KL散度多模块滑动窗口慢特征分析的故障诊断方法

针对传统分块方法根据经验划分子块导致变量特征信息无法充分利用，其单一的建模方式忽略局部信息以及离线模型无法适应时变特性的问题，提出了一种KL (Kullback-Leibler)散度多模块滑动窗口慢特征分析方法。在正常工况数据集中，利用KL散度来度量变量间的距离，同时引入最小误差平方和准则进行聚类，分成两个距离最小的子模块；在此基础上利用慢特征分析方法对每个子模块进行建模，结合滑动窗口对每次采样的数据进行更新，得到最优模型，分别计算监测统计信息，利用支持向量数据描述对故障监测结果进行融合，实现故障诊断。并将该方法应用于田纳西伊斯曼过程的监控中，得到了较高的故障检测率和较低的虚警率，验证了该方法的可行性和有效性。

KL divergence multi-block moving window slow feature analysis method for fault diagnosis

Noted the limitation of existing methods to study the delay propagation mechanism in high-speed railway network, the Bayesian network was used to analyze delay dependencies between stations based on high-speed train operation data and schedule data. The percolation theory of complex network was applied to study the evolution of delay propagation clusters. Nanjingnan Railway Station in the largest delay propagation cluster and Changshanan Railway Station in the second largest delay propagation cluster were taken as examples to analyze the network at percolation threshold. Based on core delay propagation clusters, the station delay state prediction model was established. The results show that stations can be divided into 3 categories, namely, delay generator, delay mediator and delay absorber according to their characteristic in the aspect of delay propagation. Delay generator can not only spread the delay to stations close to them, but also spread the delay to the far away stations through delay propagation chain, which makes the network taking part of the delay generator as the center and spreading delay to the delay mediator and the delay absorber in a radiational delay propagation mode.