基于轻量化GE-GRU-VAE模型的多维时间序列异常检测

Anomaly Detection in Multidimensional Time Series Based on a Lightweight GE-GRU-VAE Model

摘要: 在工业应用中，时间序列的无监督异常检测至关重要，因为它能显著减少人工干预的需求。时间序列数据通常具有非平稳性、高维、异常稀缺等特点，使得对其进行异常检测具有挑战性。本文提出了一种基于注意力机制、图嵌入技术和VAE相结合的无监督多维时间序列异常检测模型GE-GRU-VAE。首先，在GE-GRU-VAE编码器中采用MLP和多头注意力结构进行局部特征提取，获得输入数据内在特征的分布参数。其次，采用重参数化得到其低维图嵌入特征。然后，在GE-GRU-VAE解码器中采用基于图嵌入与GRU的GE-GRU模块进行时间序列重构，通过无监督学习得到最优模型。最后，根据双阈值异常判定方法判断序列是否异常。在两个公共数据集SWaT和WADI上验证了所提模型的有效性。实验结果表明，GE-GRU-VAE不仅具有较低的时间与空间复杂度，而且具有良好的异常检测精度。

Abstract: In industrial applications, unsupervised anomaly detection for time series is of crucial importance, as it can significantly reduce the need for manual intervention. Time series data typically exhibits characteristics such as non-stationarity, high dimensionality, and scarcity of anomalies, which pose challenges to anomaly detection. This paper proposes an unsupervised multi-dimensional time series anomaly detection model named GE-GRU-VAE, which combines the attention mechanism, graph embedding technology, and Variational Autoencoder (VAE). Firstly, in the encoder of GE-GRU-VAE, a Multi-Layer Perceptron (MLP) and a multi-head attention structure are employed for local feature extraction to obtain the distribution parameters of the intrinsic features of the input data. Secondly, reparameterization is used to derive its low-dimensional graph embedding features. Then, in the decoder of GE-GRU-VAE, a GE-GRU module based on graph embedding and Gated Recurrent Unit (GRU) is adopted for time series reconstruction, and the optimal model is obtained through unsupervised learning. Finally, a dual-threshold anomaly determination method is used to judge whether the series is abnormal. This paper verifies the effectiveness of the proposed model on two public datasets, SWaT and WADI. Experimental results show that GE-GRU-VAE not only has low time and space complexity but also achieves excellent anomaly detection accuracy.