Sensor-Driven Robust Fault Detection and Diagnosis for Wind Turbines Under Dynamic and Uncertain Environments Based on Digital Twin-Enhanced Fusion Learning

Abstract

Reliable fault detection and diagnosis in wind turbine systems remains a significant challenge due to the highly dynamic and uncertain nature of real-world operating environments. Variations in wind speed, turbulence intensity, mechanical loading, and sensor noise can substantially degrade the performance of conventional data-driven diagnostic methods. This article presents a robust fault detection and diagnosis framework for wind turbines operating under dynamic and uncertain conditions. All results reported in this article are obtained through simulation studies conducted in a MATLAB/Simulink environment; no physical wind turbine test rig was used. The only externally measured information employed is a set of wind-speed statistics derived from supervisory control and data acquisition (SCADA) records of large-scale wind turbines, used solely to initialize the simulated wind profile. Unlike many purely simulation-based studies, the proposed framework explicitly distinguishes among, first, MATLAB/Simulink-generated simulation data, second, SCADA-informed wind profile initialization, and, third, training-only data augmentation, thereby ensuring transparent data provenance and facilitating reproducibility. Dynamic operating conditions are represented through fluctuating wind speeds, turbulence effects, and load-dependent variations. Model parameters are selected according to physically meaningful operating ranges reported in the wind energy literature, thereby enhancing simulation realism and reducing arbitrary parameter tuning. To improve robustness against uncertainty, a digital twin-based data augmentation strategy and small measurement perturbations are incorporated during training. A hybrid feature extraction scheme combining time-domain, frequency-domain, and statistical features is employed to capture diverse fault signatures. The proposed framework integrates multiple classifiers through early-, late-, and decision-level fusion strategies with attention-based weighting mechanisms. Simulation results obtained from multiclass wind turbine fault scenarios demonstrate that the proposed framework achieves superior diagnostic accuracy and robustness compared with conventional approaches. Nevertheless, the near-perfect classification performance is interpreted cautiously, as controlled simulation conditions and strong feature separability may partially contribute to the observed results. To mitigate potential data leakage, strict scenario-level data partitioning is enforced throughout the evaluation process.