Continuous Fault Diagnosis in Resource‐Constrained IIoT: A Value‐Driven Asynchronous Federated Framework

The deployment of Federated Learning (FL) in the Industrial Internet of Things (IIoT) encounters a triple threat of systemic challenges: physical resource heterogeneity, long‐tailed communication latency, and catastrophic forgetting within streaming task environments. Existing resource‐efficient paradigms predominantly rely on static sub‐network partitioning that struggles to accommodate real‐time hardware fluctuations, whereas asynchronous protocols often impose indiscriminate staleness penalties that lead to the erroneous filtering of high‐utility gradients. Furthermore, generative federated continual learning schemes induce prohibitive downlink communication overhead. To address these limitations, this article proposes Federated Elastic Asynchronous Continual learning (FedEAC), a system‐level framework specifically designed for constrained industrial edge nodes. FedEAC integrates a dynamic structured masking mechanism that formulates sub‐networks on‐the‐fly according to real‐time physical constraints, alongside a value‐compensated semi‐asynchronous protocol that establishes a mathematical balance between temporal staleness and gradient utility. Critically, the framework offloads the computationally intensive orthogonal projection process to the central server, enabling historical knowledge retention with zero additional storage overhead at the edge. Extensive system‐level evaluations across three industrial datasets demonstrate that FedEAC achieves a peak accuracy of 85.94% and a minimum Root Mean Square Error (RMSE) of 18.28, while empirically characterizing the Pareto frontier between gradient freshness and system latency. Under extreme network conditions with an 80% straggler ratio, FedEAC realizes a 1.45 speedup compared to robust baselines and yields a 41.1% reduction in cumulative communication traffic.