基于对偶对抗学习的多维时间序列异常检测

Anomaly Detection in Multidimensional Time Series Based on Dual Adversarial Learning

时间序列中异常点的无监督检测是一个具有挑战性的问题,要求模型能够快速准确地发现异常数据。VAE类深度神经网络模型能在数据压缩和恢复中学习数据的特征,但由于训练过程中缺乏对抗性,无法更好地区分正常数据和异常数据特征,导致模型训练困难。针对该问题,本文提出一种基于对偶对抗思想的改进多维时间序列异常检测方法。首先利用滑动窗口将数据集划分为合适的长度的序列,使用正常序列数据训练模型。继而利用对偶结构加强两组编码器解码器之间的对抗性,以更好地学习正常数据特征,减少训练难度。最后,将含有异常数据的待测数据放入训练好的模型,根据待测序列在模型中的异常得分,结合阈值技术进行异常判定,并从待测数据中获得异常序列片段,计算评价指标。实验表明,本文方法Dual-AE具有模型容易训练且稳定性强的特点,并且相对于USAD方法,在水文数据集SWaT上F1分数提升了0.01,召回率提升了0.01,在WADI数据集上F1分数提升了0.09,召回率提升了0.02。异常检测性能指标上,比现有的生成式异常检测模型有明显提升。

Unsupervised detection of outliers in time series was a challenging problem,and the model was required to find outliers quickly and accurately.The VAE deep neural network model could learn the characteristics of data in data compression and recovery,due to the lack of confrontation in the training process,it couldn’t better distinguish the characteristics of normal data and abnormal data,which made the model training difficult.To solve this problem,this paper proposed an improved multidimensional time series anomaly detection method based on the idea of dual confrontation.Firstly,the data set was divided into sequences of appropriate length by using a sliding window,and the model was trained using normal sequence data.Then,the dual structure was used to strengthen the confrontation between the two sets of encoders and decoders,so as to better learn the characteristics of normal data and reduce the difficulty of training.Finally,the test data containing abnormal data was put into the trained model.According to the anomaly score of the sequence to be tested in the model,the anomaly judgment was made in combination with threshold technology.Abnormal sequence fragments were obtained from the data to be tested,and the evaluation index was calculated.Experiments show that the proposed method of Dual-AE has the characteristics of easy training and strong stability,comparing with USAD method,and the F1 score and recall rate are increased by 0.01 and 0.01 on the hydrological dataset SWaT,and on the WADI dataset,the F1 score is increased by 0.09 and the recall rate is increased by 0.02.In terms of anomaly detection performance indicators,it is significantly improved in comparing with the existing generative anomaly detection models.