A Novel Adaptive Correlation Spatio-Temporal-Frequency Graph Convolutional Framework for Multi-Source Information Fusion-Based Fault Diagnosis of Wind Turbines

Wind-turbine fault diagnosis is limited by incomplete fault information in single-sensor signals and the poor adaptability of fixed graph structures to non-stationary operating responses. This study proposes an adaptive correlation spatio-temporal-frequency graph convolutional network (AC-STFGCN) to address these limitations in multi-source wind-turbine fault diagnosis. One branch uses one-dimensional convolutions to extract local time-domain features, while the other learns time–frequency representations based on the short-time Fourier transform (STFT). Vibration and acoustic features from these branches are fused and encoded using a bidirectional long short-term memory (Bi-LSTM) network. Local temporal segments are then represented as individual graph nodes. Feature correlations between nodes determine an adaptive, sample-specific adjacency matrix that captures non-Euclidean dependencies across temporal segments and multi-domain features. PoolGAT combines hierarchical graph pooling and graph attention to emphasize fault-relevant nodes and suppress redundant information. Experiments used multi-source signals collected from a wind-turbine drivetrain fault-simulation test rig. AC-STFGCN achieved 99.08% accuracy, 99.12% precision, 99.03% recall, and an F1-score of 99.07%. Confusion matrices, receiver operating characteristic (ROC) curves, and t-distributed stochastic neighbor embedding (t-SNE) visualizations also showed improved discrimination between similar fault patterns. Further experiments at load levels of 0%, 25%, 50%, 75%, and 100% assessed the sensitivity of AC-STFGCN to load-dependent signal variations. All four diagnostic metrics remained above 98% across the evaluated load range.