Abstract:

A large-scale road network traffic flow data imputation method based on the symmetrical Residual UNet (RU-Net) model is proposed for the traffic flow data missing caused by the scarcity or failure of traffic data collectors in the process of road network traffic data collection. By gridding the traffic flow data and channelizing the timing, then forming the tensor data format for convolution operation, this method uses the coding and decoding ability of RU-Net to encode the traffic flow data, and keeps the distortion small in the decoding process, to learn the internal multi-factor coupling characteristics of the traffic flow data. Residual learning can improve the signal-to-noise ratio of traffic flow data after coding, reduce the compression rate, and further improve the repair accuracy. Experimental results show that the RU-Net model can effectively repair the large-scale network traffic flow data under different data missing rates and different missing patterns by mapping relationship between the traffic flow characteristics learning history, non-fault collectors' data, and the data to be repaired.

Key words: intelligent transportation, traffic data recovery, residual U- Net, large- scale road network, residual learning

摘要：

针对路网交通数据采集过程中，采集设备稀缺或故障等原因造成路网交通流量数据缺失问题，提出基于对称残差U型网络(Residual U-Net，RU-Net)模型的大规模路网交通流量数据修复方法.通过将路网交通流量数据网格化和时序通道化操作，构成可供卷积操作的张量数据格式；利用RU-Net编码解码能力，对交通流量数据进行编码；在解码过程中保持失真度较小，使模型学习到交通流量数据内部多因素耦合特性.通过残差学习使交通流量数据编码后的信噪比提升，压缩率降低，提升模型修复精度.实验结果表明，RU-Net模型能够利用交通流量特性学习历史和非故障采集点数据与待修复数据的映射关系，在不同数据缺失率，不同缺失模式下，高效地完成对大规模路网交通流量数据的修复.

关键词: 智能交通, 交通数据修复, 残差U型网络, 大规模路网, 残差学习