Monitoring and maintaining the health of wind turbine blades has long been one of the challenges facing the global wind energy industry. Detecting damage to a wind turbine blade is important for planning blade repair, avoiding aggravated blade damage, and extending the sustainability of blade operation. This paper firstly introduces the existing wind turbine blade detection methods and reviews the research progress and trends of monitoring of wind turbine composite blades based on acoustic signals. Compared with other blade damage detection technologies, acoustic emission (AE) signal detection technology has the advantage of time lead. It presents the potential to detect leaf damage by detecting the presence of cracks and growth failures and can also be used to determine the location of leaf damage sources. The detection technology based on the blade aerodynamic noise signal has the potential of blade damage detection, as well as the advantages of convenient sensor installation and real-time and remote signal acquisition. Therefore, this paper focuses on the review and analysis of wind power blade structural integrity detection and damage source location technology based on acoustic signals, as well as the automatic detection and classification method of wind power blade failure mechanisms combined with machine learning algorithm. In addition to providing a reference for understanding wind power health detection methods based on AE signals and aerodynamic noise signals, this paper also points out the development trend and prospects of blade damage detection technology. It has important reference value for the practical application of non-destructive, remote, and real-time monitoring of wind power blades.
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.
Wind turbines operate under harsh conditions, heightening the risk of rotating bearing failures. While fault diagnosis using acoustic or vibration signals is feasible, single-modal methods are highly vulnerable to environmental noise and system uncertainty, reducing diagnostic accuracy. Existing multi-modal approaches also struggle with noise interference and lack causal feature exploration, limiting fusion performance and generalization. To address these issues, this paper proposes CAVF-Net—a novel framework integrating bidirectional cross-attention (BCA) and causal inference (CI). It enhances Mel-Frequency Cepstral Coefficients (MFCCs) of acoustic and short-time Fourier transform (STFT) features of vibration via BCA and employs CI to derive adaptive fusion weights, effectively preserving causal relationships and achieving robust cross-modal integration. The fused features are classified for fault diagnosis under real-world conditions. Experiments show that CAVF-Net attains 99.2% accuracy with few iterations on clean data and maintains 95.42% accuracy in high-entropy multi-noise environments—outperforming single-model acoustic and vibration by 16.32% and 8.86%, respectively, while significantly reducing information uncertainty in downstream classification.
Wind turbines operating in harsh environments have a significantly higher risk of rotating bearing equipment failures. Although fault diagnosis can be achieved by collecting and analyzing acoustic signals from the equipment, various background noises in the environment inevitably interfere with the accuracy of equipment detection. Delayed maintenance may lead to further damage and increased maintenance costs. To address this issue, this work proposes a WGU-Net (Wave-GRU-U-Net) signal denoising network to resolve the background noise problem in acoustic signal acquisition for the rotating bearings of wind turbines under simulated realistic working conditions and consequently enable their fault diagnosis. The network is an improvement over the U-shaped structure of the U-Net: it inserts a gated recurrent unit network between the encoder (downsampling and cropping) and the decoder (upsampling and concatenation) to better preserve temporal dependencies in the signal, enabling reconstruction and denoising. Then, the denoised signals are fed into preset SVM and CNN networks to perform fault diagnosis for rotating bearings of wind turbines under simulated realistic working conditions. The results demonstrate that under multi-noise conditions, the signals denoised by WGU-Net exhibited high recognition accuracy, achieving averages of 97.98% with the SVM classifier and 99.3% with the CNN classifier. These results represent improvements of 9.17% and 6.9%, respectively, compared to the accuracy before denoising.