Spatio-temporal multi-fault and severity aware diagnosis in WSN via GNN transformers.

Wireless Sensor Networks (WSNs) play a perilous role in the IoT environment and other applications, where reliable sensing and timely fault diagnosis are essential for system stability. However, fault behavior in WSNs is inherently complex because of the coexistence of multiple fault types, their spatial propagation across the network topology, temporal evolution over time, and varying levels of severity. Existing fault diagnosis approaches largely focus on binary or single-fault detection and fail to jointly model network topology, long-range temporal dependencies, class imbalance, and fault severity estimation. To address these limitations, this study proposes a GNN-Transformer, a dual-stream Graph Neural Network-Transformer framework for multi-fault classification and severity-aware fault diagnosis in WSNs. The proposed model integrates a GNN stream to capture spatial correlations and fault propagation among neighbouring sensor nodes, and a Transformer-based temporal stream to model long-term temporal dependencies in multivariate temperature-humidity sensor data. To mitigate the severe class imbalance intrinsic in real-world WSN datasets, Conditional Tabular Generative Adversarial Network (CTGAN) is employed to generate realistic minority fault samples, which are further enhanced using SMOTE and Borderline SMOTE techniques. A graph-temporal fusion module with attention-based multi-scale aggregation adaptively combines spatial and temporal representations for robust fault characterization. The proposed GNN-Transformer achieves 86.79% accuracy and an F1-score of 87.1% for multi-fault classification, along with a 62.4% F1-score for fault severity estimation. These results confirm the effectiveness of the proposed topology-aware, severity-conscious framework for reliable and scalable fault diagnosis (FD) in WSNs.