Detective variables of industrial processes show nonlinear and non-Gaussian behavior. This paper proposes the kernel principal component analysis (WKPCA) based on wavelet kernel clustering to handle the nonlinearity of the process, and introduces support vector data description (SVDD) to model the process. The first step is to construct the kernel function by using Morlet wavelet because of its advantages of multi-resolution analysis and good fitness, which could enhance nonlinear mapping and anti-noise capability of the kernel function. Then this method uses cluster analysis in the feature space, and chooses the data which represent the characteristic center in every cluster, which could decrease calculation load of the kernel function. Finally the method uses the monitor statistics offered by SVDD to describe the non-Gaussian information. Application to the Tennessee-Eastman benchmark process showed effectiveness and accuracy of detecting fault and exception generated by the system.%针对工业过程检测变量具有的非线性和非高斯性等特点,提出了一种基于小波核聚类的核主元分析(WKPCA)方法来处理过程数据的非线性特性,同时引用支持向量数据描述(SVDD)对过程进行建模.本算法先根据Morlet小波具有多分辨分析和能以更高的精度逼近任意函数的特点,将其构建为小波核函数,可以增强KPCA的非线性核映射和抗噪能力,然后在映射后的特征空间中进行均值聚类分析,选择每个聚类中展现特征中心的数据,大大减少了核函数的计算量;最后通过SVDD提出监控指标来描述过程的非高斯特性.将上述方法用在一个标准仿真平台Tennessee-Eastman上,结果表明,该方法能及时有效地检测出系统产生的故障和异常情况.
展开▼
