Traffic Noise Annoyance Assessment Based On Deep Learnin

With the development of urban road traffic,road noise pollution has increasingly become a public concern,controlling and reducing the harm caused by traffic noise pollution has always been a hot spot in traffic noise control research.Traffic noise subjective annoyance has become one of the important bases for road traffic pollution control,the assessment of traffic noise annoyance usually has subjective experimental methods and objective prediction methods: subjective experimental methods usually use social investigation or listening test in the laboratory to directly obtain subjective annoyance,high reliability,but often require a lot of manpower and material resources;The objective prediction method is simple and easy to predict the subjective annoyance by extracting acoustic features and mapping the model,but the reliability is usually inferior to the subjective experimental method.In view of the disadvantages of the above noise annoyance evaluation method,combined with the above two methods,this study constructs the mapping relationship between road traffic noise and subjects’ perceived annoyance based on artificial intelligence deep learning model,so as to quickly and accurately evaluate the perceptual annoyance caused by road noise,and it is proposed to carry out research in the following three aspects:Firstly,this paper uses acoustic equipment to collect road noise and road noise is collected in the morning,afternoon and evening respectively.It also uses extra databases Sound Ideal General6000 to enrich road noise data,such as horns and motorcycle noise,to meet actual needs.Secondly,the recorded noise data was cut into 8-second noise fragments and replayed through headphones HD600 in the listening room.The ISO 11 level evaluation scale was adopted to rate the perceived annoyance to obtain the evaluation level of the perceived annoyance.949 datas are obtained,furthermore,this paper obtain the characteristics of noise fragments and analyze the relationship between subjective annoyance and noise fragments.Finally,an Unet based traffic noise annoyance evaluation model is constructed according to the data format.The mapping relationship between noise segments and subjective annoyance is established.Furthermore,the performance of the algorithm is optimized by using transfer learning.The algorithm is pre-trained by constructing a psychoacoustic annoyance dataset of traffic noise and then secondary optimization is carried out on the listening experiment dataset to further improve the robustness of the algorithm.