基于无监督学习的工业互联网时序数据事件检测

Unsupervised event detector for time series data in industrial Internet scenario

为了在不依赖事件标签的前提下对时间序列数据进行事件检测,提出一种基于概率模型的无监督时间序列数据事件抽取算法,该算法在广义似然比(GLR)算法的基础上,通过预先筛选和分组投票的策略,使得算法的事件抽取效率与抽取结果的准确程度都得到了一定的提升。此外,还提出了一种基于分裂式层次聚类的时序数据聚类算法,该算法将自上而下的分裂式层次聚类与NME算法相结合,在自适应估计结果类簇数量的前提下达到了较高的准确度。

To solve the problem of unsupervised event detection on time series data without event labels,an unsupervised event extraction algorithm based on probabilistic models was proposed,which improved the efficiency and accuracy of the event extraction results by using pre-filtering and group voting strategies based on Generalized Likelihood Ratio(GLR)algorithm.Besides,a time series data clustering algorithm based on hierarchical clustering was proposed,which combined divisive hierarchical clustering with Normalized Maximum Eigengap(NME)algorithm and achieved high accuracy while estimating the number of clusters automatically.