Task-Agnostic Generalized Meta-learning Based on MAML for Few-Shot Bearing Fault Diagnosis

Meta-learning methods have been widely applied to solve few-shot problems. However, the metamodels of current meta-learning methods may be too biased toward the tasks in the meta-training phase and are less adaptable to new tasks, especially when the number of new tasks is small. To reduce the bias of the metamodel and improve its generalizability, this paper proposes a Task-Agnostic Generalized Meta-Learning (TAGML) algorithm based on Model-Agnostic Meta-Learning (MAML) for few-shot bearing fault diagnosis. The algorithm improves MAML in terms of both network structure and optimization algorithm. Firstly, the quality of feature extraction is improved by adding a squeeze-and-excitation attention module to the network of MAML. Secondly, the following improvements are made in the optimization algorithm: (1) The stability of the training process is improved by using multi-step loss optimization in the optimization; (2) It is proposed to add the Task-Agnositic regular penalty term to the meta-optimization objective function to improve the task unbiasedness of the metamodel; (3) To speed up the convergence and further improve the model's ability to generalize to different tasks, an iterative updatable outer loop learning rate strategy is used. Experiments demonstrate that the algorithm is not only effective in identifying new fault tasks that do not appear in the meta-training phase but also has good recognition performance for generalized bearing fault scenarios with a mixture of seen and unseen class fault tasks.