摘要
为了检测出异常航迹数据从而提高航迹数据挖掘的精确性,将航迹异常检测转化为无监督学习问题,研究了基于VAE-LSTM的航迹异常检测算法。引入残差结构到LSTM中,建立残差门LSTM,通过将变分自编码器中的BP神经网络层替换为残差门LSTM层,实现对变分自编码器的改进,并构建了VAE-LSTM航迹异常检测模型。模型输入为航迹的速度、加速度、真航向和曲率半径运动特征,输出为航迹点特征的重建概率,重建概率小于概率阈值的航迹点为异常航迹点,包含异常航迹点的航迹判定为异常航迹。以长江水域内的航迹数据进行验证并与多种机器学习异常检测算法进行对比。VAE-LSTM航迹异常检测算法的召回率达到了0.935,F1值达到了0.940,各项指标均高于对比算法,验证了方法的有效性。
To detect anomalies in ship trajectory data and improve the accuracy of the data mining method for ship trajectory data,an anomaly detection algorithm for ship trajectory data based on a VAE-LSTM model is proposed by using an unsupervised classification method.Residual structure is introduced to build a residual gate LSTM,and layers of the residual gate LSTM are used to replace the BP neural network layers,in order to improve the VAE model.An anomaly detection model for ship trajectory data is established by taking four-dimensional motion features,including velocity,acceleration,true course,and radius of curvature,as input;and reconstruction probability of the features as output.The ship trajectory data containing trajectory points which reconstruction probabilities are lower than a threshold is classified as anomalous data.Ship trajectory data of the Yangtze River are used to verify the proposed algorithm.By comparing with other machine learning algorithms,the results indicate that the recall value of the VAE-LSTM model reaches 0.935,the F1-score of the VAE-LSTM model reaches 0.940,and all results of the detection performance exceed those of the comparison algorithms,which verifies its effectiveness.
作者
常吉亮
谢磊
赵建伟
杨洋
CHANG Jiliang;XIE Lei;ZHAO Jianwei;YANG Yang(National Engineering Research Center for Water Transport Safety,Wuhan University of Technology,Wuhan 430063,China;School of Energy and Power Engineering,Wuhan University of Technology,Wuhan 430063,China)
出处
《交通信息与安全》
CSCD
北大核心
2020年第6期1-8,共8页
Journal of Transport Information and Safety
基金
国家重点研发计划项目(2019YFB1600600、2019YFB1600604)资助。
关键词
智能交通
航迹数据
异常检测
变分自编码器
无监督学习
intelligent transportation
ship trajectory data
anomaly detection
variational auto-encoder
unsuper⁃vised learning