Acoustic-based method for identifying surface damage to wind turbine blades by using a convolutional neural network

Damage to the surface of the blades of a large wind turbine may lead to catastrophic blade failure. Although numerous methods have been proposed for detecting surface damage to wind turbine blades, many of them involve laboratory tests because of the difficulty in acquiring data from a commercial wind farm. This lack of data variety is an obstacle to the development of machine learning approaches for identifying the aforementioned damage. Therefore, we developed a damage detection method for wind turbine blade surfaces that is based on the physical correlation between surface conditions and acoustic signals of operating wind turbines under realistic environmental conditions. In the preprocessing stage of the aforementioned method, the short-time Fourier transform and smoothing techniques are used to analyze real-time spectrograms and the rotor speed. The derived spectrogram and rotor speed are then input into a convolutional neural network (CNN) to classify the wind turbine blade surfaces into two classes: turbines with at least one or no damaged blade. The CNN proposed in this paper is a hybrid network containing a masking module and residual classifier. The masking module suppresses redundant information in the spectrogram, and the residual classifier quantifies the difference between the masked spectrogram and a standard spectrogram. The proposed CNN can be easily trained on a small dataset with a few trainable parameters by using the physical characteristics in the residual classifier. The proposed damage detection method was evaluated using the operational noise of commercial wind turbines; according to the results, this method outperformed approaches proposed in previous studies and exhibited an accuracy of 97.11%.