Ahmed, Baseer and Wang, Jia ORCID: https://orcid.org/0000-0003-4379-9724 (2026) Spectrogram-based CNN feature optimisation for acoustic detection of wind turbine blade damage. In: Intelligent Methods, Systems, and Applications (IMSA). Institute of Electrical and Electronics Engineers (IEEE), Piscataway, New Jersey, pp. 401-406. ISBN 979-8331584887 (doi:10.1109/IMSA70415.2026.11700328)

Preview

PDF (Author's Accepted Manuscript) 54613 WANG_Spectrogram-Based_CNN_Feature_Optimisation_For_Acoustic_Detection_(IEEE AAM)_2026.pdf - Accepted Version Available under License Creative Commons Attribution. Download (395kB) | Preview

Abstract

Wind turbine blade damage, if left undetected, poses significant risks to operational safety, maintenance costs, and environmental sustainability. Inspired by the human ability to identify mechanical abnormalities through sound, this paper presents a Convolutional Neural Network (CNN)-based acoustic fault detection system capable of classifying the structural health of wind turbine blades from audio data. Real turbine recordings collected from wind farms in Çanakkale, Turkey and Edinburgh, United Kingdom were used as the basis for constructing a synthetic yet acoustically representative dataset of 100 recordings. These recordings were generated using generative AI (ChatGPT 5.0). Two spectrogram-based feature extraction pipelines were developed and compared: one based on the Short-Time Fourier Transform (STFT) and one based on Mel-frequency spectrograms. To address initial overfitting, hyperparameter optimization was applied using Dispersive Flies Optimization (DFO) for the STFT pipeline and the Optuna framework for the Mel-spectrogram pipeline. K-Fold cross-validation was employed throughout to improve model generalization given the relatively limited data size. Both optimized models demonstrated consistent and stable performance on anomaly detection. The STFT model outperformed the Mel-spectrogram model by achieving a mean cross-validation accuracy of 75% and a false negative rate of only 4% in damage detection, compared with 60% and 28% respectively for the Mel-spectrogram model. The findings support the viability of spectrogram-based deep learning for non-intrusive wind turbine condition monitoring, with potential for extension to other rotating machinery including engines, pumps, rail systems, and aircraft components.

Item Type:	Conference Proceedings
Title of Proceedings:	Intelligent Methods, Systems, and Applications (IMSA)
Additional Information:	2026 Intelligent Methods, Systems, and Applications (IMSA). DOI: 10.1109/IMSA70415.2026. 11th - 12th July 2026. Giza, Egypt.
Uncontrolled Keywords:	convolutional neural networks, wind turbines, acoustic signal processing, spectrograms, anomaly detection, short-time fourier transform, renewable energy
Subjects:	Q Science > Q Science (General) Q Science > QA Mathematics > QA75 Electronic computers. Computer science T Technology > T Technology (General)
Faculty / School / Research Centre / Research Group:	Faculty of Engineering & Science Faculty of Engineering & Science > School of Computing & Mathematical Sciences (CMS)
Last Modified:	05 Oct 2026 09:56
URI:	https://gala.gre.ac.uk/id/eprint/54613

Actions (login required)

View Item

Downloads