Wind turbine blade damage detection based on acoustic signals

In recent years, the size of wind turbine blades has increased, underscoring the critical importance of monitoring their structural health. This study explores the use of noise emitted during wind turbine operation for the assessment of blade structural integrity. During sound acquisition, the wind sound, pneumatic sound and mechanical sound are recorded together to form the wind turbine sound signal. Considering the computational challenges of spectral subtraction under extreme noise intensities, a pretrained sound source separation neural network was used to distinguish between random wind noise and mechanical noise in wind turbine sound signals. In this paper, the short-time Fourier transform (STFT) time-frequency diagrams of signals processed using the spectral subtraction method are compared with those processed by combining the source separation model and spectral subtraction. The results reveal that the combined approach provides a more detailed representation in the time-frequency diagrams. Additionally, the mel-scale frequency cepstral coefficients (MFCCs) algorithm is utilized for feature extraction in the experimental dataset, forming training and test sets for the normal and abnormal datasets. To carry out damage detection, the ResNet50 deep residual neural network model is employed. The training results of the same network model were evaluated using the datasets obtained from the four different denoising schemes in the experiments and a 95% confidence level assessment metric. The analysis of the 95% confidence intervals reveals that the proposed sound source separation model combined with the traditional spectral subtraction denoising algorithm is effective in reducing the noise of wind turbine sound signals and performs well in identifying the anomalous sound generated by blade damage. Under this approach, the 95% confidence intervals of the model training set accuracies range from 0.926 to 0.965, while the confidence intervals of the test set accuracies range from 0.869 to 0.931.