Time-Frequency Attention Mechanism Based Precise Noise Extraction and Quantitative Analysis Technique in Complex Noise Environment

With the increasingly severe noise pollution, the high-quality intelligent noise origin-tracing surveillance becomes increasingly important for urban environmental monitoring and governance. Due to the complexity of urban acoustic environments and the variety of noise types, the excessive noise levels monitored often comprise multiple components, complicating the origintracing surveillance analysis. To facilitate noise origin-tracing surveillance (NOTS) analysis and effective governance of excessive urban noises, a precise target noise extraction (TNE) technique was proposed, aiming to accurately quantify the contribution of target noise type. To this end, a generative adversarial training model combined with two-dimensional time-frequency attention (2D-TFA) was developed. And the 2D-TFA module was utilized to assign attention to each time-frequency point, guiding the model to focus on high-energy regions and thereby improving TNE. Furthermore, to meet the quantitative noise analysis requirements, the standard A-weighted sound pressure level (awSPL) will be used as the metric to assess the TNE performance. Experiments conducted on a practical noise dataset demonstrate that, incorporating generative adversarial architecture with the proposed 2D-TFA mechanism can lead to a high-performance TNE model. The evaluation results were presented to indicate that the proposed techniques outperform the baseline methods in terms of the average awSPL error of the extracted target noise, which confirms us that the proposed framework can effectively support precise noise extraction and quantitative analysis in complex noise environments.