A gear fault diagnosis method for wind turbines based on HDDL and ConvNeXt under limited sample conditions

Wind turbines operating under complex conditions endure variable loads. As the primary transmission component, the gearbox is highly susceptible to damage, making fault diagnosis crucial. However, neural networks require substantial data for training samples during diagnostics, yet researchers struggle to collect sufficient wind turbine gear vibration signals in real industrial settings. To address this diagnostic challenge posed by insufficient samples, this paper proposes a method combining hierarchical deep dictionary learning (HDDL) with ConvNeXt. HDDL employs a multi-layer sparse coding network to perform layer-by-layer feature extraction, enabling feature extraction from complex vibration signals while addressing the instability issues of traditional orthogonal matching pursuit (OMP) algorithms. This diagnostic method first performs sparse reconstruction of the signal via HDDL, followed by converting the reconstructed signal into a two-dimensional spectrogram using continuous wavelet transform (CWT). Given the diverse operating conditions of wind turbines, transfer learning principles are applied to the ConvNeXt model. Finally, the reconstructed two-dimensional time-frequency map is fed into the ConvNeXt model for recognition. In this study, wind turbine gear data are employed for validation. It is demonstrated that a recognition accuracy of 99.89% is achievable even with limited sample sizes, thereby effectively resolving the issue of low recognition rates caused by insufficient data.