A Novel Fault Diagnosis Method Based on Feature Fusion and Model Agnostic Meta-Learning

There are two limitations in the existing researches based on data-driven fault diagnosis: 1) the diversity of the original signal features is ignored; 2) the number of fault samples is limited, which usually leads to failure of data-driven fault diagnosis methods. To overcome it, this paper proposes a novel fault diagnosis method based on feature fusion and model agnostic meta learning. First, the proposed method generates the new signal features by fusing time-frequency features, working condition features, and time difference features based on the original signals to solve the first limitation. Second, the training strategy of model agnostic meta learning is optimized to solve the second limitation. Moreover, the deep learning model in the inner loop structure, ReLu activation function and cross-entropy loss function in model agnostic meta learning were applied to improve the performance of the diagnostic model. Ultimately, the proposed method was demonstrated on the dataset CWRU and PT100. The experimental results show that the proposed method has higher diagnostic accuracy and generalization ability by adding the diversity of the original signal features under the meta-training task.