Aiming at the problem of ship behavior anomaly detection based on massive track data without behavior pattern label, an unsupervised track anomaly detection method based on GRU-VAE model is proposed. The abnormal behavior of the target is found by detecting track anomaly, which is implemented in two steps: model training stage and anomaly detection stage. In the model training stage, the timing modeling ability of GRU gated cyclic Autoencoder (VAE) model is introduced. The Gate Recurrent unit-variational Autoencoder model is trained by historical track data without abnormal information labels. According to the reconstruction loss distribution, the normal distribution method or percentile method is used to delimit the confidence interval as the reconstruction loss threshold. In anomaly detection phase, the model of real-time track data set for testing, regarding the damage threshold of the track to refactor losses above points as abnormal track points, when the track is in the sequence beyond the proportion of abnormal track point accounted threshold, it is judged to be abnormal track sequence, combined with the data anomalies target behavior information are presented to the first-line staff. The experimental results on AIS data show that the highest F1 score of the model is up to 86.36%, and the recall rate is up to 95%. The high sensitivity and low miss alarm rate of this method to abnormal track meet the reconnaissance requirements of first-line units.
LI Lei, ZHANG Jing, OUYANG Qicheng, ZHOU Mingkang. Unsupervised abnormal track detection method based on GRU-VAE[J]. Command Control and Simulation, 2023, 45(5): 51-64. DOI: 10.3969/j.issn.1673-3819.2023.05.008
在该研究方向上,基于无监督网络模型的异常检测方法值得关注。文献[5]以自编码器(Autoencoder, AE)作为水位异常检测框架核心,通过学习正常数据的特征分布,将输出的重构损失作为异常分数并设定阈值,进而实现对水位数据的异常检测。变分自编码器(Variational Autoencoder,VAE)[6]以时间序列中潜在变量分布参数的重建概率作为异常检测的度量,不受数据结构限制,比自编码器在数据重构上更有优势。文献[7]采用变分自编码器模型对脑电数据进行异常检测,降低癫痫发作带来的安全风险。然而,以上两种模型针对航迹数据的时序建模能力还有待提高。长短时记忆网络(Long-Short Term Memory, LSTM)和门控循环单元(Gated Recurrent Unit networks, GRU)作为循环神经网络的重要变体,在时间序列建模问题上具有优势。在其他领域中,以上两种自编码器模型常配合使用以增强其时序建模能力。文献[8]提出一种基于长短时记忆网络模型LSTM的无监督异常检测模型,在此基础上,文献[9]将AE和LSTM相结合,进一步提高了模型在时序数据异常检测上的表现。文献[10]将LSTM作为VAE编码器的输入层和解码器的输出层,进一步提升了模型对时序数据的重构能力。而GRU[11]的参数量比LSTM更少,保持了LSTM的优异性能,同时其结构更加简单,过拟合风险更低,文献[12]将其应用于飞机振动数据的异常检测,并取得了良好效果。
本文构建了以GRU和VAE为主要结构元素的无监督神经网络模型,提出了一种基于GRU-VAE模型的无监督航迹异常检测方法。在模型训练阶段,使用历史航迹数据集对GRU-VAE模型进行训练,对数据集中所有航迹序列进行重构,将重构航迹序列和原始航迹序列的平均绝对距离作为模型输出的重构损失,基于输出重构损失的分布类型确定其在任意置信度下的置信区间,并将其作为航迹点重构损失门限;在异常检测阶段,该模型对实时航迹数据进行检测,将损失超出航迹点重构损失门限的航迹点视为异常航迹点,将异常航迹点占比超出阈值的航迹序列判定为异常航迹序列,再结合数据特征和重构损失异常情况向一线人员发送目标的异常行为信息。AIS数据实验结果表明,模型平均F1分数高达86.36%,查全率达95%。对比现有相关研究成果,本方法在查准率、查全率和模型F1分数上比实验性能稍次的基于长短时记忆网络的变分自编码器模型(Long-Short Term Memory-Variational Autoencoder,LSTM-VAE)方法[10]分别高出1.68%、1.18%和1.50%,对异常航迹具有高灵敏度和低漏警率,符合战场态势认知需求。
HUXF, HEX Y, RAOD H, et al. Research on mechanism of command and coordination for system of systems operation based on complex network theory[J]. Journal of Command and Control, 2015, 1(1): 5-13.
CUILF, GUOY H, SHAOP Z, et al. Enhancing situational awareness on the battlefield using spatio-temporal big data[J]. National Defense Technology, 2021, 42(2): 127-132.
DUANY X, LIUC Y, WEIW F. Review of key technologies for battlefield situational awareness[J]. Fire Control & Command Control, 2021, 46(11): 1-11, 19.
[4]
RUFFL, KAUFFMANNJ R, VANDERMEULENR A, et al. A unifying review of deep and shallow anomaly detection[J]. Proceedings of the IEEE, 2021, 109(5): 756-795.
YOUS, CHOB H, SHONY M, et al. Semi-supervised automatic seizure detection using personalized anomaly detecting variational autoencoder with behind-the-ear EEG[J]. Computer Methods and Programs in Biomedicine, 2022, 213: 106 542.
YANGF, HEZ W, HEF. Detection method of abnormal ship behavior based on LSTM neural network[J]. Journal of Wuhan University of Technology(Transportation Science & Engineering), 2019, 43(5): 886-892.
[9]
YUANY, LINL, HUOL Z, et al. Using an attention-based LSTM encoder-decoder network for near real-time disturbance detection[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 1819-1832.
PARKD, HOSHIY, KEMPC C. A multimodal anomaly detector for robot-assisted feeding using an LSTM-based variational autoencoder[J]. IEEE Robotics and Automation Letters, 2018, 3(3): 1544-1551.
PENGJ T, ZHUY X, WANGH, et al. GRU network for flight-data anomaly detection based on FPGA[J]. Micro-electronics & Computer, 2021, 38(11): 67-73.
[13]
DOERSCHC. Tutorial on variational autoencoders[EB/OL]. 2016: arXiv: 1606.05908. https://arxiv.org/abs/1606.05908
[14]
CHUNGJ, GULCEHREC, CHOK, et al. Empirical evaluation of gated recurrent neural networks on sequence modeling[EB/OL]. 2014: arXiv: 1412.3555. https://arxiv.org/abs/1412.3555
[15]
方积乾, 陆盈. 现代医学统计学[M]. 北京: 人民卫生出版社, 2002.
FANGJ Q, LUY. Advanced medical statistics[M]. Beijing: People’s Medical Publishing House, 2002.
[16]
何晓群. 现代统计分析方法与应用[M]. 北京: 中国人民大学出版社, 1998.
HEXQ. Modern statistical analysis methods and applications[M]. Beijing: China Renmin University Press, 1998.
[17]
LIUF T, TINGK M, ZHOUZ H. Isolation-based anomaly detection[J]. ACM Transactions on Knowledge Discovery from Data, 2012, 6(1): 1-39.
[18]
DONGX F, LIUJ X, ZHANGW N, et al. Identifying anomaly aircraft trajectories in terminal areas based on deep auto-encoder and its application in trajectory clustering[J]. Transactions of Nanjing University of Aeronautics and Astronautics, 2020, 37(4): 574-585.
[19]
LIZ Y, CHENW X, PEID. Robust and unsupervised KPI anomaly detection based on conditional variational autoencoder[C]// 2018 IEEE 37th International Performance Computing and Communications Conference (IPCCC). Orlando, 2019: 1-9.