过程控制异常值的在线检测方法研究  

Method for Outlier Detection in Process Control Field

作　　者：刘芳[1] 毛志忠[1,2] 

机构地区：[1]东北大学信息科学与工程学院,辽宁沈阳110004 [2]东北大学流程工业综合自动化教育部重点实验室,辽宁沈阳110004

出　　处：《计量学报》2013年第1期84-89,共6页Acta Metrologica Sinica

基　　金：国家“863”计划项目（2007AA042194）

摘　　要：提出了一种改进输入形式的径向基网络RBFN与自回归隐马尔可夫模型ARHMM相结合的异常数据检测方法，并通过引入核空间概念，用以解决过程工业中大量过程数据要求在线异常检测问题。该方法利用改进的RBF网络在核空间内预测待检测数据值，并根据核空间内的预测值与实际值偏差的大小，利用核ARHMM检测数据异常情况。改进的RBF网络能够方便地引入遗忘因子以及惩罚因子，从而增加算法的鲁棒性，提高检测的准确性。采用核ARHMM检测算法可以直接对数据异常情况作出准确判断，从而避免事先确定检测阈值的问题。通过实验与应用证明了该算法的实用性，与AR模型检测方法比较，该方法更适合于过程数据的异常检测问题。Aiming at the characteristics of data in process industry which are large volume of data and on-line detection, an outlier detection algorithm which combines the improved RBF network and ARHMM is proposed. In the new algorithm, improved RBF network is used to model base on major data in kernel space, and then according to the residual errors,the detection results are made by kernel ARHMM. Forgetting factor and penalty factor are introduced by improved RBF network, which can make the algorithm more robust and accuracy. In order to avoid preselecting the detection threshold, KARHMM is used to detect outlier in process industry. The practicality is proved by experimentation and application, and through the comparison with AR model, it shows that the nonlinear KARHMM algorithm is more suitable for process data.

关 键 词：计量学 过程数据 被控对象 异常数据检测 径向基函数网络 核自回归隐马尔可夫模型 

分 类 号：TB9[一般工业技术—计量学]