Abstract
<title>Abstract</title> <p>Wireless Sensor Networks (WSNs) in hostile or far away environments will often be exposed to more than one type of faults at once; Stuck-At, Random Noise, Gain/Offset, etc., and this is usually due to the degradation of both data quality and network efficiency. Distributed fault detection algorithms are usually dependent upon either comparing local thresholds, or using some form of statistical testing, and these approaches have difficulty capturing spatial relationships among neighbouring nodes as well as the effects of interacting multiple faults. Graph Neural Networks (GNNs) can usefully model the dependency relationships within graphs. This approach utilizes the first layer to encode each node's spatially informed neighbourhood features (weighted sums and/or anomalies) using a graph convolutional neural net (GCN). The second layer uses the learned features to classify the node(s) as faulty or normal via an adaptive belief rule base system. The proposed framework evaluated a synthetic benchmark WSN. The BRB-GNN, an interpretable framework for fault diagnosis in wireless sensor networks, uniquely supporting spatial modelling, rule-based interpretability, and multi-fault detection simultaneously. Evaluated across ten independent runs at fault probability p = 0.15, BRB-GNN achieved a Correct Detection Rate of 25.23% ± 1.07% and False Alarm Rate of 70.93% ± 8.98%. Ablation confirmed the hybrid design's value, with the full model (25.7% ± 0.9% CDR) outperforming GNN-only (24.8%) and BRB-only (25.5%) variants. Per-class analysis showed strongest detection for gain/offset faults (38.3%), while validation on the Intel Berkeley Lab dataset yielded 25% CDR with 0% FAR, confirming real-world applicability.</p>