郭金玉 赵文君 李元

摘 要：针对核熵成分分析算法（kernel entropy component analysis，KECA）为不同的故障选择相同的核参数影响检测效果的问题，提出了一种基于集成核熵成分分析（ensemble kernel entropy component analysis，EKECA）算法的工业过程故障检测方法。首先，选取一系列具有不同宽度参数的核函数将非线性数据投影到核特征空间，选取Rényi熵值贡献较大的特征值和特征向量，得到转换后的得分矩阵，建立多个KECA子模型;然后，将测试数据投影到各KECA子模型上，计算各KECA子模型的统计量，得到检测结果;最后，将各KECA子模型的检测结果利用Bayesian决策进行概率换算，利用集成学习法计算检测结果统一的统计量，判断其是否超出控制限，并将该算法应用于数值例子和TE过程。仿真结果表明，与传统的EKPCA，KECA等算法相比，所提方法有效提高了故障检测率，降低了误报率。新方法解决了传统KECA算法中不同故障核参数的选择问题，为提高KECA算法在非线性工业过程故障检测中的性能提供了参考。

关键词：自动控制技术其他学科;核熵成分分析;高斯核函数;Bayesian决策;集成学习法

中图分类号：TP277   文献标识码：A

doi：10.7535/hbkd.2021yx05006

收稿日期：2021-06-03;修回日期：2021-09-30;责任编辑：王淑霞

基金项目：国家自然科學基金（61673279）;辽宁省教育厅一般项目（LJ2019007）

第一作者简介：郭金玉（1975—），女，山东高唐人，副教授，博士，主要从事故障诊断、生物特征识别算法及应用方面的研究。

E-mail：969554959@qq.com

Fault detection of industrial process based on ensemble kernel entropy component analysis algorithm

GUO Jinyu，ZHAO Wenjun，LI Yuan

（College of Information Engineering，Shenyang University of Chemical Technology，Shenyang Liaoning，110142，China）

Abstract：To solve the problem caused by kernel entropy component analysis （KECA） for selecting the same kernel parameters for different faults，a fault detection of industrial process based on ensemble kernel entropy component analysis （EKECA） was proposed.Firstly，a series of kernel functions with different width parameters were selected to project the nonlinear data into the kernel feature space.The eigenvalues and eigenvectors with large contribution to Rényi entropy were selected to obtain the transformed score matrix.The multiple KECAsubmodels were established.Secondly，the test data were projected onto each KECA submodel.The statistics of each KECA submodel were calculated to obtain the detection results.Finally，the detection results of each KECA submodel were turned into probability by Bayesian decision.The unified statistics were calculated by ensemble learning strategy and judged whether it exceeds the control limit.The algorithm was applied to a numerical example and the TE process.The simulation results show that the proposed algorithm can effectively improve the fault detection rate and reduce the false alarm rate compared with traditional EKPCA，KECA and other algorithms.This method solves the problem of selecting kernel parameters for different faults in the traditional KECA algorithm and provides a reference for improving the performance of KECA algorithm in fault detection of nonlinear industrial processes.

Keywords：

other disciplines of automatic control technology;kernel entropy component analysis;Gaussian kernel function;Bayesian decision;ensemble learning method

在流程工业和制造业中，随着现代制造业的快速发展，对高质量产品的需求不断增加，保证系统正常运行成为一项至关重要的任务。虽然运行过程中的标准控制器可以补偿许多类型的干扰，但也存在控制器无法完全处理的变化，将这种特征属性或变量不允许的偏差定义为故障。……