基于Seq2Seq深度自编码器的时间序列异常检测方法研究

Method of time series anomaly detection based on Seq2Seq depth autoencoder

传统的时间序列异常检测方法大多以数据点作为检测单位,通过训练模型预测下一时刻数据,这类方法的缺点是没有考虑时间序列数据的特性,即序列模式的多样性。因此文中提出一种基于Seq2Seq深度自编码器的时间序列异常检测方法,以更好地挖掘时间序列数据中的异常序列模式。此方法使用Bi-LSTM网络作为深度自编码器,其输入输出均为序列,使用深度自编码器对时间序列进行编码压缩和解码重建。通过计算重建序列与原始序列之间的重建误差,并设置重建异常比率以获取误差阈值,将重建误差大于此阈值的时间序列视为异常序列。异常时间序列的发现取决于模型对原始序列的重建效果,通过在空气质量时间序列数据上的实验,模型初步达到了不错的检测效果,证明了所提方法的可行性。文中方法为时间序列异常检测提供了新的途径。

Most of the traditional time series anomaly detection methods take the data points as the detection unit and predict the next moment data by the training model.The disadvantages of these methods are that they do not consider the characteristics of time series data,that is,the diversity of sequence patterns.A time series anomaly detection method based on Seq2Seq depth autoencoder is proposed,which can mine anomaly sequence patterns in time series data preferably.In the method,the Bi-LSTM network is applied as the depth autoencoder,whose input and output are all sequences.The time series are encoded,compressed,decoded and reconstructed by the depth autoencoder.By calculating the reconstruction error between the reconstructed sequence and the original sequence,and setting the reconstruction anomaly ratio to obtain the error threshold,the time series whose reconstruction error is greater than this threshold are regarded as abnormal sequences.The discovery of anomaly time series depends on the model′s reconstruction effect on the original series.The results of the experiments on air quality time series data indicate that the model initially achieves a good detection effect,which proves the feasibility of the proposed method and provides a new method for the time series anomaly detection.