Abstract: Aiming at a large number of nonlinear problems in the chemical process, the main existing method is the combination of the kernel algorithm and the partial least squares (KPLS) algorithm. Compared with KPLS algorithm, the method of combining kernel algorithm and canonical correlation analysis (KCCA) algorithm can maximize the correlation between the two groups of variables to achieve better detection results. However, the current KCCA method cannot accurately decompose the data space into parts that are related and unrelated to the key performance indicator (KPI), thereby it ignores the fact that the remaining space still involves some information related to the KPI. In this paper, an improved KCCA is proposed. The method performs singular value decomposition (SVD) on the calculable loadings of kernel matrix, a projection model is obtained in which the kernel matrix is appropriately decomposed into KPI -related and -unrelated parts, and then two statistics are accordingly designed for fault detection. Finally, the Eastman Eastman (TE) process was used to verify the effectiveness and superiority of the proposed method.

Key words: Canonical Correlation Analysis（CCA）, Fault detection, Nonlinear, Key performance indicator (KPI)

摘要： 针对化工过程中存在大量非线性问题，目前存在的主要方法为将核算法与偏最小二乘算法相结合（KPLS），相比于KPLS算法，将核与典型相关分析相结合的方法（KCCA）能够最大化两组变量相关性，以达到更好的检测效果。然而KCCA方法不能准确地将数据空间分解成与关键性能指标（KPI）相关和不相关的部分，从而忽略了剩余空间仍然涉及与KPI相关的一些信息的事实。文中提出了一种新的改进的KCCA（MKCCA）方法，该方法对核矩阵的可计算负载进行奇异值分解（SVD），得到一个投影模型，将核矩阵适当地分解为KPI相关部分和不相关部分，然后设计两个统计量进行故障检测。最后利用田纳西伊士曼（TE）过程验证了所提出方法的有效性和优越性。

关键词: 典型相关分析（CCA）, 故障检测, 非线性, 关键性能指标（KPI）