Acoustic monitoring of wind turbine blades using wavelet packet analysis and 1D convolutional neural networks
Fangfang Liu,Wenxian Yang,Kexiang Wei,Andrew Ball,Robert Cattley,Bo Qin
DOI: https://doi.org/10.1177/14759217241293006
2024-11-21
Structural Health Monitoring
Abstract:Structural Health Monitoring, Ahead of Print. Wind turbine blades (WTBs) are susceptible to faults in the harsh wind farm environments, making their safety a matter of paramount importance. Unfortunately, existing composite blade monitoring methods face various limitations in practical use. To address this issue, the study presents an intelligent fault detection method to assess the health of both the structural integrity and its skin. This has never been tried before by scholars. The study begins with the collection of acoustic signals from the blade chamber. Second, these signals are processed using wavelet packet decomposition (WPD) and Fast Fourier Transform to generate two-dimensional feature matrices. Third, apply the obtained matrices to train a one-dimensional convolutional neural network (CNN), enabling advanced feature extraction and classification, which forms the basis of the WPD-CNN model. Finally, the proposed method was experimentally verified. It has been found that the proposed WPD-CNN model achieves fault detection accuracies ranging from 87.14% to 98.57%, depending on the types of feature matrices used for training the model. These results highlight the model's strong performance in diagnosing WTBs. Additionally, the study emphasizes the advantage of using frequency spectrum-derived features over traditional time-domain waveform features for effective WTB fault detection.
engineering, multidisciplinary,instruments & instrumentation