基于深度特征聚类的高排放移动污染源自动识别  

Automatic Identification of High-emitting Vehicle Based on Deep Feature Clustering

作　　者：许镇义 王仁军 张聪 王瑞宾 夏秀山 XU Zhen-yi;WANG Ren-jun;ZHANG Cong;WANG Rui-bin;XIA Xiu-shan(Hefei Comprehensive National Science Center Artificial Intelligence Research Institute,Hefei 230088,China;School of Computer Science and Technology,Anhui University,Hefei 230601,China;Hefei Municipal Environmental Protection Bureau,Hefei 230601,China;Institute of Advanced Technology,University of Science and Technology of China,Hefei 230088,China)

机构地区：[1]合肥综合性国家科学中心人工智能研究院,合肥230088 [2]安徽大学,计算机科学与技术学院,合肥230601 [3]合肥市生态环境局,合肥230601 [4]中国科学技术大学,先进技术研究院,合肥230088

出　　处：《交通运输系统工程与信息》2021年第6期298-309,共12页Journal of Transportation Systems Engineering and Information Technology

基　　金：国家自然科学基金(62103124)。

摘　　要：传统的高排放移动源识别方式是将采集的尾气数据与预先设定的排放阈值进行比较判定,但是,排放阈值的设定很大程度上取决于人为标准,并且忽视了外部因素对尾气排放的影响,无法真正反映移动源排放水平。针对此问题,本文结合机器学习算法,提出一种基于深度特征聚类的高排放移动源识别方法。首先,利用随机森林算法筛选出不同污染物(CO、HC、NO)排放的主要影响特征;其次,对多维影响特征进行聚类分析,获取高排放类别标签;最后,训练得到基于深度森林的移动污染源分类模型,自动识别高排放目标源。通过对比实验,在合肥市机动车污染遥测数据集上验证了所提方法的有效性。The traditional approach of identifying high emission mobile sources is to compare the collected tailpipe data with pre-defined emission thresholds.However,the setting of emission thresholds depends mainly on human standards,and this method ignores the influence of external factors on tailpipe emissions,which cannot exactly reflect the emission level of mobile sources.To address this problem,this paper combines different machine learning algorithms and proposes a method for identifying high emission mobile sources based on deep feature clustering.The random forest algorithm is first used to filter out the main impact features of different pollutant(CO/HC/NO)emissions.Then,the multidimensional impact features are clustered to obtain the high emission category labels.A deep forest-based mobile source classification model is trained to automatically identify the high emission target sources.The experiment results on the telemetry dataset of mobile source pollution in Hefei verify the effectiveness of this method.

关 键 词：信息技术 高排放识别 特征聚类 深度森林 

分 类 号：U495[交通运输工程—交通运输规划与管理]