A Comparative Hybrid Deep Learning Approach for Robust Stator–Rotor Fault Identification in DFIGs Using FFT‐Based Features and CNN–LSTM–Transformer Models Under Nonstationary Wind Turbulence

Abstract

Reliable fault diagnosis in doubly fed induction generators (DFIGs) is essential for the stable and efficient operation of wind energy systems under nonstationary conditions caused by wind turbulence. This paper proposes a hybrid deep learning framework that combines FFT‐based frequency‐domain feature extraction with a CNN–LSTM–Transformer architecture for detecting and estimating the severity of electrical faults in DFIGs. The study focuses on stator inter‐turn short‐circuit faults and rotor winding open‐circuit fault. A comprehensive dataset was generated using MATLAB/Simulink under diverse operating conditions, including multiple fault severities, realistic load profiles, and single and simultaneous fault scenarios. An overlapping sliding window technique was employed to capture transient fault signatures and enhance sensitivity to early‐stage faults. The CNN component extracts discriminative spectral features, the LSTM models temporal dependencies, and the Transformer attention mechanism highlights fault‐relevant patterns while reducing the influence of load variations. The proposed framework was evaluated against several deep learning models, including CNN, LSTM, CNN–LSTM, Temporal Convolutional Network (TCN), Transformer, and CNN–Transformer architectures using standard performance metrics. Validation results obtained from independent test datasets indicate that the proposed CNN–LSTM–Transformer model achieves the best performance, with lower prediction errors (MSE ≈ 0.017 ± 0.001) and higher coefficients of determination (R2 up to 0.992 ± 0.001), consistently outperforming the comparative models across all diagnostic scenarios. These findings indicate that the proposed framework is a robust and effective solution for early fault detection and predictive maintenance in DFIG‐based wind turbine systems.