Research On Separation And Evaluation Of Road And Track Mixed Traffic Noise Based On Deep Learnin

With the rapid socio-economic development in China,in order to meet the increasing travel demand of urban residents,there are more and more areas where different transportation modes such as road,subway and light rail coexist in a limited urban space.Road traffic noise and rail traffic noise are different in terms of sound generation mechanism,duration and assessment methods,and the emergence of mixed noise from multiple transportation modes,thus bringing challenges to the assessment and management of environmental noise in areas where road and rail traffic coexist in cities.Therefore,in order to reduce the impact of mixed traffic noise in these areas on the quality of life of surrounding residents,it is necessary to use sound source separation techniques in these areas to obtain the pollution contribution of different noise sources,so that targeted measures can be taken for effective management.According to the above-mentioned noise assessment problem in urban traffic coexistence areas,this study uses deep learning algorithms based on artificial intelligence technology to separate mixed traffic noise in order to achieve pollution assessment of different noise sources.In this paper,the research will be carried out in two aspects as follows:First,a baseline dataset is designed for deep learning model experiments based on the traffic noise separation task.In this study,the existing two-path recurrent neural network in the speech separation task is used as the baseline model.Then,a two-branch dual-path recurrent neural network model is proposed to predict road traffic noise and rail traffic noise separately using the branch structure,and an interaction module is introduced between the branches to use the information learned from the other branch to offset the unwanted information and recover the missing information.The A-weighted mean square error is used as the objective function to separate the signals of different traffic noise sources,and the A-weighted sound pressure level is used for quantitative evaluation.Second,experiments are conducted based on narrowband traffic noise dataset to compare the performance of the proposed model and the baseline model,and spectral analysis is performed by 1/3 octave.Because of the high frequency limit of rail traffic noise,the proposed model is also extended to a wide band for experiments.Then,the effects of different vehicle speeds,vehicle types,signal-to-noise ratios and road sections on the rail traffic noise separation performance are investigated.Finally,an executable program is designed according to the traffic noise separation task to test the model separation effect by mixing traffic noise with real scenes.