Abstract: Traditional rotation forest based on principal component analysis does not consider time series correlation of features. Therefore, a fault diagnosis method based on kernel canonical rotation forest is proposed to improve fault diagnosis accuracy in nonlinear dynamic processes. For the proposed method, random forest features are projected to the high dimension linear reproducing kernel Hilbert space by using unknown nonlinear mapping. Canonical variate analysis is used to extract dynamic correlation information, and to produce irrelevant features. Kernel function is used to solve the unknown nonlinear mapping problem. In order to avoid kernel matrix singular problems in traditional kernel canonical variate analysis, canonical variables are extracted in kernel principal space, and used to train decision trees. The proposed method takes nonlinear correlation and dynamic correlation of random forest features into account. Meanwhile, the difference between decision trees is increased, which is helpful to improve the accuracy of fault diagnosis. The effectiveness of the proposed method is demonstrated through a case study of the Tennessee Eastman process.

Key words: Fault diagnosis, random forest, rotation forest, canonical variate analysis, kernel principal component analysis

摘要： 基于主元分析的旋转森林没有考虑特征的时序相关性，为此，提出一种基于核典型旋转森林的故障诊断方法，以提高非线性动态过程的故障诊断精度。所提方法通过未知非线性映射将随机森林特征投影到高维线性再生核Hilbert空间，利用典型变量分析提取变量的动态相关信息，得到不相关特征。采用核函数解决非线性映射未知的问题，为了避免传统核典型变量分析存在的核矩阵奇异问题，该研究在核主元空间提取典型变量，以训练决策树。该方法考虑了随机森林特征的非线性相关性和动态相关性，增加了决策树之间的差异性，有助于提高故障诊断精度。以田纳西-伊斯曼过程为例对所提方法的有效性进行了验证。

关键词: 故障诊断, 随机森林, 旋转森林, 典型变量分析, 核主元分析