Multivariate Time-series Anomaly Detection using SeqVAE-CNN Hybrid Model
Multivariate Time-series Anomaly Detection using SeqVAE-CNN Hybrid Model
复制标题
使用 SeqVAE-CNN 混合模型进行多变量时间序列异常检测
DOI：
--
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Ho
中科院分区：
文献类型：
--
作者：
Tae;Dongkun Lee;Yuchae Jung;Ho
异常检测已成为信息技术、制造业、金融等众多行业公认的重要研究领域。最近，利用包括机器学习算法在内的当前深度学习方法进行了各种异常检测研究。然而，由于异常数据的不平衡性和多变量的复杂性，多变量时间序列异常检测是一个具有挑战性的问题。在本文中，我们使用无监督方法提出了基于深度学习的seqvee - cnn模型。我们的模型结合了变分自编码器（VAE）和卷积神经网络（CNN），利用Seq2Seq结构捕获多变量时间序列中的时间相关性和空间特征。为了证明我们的方法的性能，我们在来自不同领域的8个数据集上评估了我们的模型。实验结果表明，我们的模型在8个数据集中的6个数据集上记录了最高的AUROC和F1分数，比其他模型具有更好的异常检测性能。
Anomaly detection has been recognized as an important research area in many industries such as Information Technology, manufacturing, finance, etc. Recently, diverse research for anomaly detection has been conducted utilizing current deep learning methods including machine learning algorithms. However, multivariate time-series anomaly detection can be challenging problems because of the imbalance of anomaly data and the complexity of multivariate. In this paper, we propose a SeqVAE-CNN model based on deep learning using an unsupervised approach. Our model combines Variational Autoencoder (VAE) with Convolutional Neural Networks (CNN) as utilizing Seq2Seq structure to capture temporal correlations and spatial features in multivariate time-series. To demonstrate the performance of our approaches, we evaluate our model on 8 datasets from various domains. The experimental results demonstrate that our model has better performance of anomaly detection than other models by recording the highest AUROC and F1 scores on six of the eight datasets.