Accelerating graph-based deep learning for self-calibration in large-scale uncontrolled wireless sensor networks for environmental monitoring

Low-cost Wireless Sensor Networks (WSNs) have emerged as a valuable complement to high-performance monitoring stations for environmental monitoring. These low-cost networks significantly enhance spatial and temporal resolution, which is crucial given that many highly localised events require early warnings to mitigate their impacts. Recent advancements in machine learning (ML) have enabled self-calibration for WSNs, but existing approaches struggle to capture the spatiotemporal relationships among sensors. Graph Convolutional Networks (GCNs) have the ability to capture complex interactions of WSNs by recursively aggregating information from each node’s neighbours. However, as the network grows, this neighbourhood aggregation becomes computationally expensive, limiting scalability. This requires sampling strategies to reduce node count per aggregation step without sacrificing performance. To address computational inefficiency in GCNs, this work introduces an adaptive importance sampling (AIS) method that adjusts sampling probabilities based on node degree and node importance within the network. By prioritizing the most influential nodes during neighbourhood aggregation, this approach reduces the number of nodes sampled at each step, lowering computational complexity. To maintain layer connectivity and avoid disconnection of sampled nodes AIS-GCN samples from the set of 1-hop neighbours at each layer, ensure smooth information flow between layers. Experimental results show that AIS-GCN achieves a test accuracy of 0.9365, outperforming traditional GCN (0.9219), FastGCN (0.9288), HDSGNN(0.9317) and performing comparably to AS-GCN (0.9396 ), demonstrating better scalability without sacrificing performance. Notably, AIS-GCN reduces training time by over 69% compared to standard GCN, making it an effective solution for large-scale WSNs and resource-constrained settings.