Image credit: Unsplash
Abstract
Learning-based mechanical fault diagnosis (FD) methods have been widely investigated in recent years. To overcome the shortages of centralized learning techniques from the perspective of data privacy and high communication overhead, federated learning (FL) is emerging as a promising method for FD. However, a large number of labeled fault data is required for the FL technique, which is not accessible in real-world industrial Internet-of-Things (IIoT) scenarios. To address the data scarcity challenge (i.e., few-shot), we propose a collaborative learning method that incorporates meta-learning into the federated learning framework. Specifically, our approach learns an effectively global meta-learner, which can quickly adapt to a new machine or a newly encountered fault category with just a few labeled examples and training iterations. Further, we theoretically analyze the convergence of the proposed algorithm in a non-convex setting. We conduct an extensive empirical evaluation of two real-world fault diagnosis datasets and they demonstrate that our proposed method achieves significantly faster convergence and higher accuracy, compared with the existing approaches.