Fault Diagnosis for Gearbox of Wind Turbine Based on Transfer Learning and Improved Res2Net

The gearbox of wind turbine, as a critical component, plays an essential role in ensuring the stable operation of wind power systems through effective fault diagnosis. The limited availability of labeled vibration data from wind turbine gearboxes, making it difficult for conventional fault diagnosis methods to effectively address this challenge. In this paper, a fault diagnosis method is proposed for gearboxes of wind turbine based on transfer learning and an improved Res2Net50 network. First, continuous wavelet transform (CWT) and Gramian angular field (GAF) are applied to extract features from vibration signals. Subsequently, transfer learning is employed by adapting a pre-trained model from the Southeast University gearbox dataset to the China Huadian Electric Power Research Institute dataset, followed by fine-tuning for accurate fault classification across different wind turbine systems. Additionally, this paper proposes a multi-branch, multi-scale residual structure based on Res2Net to effectively classify wind turbine gearbox faults. Experimental results demonstrate that the proposed method achieves high accuracy and robustness on the CHDER gearbox dataset, successfully addressing the challenge of classification accuracy in small sample datasets.