Wind Turbine Gearbox Fault Diagnosis Based on PNCC-STFT Voiceprint Feature Extraction and GDC- ResNet

Due to harsh operating environments and wear caused by long-term use, wind turbine gearboxes often suffer from various faults, which lead to the emission of different noises or abnormal sounds during operation. This seriously affects the normal operation of wind turbines and their power generation efficiency. Given the limitations of traditional diagnostic methods in engineering practice, this paper proposes a fault diagnosis method for wind turbine gearboxes, which integrates the Residual Neural Network with a Global Dynamic Channel Coordinate Attention Module (GDC-ResNet) and the PNCC-STFT voiceprint feature extraction combining Power Normalized Cepstral Coefficients (PNCC) and Short-Time Fourier Transform (STFT). This method is specifically developed to address the issues of feature extraction from gear noise signals and intelligent diagnosis in noisy environments. First, noise data of wind turbine gearboxes under six different states were collected using an LMS noise acquisition instrument and noise sensors. Subsequently, the PNCC-STFT voiceprint feature extraction method was employed to extract voiceprint spectrograms from the gearbox noise signals. Given that the ResNet baseline model exhibits low classification accuracy for gearbox voiceprints, a high-performance gear fault diagnosis model, GDC-ResNet, is proposed by incorporating an attention mechanism and dynamic convolutional kernels into ResNet. Experimental results show that the GDC-ResNet model achieves a classification accuracy of 99.3% for the PNCC-STFT voiceprints of gearboxes, representing an improvement of 1.8% compared with the ResNet baseline model. The proposed classification model can effectively classify gearboxes with variable rotational speeds under different states in the dataset, enabling more efficient diagnosis of faults in wind turbine gearboxes.