Robust Phasor Measurement Unit (PMU)-based fault diagnosis in transmission lines using a spatio-temporal hybrid transformer with graph contrastive learning

The increasing deployment of Phasor Measurement Units (PMUs) in real-time power system monitoring demands accurate, fast, and reliable fault diagnosis frameworks. This paper presents a novel Spatio-Temporal Hybrid Transformer with Graph-Embedded Contrastive Learning and Bayesian Uncertainty, comprising five modules optimized for robust transmission line fault diagnosis. A Multi-Scale Temporal Transformer (MSTT) processes PMU time-series data using parallel dilated self-attention to capture multi-resolution temporal dependencies, yielding unified temporal embeddings. A Graph Signal Embedding Unit (GSEU) integrates the spatial grid topology using a graph convolutional network applied to adjacency matrices derived from electrical distances, generating topology-aware spatial feature embeddings. The Dual-Space Contrastive Learning Module (DSCLM) applies domain transform augmentations, such as Fast Fourier Transform (FFT) and wavelet transforms, to generate contrastive pairs whose alignment through contrastive loss enhances robustness to domain shifts. A Bayesian Decision Fusion Network provides calibrated class probabilities and uncertainty-aware decision-making by fusing outputs from Decision Tree, Random Forest, and K-Nearest Neighbor classifiers through Bayesian posterior weight estimation. An Adversarial Fault Simulator (AFS) employs Generative Adversarial Networks (GANs) to generate adversarial PMU samples, which stress-test the model and validate its robustness against perturbations. Evaluated on benchmark transmission line fault datasets, the proposed method achieved 99.3% accuracy, outperforming the base ensemble by 2.2% and reducing fault detection latency from 0.4977 s to 0.3215 s. Adversarial testing further demonstrated a 17.4% accuracy increase on disturbed inputs. These results collectively highlight improved spatio-temporal feature learning, enhanced generalization to noisy conditions, reliable uncertainty estimation, and robust adversarial resilience.