在我国快速城镇化发展过程中，城市环境噪声污染危害日益凸显，然而由于城市噪声产生的多元性和实时性，现有噪声监管体系难以及时、全面、精细化监测和实时动态管控噪声环境。城市噪声地图是将声源数据与交通信息等综合分析计算后生成反映城市噪声水平状况的数据地图，被公认是一种相对有效、廉价及全面定量化噪声的决策管理工具。本文通过引入实时噪声、实时交通和空间兴趣点数据，实现动态噪声的大数据建模与可视化，对噪声地图推广与城市噪声精细化管理具有重要意义。论文分别从运用已有静态噪声模型和改进和基于移动传感器监测的空间数据挖掘进行噪声大数据建模研究。第一，以LUNOS静态噪声模型为基础，引入噪声自动监测数据、实时交通数据和兴趣点数据，运用神经网络建模方法开展噪声与相关大数据自学习建模。第二，基于智能机噪声监测数据，引入实时交通数据和兴趣点大数据，综合运用神经网络、空间相似度分析和支持向量机等大数据建模方法，实现对城市环境噪声的建模与噪声分类工作。论文以大连和北京为研究案例。基于LUNOS改进的噪声建模结果表明，通过引入相关大数据，发现噪声和道路长度、车辆数、等待时间、车速呈弱正相关性，与路过时长呈正相关性，拥堵等级是对噪声影响最显著的变量。通过建立区分拥堵等级的线性回归建模，在拥堵等级为1、2、3时率定数据R从0.54提高到0.77、0.91，验证数据R从0.53提高到0.80、0.71；通过神经网络建模，选定3层3神经元的最佳网络结果作为噪声预测模型，模型R达到0.838。基于移动传感器监测和空间相似度分析的噪声研究结果表明，基于8个参考点的最优Hsim函数率定误差均值9.42dBA，方差12.14dBA，验证误差均值9.46dBA，方差12.19dBA；基于移动时间窗口的噪声神经网络模型率定数据R均值达0.88，能够较好模拟噪声动态变化，减少噪声源在不同时段的表现差异性；通过支持向量机和决策树的噪声种类分类，可以实现对区域交通噪音或社会噪音作为主要声源的区分，并通过可视化方法可以观察噪声主要类型空间分布特征。

With the rapid development of the urbanization in China, the jeopardy of urban environmental noise pollution is getting obvious, however, due to the complexity of urban scale noise, a timely, comprehensive, and intensive noise monitoring system is needed now to control the environmental noise dynamicly. Cities noise map is a comprehensive analysis and calculation on the sound source data, traffic information and other factors to reflect the noise level of a data map, which is recognized as a relatively effective, inexpensive and comprehensive display of the noise strategic decisions’ management tool. In the process of noise simulation, with high cost in drawing and lack of parameters localization, a noise map includes only traffic noise and static noise in a particular moment, so it’s of great significance for the noise map’s promotion and fine management on city noise to introduce real-time noise, real-time traffic and POI data, and to realize the big data modeling of dynamic noise and visualization.This paper is conducted to research the big date of noise from two aspects including the improvements on the existing noise modeling and monitoring methods based on the noise sensor. The research of urban noise map is based on LUNOS model and neural network method. LUNOS noise model is the basis of static noise. By introducing automaticly noise monitoring data, real-time traffic data and POI data, noise and related big date are associated to model by self-learning with the method of neural network modeling. The research of urban noise map is based on mobile sensors to monitor and Support Vector Machine. Urban environmental noise modeling and noise classification are achieved with the intelligent machine monitoring noise, the introduction of real-time traffic data and POI big data, the integrated use of neural network, spatial similarity analysis, vector machines supported and other big data modeling methods.The paper takes a deeper research on Dalian and Beijing as case studies, and big data modelings of noise were carried out in the two cities respectively. Results of the improved noise modeling based on LUNOS show that by introducing relevant big data, noise has a weak positive correlation with road length, number of vehicles, waiting time and vehicle speed, and a positive correlation with the time passing by, while the congestion level is the most significant variable that influces the noise. By distinguishing the linear regression modeling of congestion level, it’s gained that when the congestion level is 1, 2, and 3, the R of calibration is 0.54，0.77，0.91， the R of validation is 0.53，0.80，0.71 respectively. Through Artificial Neural Network modeling, the optimal network results of the three layers and three neurons structure were finally selected to be noise prediction model. The R of it can reach 0.838, which is much butter than most linear models( when the congestion level is 1, 2), but a little bit lower than linear nodels of level 3. The noise research results based on the mobile sensors monitor and spatial similarity mining shows tha Hsim function with 8 connected point is the best model，the error of calibration is 9.42dBA(mean) and 12.14dBA(var), the error of validation is 9.46dBA(mean) and 12.19dBA(Var). Moving time window’s R of calibration reaches0.88, ten neurons neural network noise models can better simulate dynamic changes of noise, and reduce the differences when the noise source performs at different times. By the species classification of Support Vector Machines and numbers of decision, it can be achieved to distinguish the regional traffic noise or social noise that are both the main sound source, and it can also observe the main types of noise in the spatial distribution through the method visualization.