Abstract: Due to the high dimensionality of multivariate time series data and the complex and variable distribution, anomaly detection on such data faces great challenges. Existing anomaly detection models generally suffer from high misjudgment rates and training difficulties when facing multivariate time series data. Moreover, most models only consider the spatiotemporal features of time series samples, which are not sufficient to represent the feature of time series . To solve the above problems, a multivariate time series anomaly detection model based on multi-domain feature extraction (MFE-TS) is proposed. First, starting from the original data domain, the LSTM network and the convolutional neural network are used to extract the temporal correlation and spatial correlation features of the multivariate time series respectively. Secondly, Fourier transform is used to convert the original time series into frequency domain space, and Transformer is used to learn the amplitude and phase features of the data in frequency domain space. Multi-domain feature learning can more comprehensively model time series features, and significantly improves the anomaly detection performance. In addition, the masking strategy is introduced to further enhance the feature learning ability of the model and make the model have a certain degree of noise resistance. Experimental results show that MFE-TS exhibits superior performance on multiple real multivariate time series data sets, while still maintaining good detection accuracy in noisy data sets.

Key words: multivariate time series, anomaly detection, unsupervised learning, multi-domain feature extraction

摘要： 由于多变量时间序列数据样本具有高维性，且分布复杂多变，对于此类数据的异常检测存在巨大挑战。现有异常检测模型在面对多变量时间序列数据集时普遍存在误判率高，训练困难等问题，且多数模型仅考虑时间序列样本的时空特征，对时间序列特征的学习并不全面。为了解决以上问题，提出一种基于多域特征提取的多变量时间序列异常检测模型(MFE-TS)。首先从原始数据域出发，使用LSTM网络与卷积神经网络分别提取多变量时间序列的时间相关性和空间相关性特征。其次，用傅里叶变换将原始时间序列转换到频域空间，并利用Transformer学习数据在频域空间的幅度与相位特征。通过多域特征学习能更全面地建模时间序列特征，从而提高对多变量时间序列的异常检测性能。此外，引入掩码策略，进一步增强了模型的特征学习能力，并使模型具备一定的抗噪性。实验结果表明，MFE-TS在多个真实多变量时间序列数据集上展现了优越的性能，同时在含有噪声的数据集中仍能保持较好的检测精度。

关键词: 多变量时间序列, 异常检测, 无监督学习, 多域特征提取