Session: 09-17-01: AI for Energy

Paper Number: 167413

Transfer Learning for the Classification and Characterization of Wind Farm Acoustical Emissions 

Analyzing acoustical data from wind farms presents challenges related to distinguishing turbine-generated sounds from environmental noise and classifying wind turbine noises based on its acoustic characteristics. While machine learning (ML) techniques for general environmental sound classification have been developed using extensive human-labeled databases (e.g., YAMNet), few studies have specifically addressed wind farm noise (WFN) classification. Existing approaches primarily focus on identifying amplitude modulation (AM) using both traditional methods, such as low-frequency peak prominence (IOA method), and ML-based techniques that leverage targeted AM acoustic features and deep acoustic representations.

To overcome these challenges, we are developing a multi-echelon ML framework for WFN identification and classification using publicly available wind farm datasets and open-source software. The first echelon automates the identification of WFN samples that are free of environmental noise. The second echelon employs ML algorithms to classify these wind farm noise samples based on AM characteristics, prominent tones, and other acoustic features relevant to human perception and regulatory compliance metrics.

Using a publicly available dataset of 6,000 ten-second WFN samples (https://open.flinders.edu.au/articles/dataset/Benchmark_wind_farm_noise_data_set/19618023/1), we extracted samples with minimal environmental noise using YAMNet analysis, spectrogram inspection, and auditory validation. For the samples identified as being free of disturbances, we computed 31 AM-related parameters derived from a machine learning approach to AM detection (https://github.com/ducphucnguyen/WFN_AM_Detection). These parameters were then used for cluster analysis grouped the samples into acoustically distinct categories, such as broadband noise with minimal AM, broadband noise with prominent AM (“swish”), and low-frequency-dominant noise with pulsating tones. Differences in these clusters in terms of prominence and modulation depth, as quantified using the initial steps of the IOA AM computation method (https://sourceforge.net/projects/ioa-am-code/) were compared with results from a small psychoacoustic pilot study. In this study, we conducted listening tests on representative noise samples from each cluster. Young adults with normal hearing and no prior listening training participated in the study, conducted in an audiometric test suite. Participants rated noise characteristics and annoyance levels using a graphical user interface. Our findings provide insights into human perception of wind farm noise variations and contribute to the development of AI-driven methodologies for improved noise classification and regulatory assessment.

Results from this preliminary work is then expanded to wind farm acoustics data collected from local wind farms under a wide range of distances and environmental conditions. Efficacy of transfer learning using the pretrained models developed with the original publicly available dataset, is presented along with strategies for improving the results.

This research advances AI-driven acoustic analysis by enhancing automated WFN classification and providing a foundation for refining noise assessment frameworks in wind energy applications.

Presenting Author: Heather Lai Suny New Paltz

Presenting Author Biography: Heather Lai is an Associate Professor of Mechanical Engineering at SUNY New Paltz in New Paltz, NY

Research interests include wind farm acoustics, room acoustics, psychoacoustics, modeling of dynamic behavior of viscoelastic materials, and properties of 3D printed flexible materials.

Teaching in the undergraduate mechanical engineering program includes courses such as dynamics, system dynamics and associated labs, mechanical measurements and FEA.

Authors:

Heather Lai Suny New Paltz
Chih-Yang Tsai SUNY New Paltz