Spatio-temporal connected graph convolutional network for SCADA-data-based wind turbine fault diagnosis under time-varying operating conditions

Wind turbine fault diagnosis based on Supervisory Control and Data Acquisition (SCADA) data is challenging due to continuously varying operating conditions. Existing approaches are often limited in representing sensor interactions across multiple temporal scales and their changing spatio-temporal correlations. This study develops a spatio-temporal connected graph convolutional network (STCGCN) to address these limitations. A Gaussian-kernel-based predefined graph is first established to describe sensor dependencies across different time steps. An adaptive graph is then learned from temporal embeddings and latent representations to characterize correlations that vary with operating conditions. Information from the two graph branches is combined through a gating mechanism, followed by dilated causal convolution for long-range temporal modeling. Evaluation on two real-world SCADA datasets covering blade icing and main bearing faults shows that STCGCN can effectively represent complex spatio-temporal relationships and provides reliable diagnostic performance under time-varying conditions.