基于集成LSTM自编码器的多维时间序列异常检测OACSTPCD
MULTIVARIATE TIME SERIES ANOMALY DETECTION METHOD BASED ON LSTM AUTOENCODER ENSEMBLE
针对长短时记忆网络自编码器(LSTM-AE)在多维时间序列(MTS)上异常检测效率低的问题,提出一种基于集成LSTM-AE(LAE)的MTS异常检测模型.该模型集成多个LSTM-AE分别重构正常MTS各子序列,并将各重构误差作为MTS的局部特征;利用全连接网络自编码器(FCAE)对各重构误差数据进行拟合,学习MTS数据的全局特征;根据FCAE的重构误差进行异常检测.在三个公共MTS数据集上的实验表明,与基准方法相比,在Precision、Recall和F1_score三个评价指标下分别最大提升0.058 4、0.118 4和0.078 6.
Aimed at the problem that the long-short term memory AutoEncoder(LSTM-AE)is inefficient in anomaly detection on multivariate time series(MTS),a model named LSTM-AE Ensemble(LAE)is proposed.LAE integrated multiple LSTM-AEs to reconstruct each sub-sequence of normal MTS,and treated each reconstruction error as a local feature of the MTS.A fully connected network AutoEncoder(FCAE)was used to fit the reconstruction error data,so that LAE could capture the global features of MTS data.Anomaly detection was carried out according to the reconstruction error of FCAE.Experiments on three public MTS datasets show that compared with the benchmark method,LAE has better MTS anomaly detection performance with the maximum improvement 0.058 4,0.118 4,and 0.078 6 respectively on the terms of Precision,Recall,and F1_score.
李亚静;霍纬纲;丁磊
中国民航大学计算机科学与技术学院 天津 300300
计算机与自动化
多维时间序列异常检测LSTM-AE集成学习
Multivariate time seriesAnomaly detectionLSTM-AEEnsemble learning
《计算机应用与软件》 2024 (001)
285-290 / 6
中央高校基本科研业务费项目中国民航大学专项(3122019190).
评论