故障检测中核参数优化方法性能评估
Performance evaluation for kernel parameter optimizations in fault detection methods

近年来, 基于核主元分析与核偏最小二乘的方法经常被应用于过程监控与故障检测领域以克服工业过程的非线性. 研究发现此类方法的检测性能很大程度上受核参数的影响, 而目前学术界对该参数的优化方法研究较少. 因此, 本文以最常用的高斯核方法为例, 首先总结了三类常用的核参数优化方法: 二分法、基于BP神经网络的重构法和基于样本分类的重构法; 其次重点分析每个方法的特点和它们之间的联系, 并评估它们的性能; 最后将上述方法设计成一个核参数优化系统应用于热连轧过程的故障检测中. 应用结果表明, 优化后的核参数能显著提高故障检测性能.
In recent years, kernel principal component analysis and kernel partial least squares-based methods have been widely applied to the process monitoring and fault detection(PM-FD) field to address the nonlinearity in complex industrial processes. It has been found that the performance of these fault detection methods can be significantly affected by the kernel parameter, however, few researches have been focused on this area. Motivated by these observations, on the basis of the commonly used Gaussian kernel, this paper firstly revisits three types of kernel parameter optimization methods, namely, the dichotomy-based, BP neural network reconstruction-based and sample classification reconstruction-based methods. Then, their individual characteristics, interconnections, and performance are analyzed in-depth. Finally, above parameter optimization methods are integrated into a kernel parameter optimization system and applied to the PM-FD in a hot strip rolling process. Compared with selecting parameter empirically, the experiment results shows that the optimized kernel parameter can improve the PM-FD performance