Federated Meta-Learning for Few-Shot Fault Diagnosis with Representation Encoding

• DOI: 10.48550/arxiv.2310.09002
• Publication Date: 2023-01-01
• Authors: Jixuan Cui; Jun Li; Zhen Mei; Kang Wei; Wei E. I. Sha; Ming Ding; Wen Chen; Song Guo
• Keywords: Initialization; Representation; Feature Learning

Deep learning-based fault diagnosis (FD) approaches require a large amount of training data, which are difficult to obtain since they located across different entities. Federated learning (FL) enables multiple clients collaboratively train shared model with data privacy guaranteed. However, the domain discrepancy and scarcity problems among deteriorate performance global FL model. To tackle these issues, we propose novel framework called representation encoding-based federated meta-learning (REFML) for few-shot FD. First, strategy based on encoding is developed. It harnesses inherent heterogeneity clients, effectively transforming it into an advantage out-of-distribution generalization unseen working conditions or equipment types. Additionally, adaptive interpolation method that calculates optimal combination local models as initialization proposed. This helps further utilize information mitigate negative effects discrepancy. As result, high diagnostic accuracy can be achieved types limited data. Compared state-of-the-art methods, such FedProx, proposed REFML achieves increase in by 2.17%-6.50% when tested same type 13.44%-18.33% totally types, respectively.