Federated multi-scale contrastive learning for semi-supervised anomaly detection in multimodal industrial Internet of Things

Industrial Internet of Things (IIoT) systems generate data from heterogeneous sources across distributed cyber–physical environments, where anomalous behavior may appear as deviations within individual data streams or as inconsistencies among correlated modalities. Existing multimodal anomaly detection approaches remain limited by their dependence on centralized data aggregation, labeled attack samples, or loosely coupled modality-specific learning, thereby restricting their ability to model cross-modal inconsistencies in label-scarce, non-IID, and modality-partitioned IIoT deployments. To address these limitations, this paper proposes a federated multi-scale contrastive learning framework for semi-supervised multimodal anomaly detection. The proposed method integrates modality-specific adversarial autoencoders with federated optimization to learn compact, prior-regularized latent representations of normal system behavior without sharing raw data. Server-side multi-scale cross-modal alignment combines bidirectional InfoNCE with timestamp-valid negatives and a hardest-negative relative-margin objective to refine synchronized modality-specific representations across multiple temporal contexts. At inference, modality-specific reconstruction deviations are combined with learned cross-modal compatibility to construct a system-level anomaly score, with robust normalization and detection thresholds determined exclusively from normal validation data. Experiments on the SWaT and TON_IoT datasets show that the proposed framework achieves F1-scores of 0.9449 ± 0.0064 and 0.9819 ± 0.0033, with corresponding false-positive rates of 0.0119 ± 0.0035 and 0.0385 ± 0.0005, respectively. Additional evaluations under increased client counts, partial participation, imperfect temporal correspondence, missing observations, and alternative architectural configurations further demonstrate stable performance and support the contribution of multi-scale cross-modal alignment and compatibility-aware inference. The results demonstrate the effectiveness of the proposed data-locality-preserving framework for semi-supervised anomaly detection in heterogeneous multimodal IIoT environments.