基于集成LSTM-AE的时间序列异常检测方法

Time series anomaly detection method based on integrated LSTM-AE

针对基于LSTM-AE(长短期记忆自编码器)的时间序列异常检测方法在异常检测阶段对正常序列和异常序列的重建误差不能明显分化,致使重建误差在阈值附近的序列无法正确判断正常与异常的问题,提出一种集成LSTM-AE框架——LAEE(长短期记忆自编码器集成框架).将训练阶段拆分为预训练和预检测两个阶段,在预训练阶段训练多个隐层维度不同的LSTM-AE,通过预检测阶段的表现挑选基检测器,并计算其各自权重;在异常检测阶段,通过对每个基检测器产生的重建误差进行加权集成获得新的重建误差矩阵,进行异常识别.通过在两类数据集上的实验结果表明:所提方法使得检测目标正常序列与异常序列重建误差分化程度加大,提高了异常检测精度.

Aiming at the problem that when using long short term memory-autoencoder(LSTM-AE) for time series anomaly detection,the reconstruction errors between normal sequence and abnormal sequence could not be differentiated significantly in the phase of anomaly detection,resulting in the sequence with reconstruction errors near the threshold could not be correctly identifed,a integrated LSTM-AE framework was proposed,which was LSTM autoencoder ensembles(LAEE).The training stage was divided into pre-training and pre-detection.In the pre-training stage,LSTM-AEs with different hidden layer dimensions were trained,the base detectors were selected through the detection performance of the pre-detection stage,and their weights were calculated.In the phase of anomaly detection,a new reconstruction error matrix was obtained by weighted integration of the reconstruction errors generated by each base detector for anomaly identification.Experimental results on two kinds of data sets show that the proposed method enhances the reconstruction error distinction between normal sequence and anomaly sequence,and improves the accuracy of anomaly detection.