基于KPCA的多变量时间序列数据异常检测方法研究

摘要

A new MTS (Multivariate Time Series) anomaly detection method based on KPCA (Kernel Principal Component Analysis)technique is introduced, which use the principal component direction vector in the high dimension feature space as the data feature and obtains it through KPCA technique. The distribution of the principal component direction vector is denoted by vMF distribution, and the model parameter is learned through actual data training. The anomaly is detected while probability of the anomaly metric is upper than the predefined threshold. Compare with the conventional method, the new method is independent of the priori expert knowledge, and can adjust the model parameter through automatic leaning. The method can be used to different system by choosing different kernel function. The effectiveness of the method is proven by actual experiment.%介绍了一种采用的KPCA技术获取多变量时间序列数据高维特征空间的主成方向矢量,使用主成方向矢繁内积作为异常的度量,并采用vMF分布表征主成方向矢量分布来进行多变量时间序列数据异常检测的方法;检测过程中使用历史数据训练获取分布模型的参数估计,通过计算实际数据主成方向矢量在训练模型的概率来判断异常的发生;与传统的异常检测方法相比,该方法不依赖先验的专家知识,且能够通过训练学习自动调节模型参数,可用于不同系统的异常检测中;实验表明,该方法具有较高的有效性.