A novel cross-domain fault diagnosis method based on model agnostic meta-learning. (August 2022)

• Record Type: Journal Article
• Authors: Yang, Tianyuan; Tang, Tang; Wang, Jingwei; Qiu, Chuanhang; Chen, Ming
• Is Part Of: Measurement. Volume 199(2022)
• Publication Date: 2022-08
• Subjects: Fault diagnosis -- Meta-learning -- Cross-domain -- MAML -- Recurrence plot; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111564
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A data processing method based on Fourier transform and recurrence plot (FT-RP) is proposed. Two-dimensional data from the transformation of the time series is used as the input feature data. The MAML is improved with a backbone network combining residual network and attention mechanism, etc. The large-margin Gaussian Mixture (L-GM) loss function replaces the fully connected layer as the classifier, resulting in the MAML with increased cross-domain diagnostic performance. The approach of meta-learning for fault diagnosis is discussed, as well as how to construct the meta-learning task on the typical public bearing dataset. Multiple sets of cross-domain meta-learning are validated, and the analysis and tests on a novel dataset with multi-working conditions confirm the proposed method's accuracy and superior generalization capabilities. Abstract: In real industrial scenarios, the working conditions of mechanical equipment are always highly variable and the amount of data that can be collected is limited, which renders a severe challenge to most existing deep learning-based intelligent fault diagnosis methods. To overcome the problem of fault diagnosis under varying working conditions and limited data, we develop a novel cross-domain diagnosis method based on model agnostic meta-learning (MAML). Firstly, a data preprocessing method based on Fourier transform and recurrence plot (FT-RP) is performed to obtain domain-independent input data. Then, the construction strategy of meta-task in cross-domain diagnosis is summarized according to the characteristics of mechanical devices with many working conditions but few fault categories. Meanwhile, the training strategy of MAML is optimized to adapt to the cross-domain problem, and the residual shrinkage network and the large-margin Gaussian Mixture (L-GM) loss are applied to improve the classification accuracy of the model. Ultimately, the effectiveness of the proposed method is demonstrated by three case studies with excellent accuracy and generalization performance for fault classification tasks under new conditions after learning prior knowledge in known conditions. It is also concluded that the performance of the model can be improved by adding data from other working conditions or other datasets to the meta-training task.