基于改进KFDA 和RWn -SVM 的化工过程故障快速诊断

王斌 工程师 匡蕾 高级工程师

(江苏省安全生产科学研究院 化工安全与信息研究所，江苏 南京 210009 )

学科分类与代码：6203070(安全系统工程) 中图分类号：X928.03 文献标志码：A

基金项目：江苏省自然科学基金(SBK201222273) 。

【摘要】 为提高复杂化工过程故障在线诊断水平，提出一种基于改进的核费舍主元分析

(KFDA)和鲁棒损失小波n -支持向量机(RWn -SVM)集成的故障快速诊断方法。首先利用近

邻边界法对 KFDA 进行监督降维，扩大样本数据最近异类点和最远同类点的边界，保证了

样本投影后不同类别的数据不发生重叠。其次将参数b 纳入支持向量机最优化问题的原问题

求解中，使回归估计函数表达式中不含参数b ，从而简化了计算。设计的鲁棒损失函数和小

波核函数能有效压制复杂数据样本中多种噪音和奇异点，具有很强的鲁棒性和实用性。最后

对经典的Tennessee Eastman Process (TEP)化工过程进行实例分析，结果表明改进方法的故

障诊断速度和正确率均好于普通方法，故障诊断结果能够反映化工过程的实际情况。

【关键词】 化工过程；核主元分析；支持向量机；快速故障诊断

*本文已被《中国安全科学学报》录用 2013 年23 卷第7 或8 期)

A Fast Fault Diagnosis Method of Chemical Process Based on Improved Kernel

Fisher Discriminant Analysis and Robust Wavelet n Support Vector Machine

WANG Bin KUANG Lei

(Institute of Chemical Safety & Information Technology, Jiangsu Academy of Safety

Science and Technology, Nanji ng Jiangsu 210009, China)

Abstract: To improve the on-line fault diagnosis level of complex chemical process, an

improved fast fault diagnosis method based on improved kernel Fisher discriminant analysis and

robust wavelet n -support vector machine was proposed. Firstly, by using the neighbor boundary

method for supervised dimensionality of KFDA, it can help to expand the boundary between

nearest heterogeneous point and farthest congeneric point of sample, which to ensure the sample

proj ection data of different classes do not overlap. Secondly, the final expression of regression

estimation function does not contain the parameter b by bring b into the solution procedure of

the original problem of support vector machine optimization problems, which simplifies the

calculation. A robust loss function and the wavelet kernel function were designed to suppress

various noise and singular points of complex data sample