Few shot fault diagnosis method of rolling bearing based on CSML

Currently, neural network models have seen widespread adoption in the sphere of fault diagnosis. In the training process of the model there are still problems such as high demand for training data and insufficient generalisation of the model under different working conditions. In order to solve the above problems, in this paper, an algorithm named Cosine Annealing-Sgd-Meta-Learning (CSML) is proposed, which combines the cosine annealing method with the principles of stochastic gradient descent meta-learning. The method uses the cosine annealing algorithm to adjust the learning rate in the outer loop so as to better help the meta-learning task to find the initialisation parameters. In the inner loop, the learning rate is updated using the SGD optimisation algorithm and a wide convolutional network is used to improve the model's ability to extract fault information from the meta-task under known operating conditions, and then the basic learner is updated. High diagnostic accuracy was obtained on the Western Reserve University bearing dataset, which verified its feasibility and effectiveness.