Application of audio spectrogram transformer machine learning model for audio tagging of construction activities

Abstract

Major construction projects are approved based on an Environmental Impact Statement, which includes modeled predictions of noise impacts based on planned program. However, actual on site construction activities can differ significantly from planned works, resulting in modeled acoustic impacts (which have been used to mitigate impacts and inform stakeholders) out of date. A potential solution to this may be the use of machine learning models, to initially classify, and later predict, actual on site activities and commensurate impacts on nearby stakeholder and communities caused by site works. By leveraging emerging lower cost, smaller noise monitoring devices more data can be collected at receivers, and classified to determine the contributing sources of sound. SiteHive has worked with the University of Technology Sydney to design and develop a machine learning model to classify construction works in real-time on site, integrated as part of the SiteHive Hexanode multi-sensor environmental monitoring device. This presentation will showcase the design and development undertaken to date, and highlight results as tested as part of a major works program.