Transfer learning for the classification and characterization of locally collected wind farm acoustical noise

Analyzing noise from wind farms presents challenges, especially in distinguishing turbine-generated sounds from background environmental noise. Traditional techniques, such as the low-frequency peak prominence method recommended by the Institute of Acoustics, have been used to detect amplitude modulation—a sound pattern linked to increased annoyance. More recent methods use machine learning to extract detailed acoustic features from neural networks trained on sound data, such as VGGish and YAMNet, which improve the ability to classify wind farm noise. This study builds on previous work by applying a multi-stage machine learning framework that uses features generated by YAMNet. In the first stage, the system identifies portions of recordings where wind turbine noise is dominant and filters out segments with unrelated environmental sounds. In the second stage, machine learning techniques are used to detect amplitude modulation, allowing further classification of the sounds based on tonal content, modulation patterns, and other characteristics important to human perception and regulatory guidelines. Earlier studies relied on publicly available sound datasets. This work expands on those efforts by training and testing the model using noise recordings collected directly from local wind farm sites. These localized data improve the model’s accuracy and make it more useful for real-world noise monitoring and community impact evaluation.