Samenvatting

This research explored the use of deep learning to classify impact events on wind turbine blades using spectrogram images. Audio data were recorded through indoor and outdoor simulations using smartphones. The data preparation process involved converting stereo recordings to mono, applying downcasting to reduce file size, augmenting the data with environmental sounds, removing noise with spectral gating, and filtering signals using a 100 to 500 Hz band-pass filter. These audio signals were then converted into grayscale spectrograms with fixed dimensions to produce consistent inputs suitable for model training. Two models were evaluated: a standard convolutional neural network(CNN) and a hybrid model named CARNN that combined convolutional layers, attention mechanisms, and recurrent neural networks. Both models were trained on indoor data and tested using outdoor recordings to assess generalization. Each model was evaluated across five independent test trials. On average, the CARNN model achieved 95.18% precision, 95.32% specificity, and an inference time of 0.0841 seconds per sample. The CNN model recorded 93.42% precision, 93.70% specificity, and a slightly faster inference time of 0.0803 seconds. Overall, CARNN outperformed CNN in both classification accuracy and reliability. These outcomes emphasize the importance of structured data preparation and model design in improving classification performance. Although still in an early stage, this project offers a technical basis for further work on automated impact detection to support wind turbine maintenance.