Complex fault diagnosis in wind turbine bearings: a hybrid approach combining the improved feature mode decomposition and convolutional neural networks

The complex noise interference and diverse fault-induced signals in vibration data from wind turbine equipment pose significant challenges for bearing fault diagnosis, including cumbersome methodologies, prolonged processing times, and compromised accuracy. To address these limitations, this study proposes a novel composite fault diagnosis framework that integrates Feature Mode Decomposition (FMD), Fast Spectral Kurtosis (FSK), and Convolutional Neural Network (CNN). While conventional Empirical Mode Decomposition (EMD) exhibits limited noise robustness and struggles to extract subtle fault signatures in composite failure scenarios, our approach employs FMD to decompose fault-related intrinsic mode functions (IMFs)and further filters the IMF components using fast spectral cliffs with enhanced feature separability. Subsequently, the Short-Time Fourier Transform (STFT) is applied to derive time-frequency representations, followed by Fast Spectral Kurtosis analysis to identify optimal demodulation bands for non-stationary signals. The energy spectrum of denoised signals is converted into grayscale images, serving as input to a tailored CNN architecture for hierarchical feature learning. Experimental validation demonstrates that this hybrid methodology achieves a fault recognition accuracy of 98 % under compound fault conditions, outperforming conventional EMD-based approaches in terms of noise immunity and diagnostic precision. Comparative analysis reveals an 8 % improvement in detection reliability over standalone deep learning models, particularly in low signal-to-noise ratio (SNR) environments. The proposed framework offers a robust solution for multi-fault identification in industrial Bearing machinery, demonstrating superior generalization capability across varying operational conditions.