Projet de fin d'étude : Noise extraction from wind turbines using deep learning methodes

Etudiant : RKHISS ABDELAZYZ

Filière : Master Informatique Décisionnelle et Vision Intelligente (MIDVI)

Encadrant : Pr. SABRI ABDELOUAHED

Annèe : 2025

Résumé : our report on acoustic separation of wind turbine noise using neural networks highlights several important observations. We found that the use of LSTM architectures is relevant for processing temporal data, taking into account the temporal dependency between data sequences. Our results show that the LSTM model succeeds in predicting with an MAE of 2 dB between actual and predicted SNR, but presents difficulties as the values become positive, showing some saturation with a larger error towards negative SNR. Regarding the separation of particular noise (BP) and residual noise (BR), our model was unable to generate a veritable extraction of the particular noise, with random variations in his predictions, whereas the real noise was stable. Overall, our results demonstrate the advantages and limitations of using MLPs and LSTMs to extract acoustic noise from wind turbines. Further improvements are needed to overcome the problems of saturation and error with over-fitting. These results open the way to future research aimed at improving the performance of acoustic separation models for better characterization and reduction of wind noise. In order to improve the results, we applied normalization with the z-score, which led to an improvement in the performance of previous models. In addition, we experimented with different sequence sizes for the LSTM model to optimize its ability to capture temporal dependencies. ==> It is interesting to note that other hypotheses have been considered, such as changing the scale of the data by using p² units instead of dB, or the idea of switching from a regression task to classification. However, these ideas have not yet been tested. ==> On the other hand, I have tested the particular noise (LBP) as a target variable, rather than SNR, with good results so far, using just a MLP. This shows the importance of exploring different approaches and target variables to improve model performance.