Research On Recognition And Separation Of Noise Signals In Subway Stations Based On Deep Learning And Blind Source Separation

China has the longest operating mileage of urban rail transit lines and the most urban rail transit stations in the world currently.Stations are an important part of urban rail transit and the noise in subway stations is of vital importance to the comfort,health and safety of passengers and staff.Therefore,it is necessary to study the composition and contribution of noise in the station by identifying and separating the noise of various sound sources.This paper was supported by the Ministry of Science and Technology of the People’s Republic of China(grant number 2017YFB1201104).This paper focuses on the study of the noise recognition algorithm based on deep learning and the noise separation algorithm based on blind source separation to identify the type of noise source in the station and separate the noise of virous sound sources.The noise types of urban rail transit stations and the contribution of various sound sources can be analyzed by inputting the noise signals into the deep learning model and the blind source separation models,which can provide a scientific basis for noise reduction measures.The main work and research results of this paper include:(1)Noise tests were carried out on the platform of subway stations.Station A on Beijing Subway Line 2,Station B on Beijing Subway Line 5,Station C on Beijing Subway Line 10 and Station D on Beijing Subway Line 10 were selected to carry out the noise test and collect acoustic signals under various conditions.The timedomain and frequency-domain characteristics of the noise at the measuring point are analyzed when the train pull into or pull out the station,the broadcast is in progress and the door opens or closes.(2)A database of noise spectrograms was established and noise recognition was carried out based on the YOLO-v3 target detection algorithm.The YOLO-v3 target detection algorithm was used to train spectrograms in the database.To effectively evaluate the performance of the network model,some indexes selected to evaluate the recognition capability of the network model.At the same time,the factors affecting the performance of the YOLO-v3 target detection algorithm were analyzed and the recognition capability were assessed under different detection scales of YOLO-v3 network model and different databases of noise spectrograms.(3)A model for separating noise signals in stations was established to realize the separation of noise signals under noise coincident conditions based on the blind source separation theory.Three common blind source separation algorithms were selected to separating linear instantaneous mixed signals and linear convolutional mixed signals.TDSEP blind source separation algorithm was considered to have the best separation performance by assessing the separation capability of the three algorithms.The separated signal obtained from the TDSEP blind source separation algorithm had a strong correlation with the original signal.(4)The measured noise of the four stations was identified and separated and the composition and contribution of sound sources in the station was obtained,and the guidance method of noise reduction were proposed.The YOLO-v3 network model was used to recognize the sound signal characteristics in the database of noise spectrograms and the period of time for four conditions were labeled.TDSEP blind source separation algorithm was used to separate overlapped signals and the composition and contribution of sound sources in the station were obtained,which can provide guidance for the proposal of noise reduction measures.