Lightweight Federated Domain Generalization With Global-Local Contrastive Learning for Machine Fault Diagnosis

Published: 01 Jan 2025, Last Modified: 10 Oct 2025IEEE Internet Things J. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Federated learning (FL) is a promising paradigm for industrial fault diagnosis with distributed data in Internet of Things. Despite notable progress, existing FL-based diagnostic methods still face several critical challenges. First, these methods require frequent communication between clients and servers by transmitting model parameters, leading to high computational and communication costs. Second, condition monitoring data from different clients are often collected under varying operating conditions, resulting in data heterogeneity across clients that severely degrades the diagnostic performance of FL approaches. To address these challenges, we propose a global–local contrastive learning-assisted lightweight federated domain generalization network for machine fault diagnosis. The proposed approach significantly reduces computational and communication overhead through lightweight model design and prototype-based interaction strategy. To mitigate the adverse effects of data heterogeneity, we design a local prototype contrastive learning module that effectively eliminates cross-client feature distribution discrepancies and learns client-invariant features. To generate robust global prototypes during model aggregation, a global prototype self-learning module is designed, which enables self-correction of global optimal prototypes and significantly enhances the generalization capability of the global model for unseen target clients. Experimental results on three cases demonstrate the competitive performance of the proposed method, highlighting its potential for real-world fault diagnosis applications with limited resources.
Loading