Wind turbine blade damage detection based on acoustic signals

Area of Science:

• Acoustics and Signal Processing
• Mechanical Engineering and Structural Health Monitoring
• Artificial Intelligence and Machine Learning

Background:

• Increasing wind turbine blade size necessitates advanced structural health monitoring techniques.
• Operational noise from wind turbines contains valuable information for assessing blade integrity.
• Traditional spectral subtraction methods face challenges with extreme noise levels in turbine sound signals.

Purpose of the Study:

• To investigate the efficacy of a pretrained sound source separation neural network combined with spectral subtraction for wind turbine sound signal denoising.
• To compare the time-frequency representations of signals processed by spectral subtraction alone versus the combined source separation and spectral subtraction approach.
• To evaluate the performance of a ResNet50 deep residual neural network for damage detection using features extracted from denoised signals.

Main Methods:

• Acquisition of wind turbine sound signals encompassing wind, pneumatic, and mechanical noise.
• Application of a pretrained sound source separation neural network to isolate mechanical noise from background wind noise.
• Processing of signals using both traditional spectral subtraction and a combined source separation-spectral subtraction method, followed by Short-Time Fourier Transform (STFT) analysis.
• Feature extraction using Mel-scale Frequency Cepstral Coefficients (MFCCs) to create training and testing datasets for normal and abnormal conditions.
• Damage detection using the ResNet50 deep residual neural network model.

Main Results:

• The combined source separation and spectral subtraction approach yielded more detailed time-frequency diagrams compared to spectral subtraction alone.
• The proposed denoising method effectively reduced noise in wind turbine sound signals, enhancing the identification of anomalous sounds from blade damage.
• The ResNet50 model achieved high accuracy in damage detection, with 95% confidence intervals for training set accuracies ranging from 0.926 to 0.965 and for test set accuracies from 0.869 to 0.931.

Conclusions:

• The integration of a sound source separation neural network with spectral subtraction is a highly effective strategy for denoising wind turbine operational sound signals.
• This advanced denoising technique significantly improves the capability to detect structural anomalies in wind turbine blades through acoustic monitoring.
• The study demonstrates the potential of AI-driven acoustic analysis for reliable and efficient wind turbine structural health monitoring.