasticity dilemma.