Two-Stage Noise Origin-Source Tracking and Quantitative Analysis for Intelligent Urban Environment Monitoring

With the increasingly serious noise pollution in cities, the intelligent noise origin-tracing surveillance becomes increasingly important for urban environmental monitoring and governance. This article proposed a two-stage noise origin-source tracking and quantitative analysis framework to decompose complex urban noise into different noise components to analyze their quantitative contributions to the overall noise pollution. In the first stage, an improved Transformer and a deep encoder/decoder network were utilized to handle the problem that traditional end-to-end dual-path speech separation networks cannot directly model the speech sequence context, thus enabling better characterizing the signal features. At the same time, a new loss function was proposed to combine the extended optimal stretched optimal scale-invariant signal-to-noise ratio (SOSISNR) and time-domain waveform loss to improve the separation performance. In the second stage, a Conformer-based time-frequency domain generative adversarial networks was used to repair and suppress impurities in the separation results. The evaluation results were presented to indicate that the proposed methods outperform the baseline methods in terms of stable separation performance.