The acoustic impact of French wind farms is currently estimated by measuring their sound emergence. These measures require the implementation of on/off cycles of the wind farm in order to determine the ambient noise (wind turbines in operation) and the residual noise (stopped wind turbines). These procedures generate very high costs for operators, which strongly limit the duration of emergence measurement periods (2 or 3 weeks). This reduced duration, compared to a full year of different

more »

... conditions, is to the detriment of the representativeness of the estimation of sound emergence. In order to remedy this limitation, we propose to estimate the noise emergence of wind turbines in real time, continuously and without stopping the machines, using a source separation method based on a machine learning technique: Non-negative Matrix Factorization. This technique is tested on a corpus of simulated sound scenes that allows a total control of their composition and especially their emergence. A numerical experiment is conducted to determine, among the various influential parameters of this method, the optimal form that achieves the best estimates of sound emergence over the entire sound corpus. Initial results indicate that this approach generates average estimation errors similar to current methods but depends on the emergence of wind turbine noise. This method makes it possible, under validation by more complex corpora, to estimate the noise emergence of wind farms continuously without having to shut them down which is not the case in the current method. Keyword: wind turbine noise, sound emergence, in situ measurements, non-negative matrix factorization 10.48465/fa.2020.0107