A novel wind turbine fault diagnosis method based on RTH algorithm and lightweight ShuffleNet with dense structure

Addressing the challenges of extracting fault features from non-stationary and noisy vibration signals in wind turbines, this paper proposes a novel wind turbine fault diagnosis method based on the Red-Tailed Hawk (RTH) algorithm and a lightweight ShuffleNet with a dense structure, which can realize rapid and accurate fault identification for wind turbine bearing during its operation and maintenance. The proposed method first employs RTH-optimized Variational Mode Decomposition (VMD) to adaptively decompose the original signal, decoupling fault features from irrelevant components. Subsequently, a weighted Intrinsic Mode Function (IMF) evaluation metric is introduced to filter out irrelevant feature components and reconstruct the signal. Bispectrum transformation is then applied to capture nonlinear phase coupling, further enhancing fault characterization. Finally, a lightweight Dense-ShuffleNet (DS) model is utilized to significantly reduce runtime without compromising accuracy. After denoising the signal using VMD optimized by the RTH algorithm, the model accuracy improved from approximately 90% to around 98.0%. In comparison with alternative data transformation approaches, the Bispectrum method enabled the model to attain the optimal classification performance. Through analysis and validation via t-SNE visualization and Friedman test, the proposed method exhibited significant advantages over other models in terms of both runtime efficiency and prediction accuracy. Experimental results under high-noise conditions and on additional datasets further demonstrated the generality and robustness of the method presented in this study.