Deep learning-based active noise control on construction sites

Abstract

Although construction noise pollution has always been a severe issue for governments in metropolitan cities, there is no effective and practical solution to control it. Due to the high nonlinear and transient nature of machinery noises on construction sites, traditional active noise control (ANC) algorithms show a marginal ability to mitigate them. In this paper, a novel high-performance deep-learning-based feedforward ANC controller is proposed to attenuate construction-related noise by considering the delay and nonlinear behavior of acoustic devices. The developed network, with around 128,500 parameters, can be expanded to a multi-channel ANC method without increasing computational costs, and it is suitable for ANC in open space environments like construction sites. Broadband noise attenuation of around 8.3 dB was achieved of a wide variety of construction noises with minor degradation at very high-frequency ranges (7.5–8 kHz). The presented network outperformed traditional and state-of-the-art ANC algorithms.