The accelerated expansion of Industrial Internet of Things (IIoT) systems has concurrently precipitated the generation of substantial multivariate time series data, where the implementation of precise anomaly detection mechanisms is indispensable for ensuring operational safety, and mitigating risks in complex industrial ecosystems. Current methodologies predominantly leverage local spatial and temporal representations derived from adjacent nodes and recent time points. This approach, which focuses on local processing, frequently overlooks global topological relationships and temporal patterns—both crucial for precise anomaly detection. This limitation arises from insufficient modeling of global sensor-time dependencies and inadequate integration of spatial-temporal interdependencies, resulting in high false-positive rates. To mitigate these issues, we introduce FL-STAM, a federated learning-enhanced framework with a spatio-temporal attention mechanism for unsupervised MTS anomaly detection. First, a parallel graph attention architecture independently extracts global sensor dependencies and temporal patterns through serial-oriented and time-oriented modules. Second, a dual-branch transformer with Wasserstein distance quantification explicitly models spatial-temporal association discrepancies, amplifying discriminative features between normal and anomalous patterns. Third, this paper adopts a privacy-preserving federated learning paradigm, which can realize collaborative model training among distributed devices while effectively preventing the exposure of local data. Extensive experiments on five IIoT datasets (SMAP, MSL, SMD, PSM, SWaT) demonstrate FL-STAM’s superiority, achieving state-of-the-art (SOTA) F1 scores of 98.52% on PSM and 97.86% on SWaT. Ablation studies verify component effectiveness, notably the parallel graph mechanism enhancing accuracy by 6.86% over baselines.