<![CDATA[Industrial bearing health monitoring is hindered by four interrelated challenges: high class imbalance, the absence of fault-type annotations, stringent data privacy constraints prohibiting centralized aggregation, and non-independent and identically distributed (non-IID) degradation dynamics across geographically dispersed assets. To address these, we propose Fed-TGCN, a novel weakly supervised federated learning framework grounded in temporal graph neural networks. Each client represents a leave-one-bearing-out fold, comprising two training bearings, one validation bearing, and one held-out test bearing, constructs a hybrid spatio-temporal graph from six physics-informed statistical features derived from raw vibration signals; edges encode both sequential dependencies and feature-space similarity via k-nearest neighbors. Pseudo-anomaly labels are generated locally through adaptive thresholding of a degradation score using exponentially weighted moving average, eliminating reliance on expert annotations. Under a strict leave-one-bearing-out protocol on the NASA IMS dataset (12 bearings), local Temporal Graph Convolutional Networks are trained in isolation and aggregated globally via FedAvg. Our method achieves an Average Precision of 0.675 ± 0.276 and Matthews Correlation Coefficient of 0.636 ± 0.285, maintains stronger performance consistency across heterogeneous bearing conditions than isolated and non-graph baselines (ΔMCC = +0.130, p < 0.01). Ablation studies confirm the necessity of temporal modeling (MCC drops by 0.069 without GRU). To the best of our knowledge, this is the first work integrating weakly supervised, graph-based federated learning for bearing prognostics under, demonstrating that parameter coordination but not the data sharing which enables degradation-invariant representation learning across heterogeneous assets.]]>
