卢剑, 张学东, 张健钦, 郭小刚, 张悦颖. 利用卷积神经网络识别交通指数时间序列模式[J]. 武汉大学学报 ( 信息科学版), 2020, 45(12): 1981-1988. DOI: 10.13203/j.whugis20200035
引用本文: 卢剑, 张学东, 张健钦, 郭小刚, 张悦颖. 利用卷积神经网络识别交通指数时间序列模式[J]. 武汉大学学报 ( 信息科学版), 2020, 45(12): 1981-1988. DOI: 10.13203/j.whugis20200035
LU Jian, ZHANG Xuedong, ZHANG Jianqin, GUO Xiaogang, ZHANG Yueying. Identification of Traffic Index Time Series Pattern by Using Convolution Neural Network[J]. Geomatics and Information Science of Wuhan University, 2020, 45(12): 1981-1988. DOI: 10.13203/j.whugis20200035
Citation: LU Jian, ZHANG Xuedong, ZHANG Jianqin, GUO Xiaogang, ZHANG Yueying. Identification of Traffic Index Time Series Pattern by Using Convolution Neural Network[J]. Geomatics and Information Science of Wuhan University, 2020, 45(12): 1981-1988. DOI: 10.13203/j.whugis20200035

利用卷积神经网络识别交通指数时间序列模式

Identification of Traffic Index Time Series Pattern by Using Convolution Neural Network

  • 摘要: 城市交通时间序列模式识别主要基于距离计算和特征提取两种方法,但前者结果受到时间序列非线性特征的影响,后者难以提取多时间段Shapelet子序列。为准确识别城市交通指数数据模式,提出了基于卷积神经网络对交通时间序列数据进行模式识别的方法。首先将时间序列数据预处理成N维矩阵,确定神经网络模型的输入数据; 然后通过反复训练特征网络提取输入层的特征信息,确定交通指数的模式类型; 最后利用Softmax分类器进行模式分类,实现交通指数类别划分。实验以2016—2018年北京市全市交通指数时间序列模式识别为例,对比发现,该方法能更准确地将交通指数时间序列数据划分为符合真实情况的5类模式。该方法对时间序列数据模式识别更准确,适合呈现时间阶段性变化的数据间相关性研究和模式发现。

     

    Abstract: Urban traffic time series pattern recognition is mainly based on two methods: Distance calculation and feature extraction. The recognition results based on distance calculation are affected by the non-linear features of the time series, and it is also difficult to extract multi-time period Shapelet sub-sequences based on the feature extraction algorithm. Therefore, in order to accurately identify the data pattern of urban traffic index, a method for pattern recognition of traffic time series data based on convolutional neural network(CNN) was proposed. Firstly, the time series data are pre-processed and transformed into N-dimensional matrix to determine the input data of the neural network model. Then, the characteristic information of the input layer is extracted through repeated training of the characteristic network, and the pattern type of the traffic index is determined. Finally, Softmax classifier is used to classify the modes and realize the classification of traffic index. The time series pattern recognition of Beijing's traffic index from 2016 to 2018 was taken as an example. Compared with the traditional machine learning algorithm based on distance calculation, it was found that the CNN can more accurately divide the time series data of traffic index into five types of patterns that are consistent with the real situation. The experimental results show that the CNN is more accurate and effective for pattern recognition of time series data, and is suitable for correlation research and pattern discovery of time-phase change data.

     

/

返回文章
返回