Anomaly detection in multi-sensor networks via dynamic graph structure and deep feature fusion

Multisensor networks are widely used for monitoring in industrial IoT and critical infrastructure. In complex spatiotemporal and dynamic environments they face two main challenges. First, fusing high-dimensional multimodal data is difficult. Second, frequent sensor-topology changes hinder dynamic graph modeling. Traditional static-graph methods cannot capture complex dynamic interactions, and deep fusion models are computationally heavy and often lack interpretable anomaly explanations. To address these issues, we propose the DGM-PIIoT framework. It incrementally updates the sensor graph with a time-aware GNN and uses a multimodal attention mechanism to jointly model spatial, temporal, and topological features of heterogeneous sensors. Experiments on an industrial pump-station dataset show significant improvement in anomaly detection performance (F1-score 92.70%) over existing methods.