Anomaly detection for time series using temporal convolutional networks and Gaussian mixture model
Anomaly detection for time series using temporal convolutional networks and Gaussian mixture model
复制标题
使用时间卷积网络和高斯混合模型的时间序列异常检测
DOI：
10.1088/1742-6596/1187/4/042111
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Xinyu Zhang
中科院分区：
文献类型：
--
作者：
Jianwei Liu;Hongwei Zhu;Yongxia Liu;H. Wu;Yunsheng Lan;Xinyu Zhang
异常检测作为时间序列分析中的一个重要研究领域，在网络安全、医疗健康、物联网、故障诊断等诸多场合都有着实际而重要的应用，但由于时间序列固有的数据量巨大、正常数据与异常数据不平衡等特点，为时间序列的异常检测增加了额外的约束和挑战。提出了一种新的异常检测框架，该框架利用时间卷积网络提取时间序列的特征，并结合高斯混合模型和贝叶斯推理检测系统的异常。为了评估该方法的有效性，在两个典型的时间序列数据集（包括脑电信号数据集和电气设备电流数据集）上进行了实验。实验结果表明，时间卷积网络能够有效地提取时间序列的显著特征，而基于贝叶斯推理的高斯混合模型在异常检测中具有良好的泛化能力和可靠性。同时，所设计的异常检测结构和分析方法也表明了该方法在其他时间序列特征提取和异常检测中的有效性和推广性。
Anomaly detection, as an important research field in the analysis of time series, has practical and significant applications in many occasions, such as network security, medical health, Internet of Things (IoT), fault diagnosis and so on. However, due to the inherent characteristics of time series, such as tremendous data volumes, the imbalance of normal data and abnormal data, additional constraints and challenges are added for anomaly detection for time series. We present a novel anomaly detection framework, which applies temporal convolutional networks to extract features of time series and combined Gaussian mixture model with Bayesian inference to detect anomalies of systems. In order to evaluate the effectiveness of our approach, experiments are carried out on two typical time series datasets including EEG dataset and current dataset of electrical equipment. The experiments indicate that temporal convolutional network can contribute to extracting salient features of time series and Gaussian mixture model with Bayesian inference has good generalization and reliability for anomaly detection. Meanwhile, the designed architecture and analysis approach of anomaly detection reveal the method’s effectiveness and generalization in the feature extraction and anomaly detection for other time series.