基于等距离映射的非线性动态故障检测方法

Nonlinear Dynamic Fault Detection Method Based on Isometric Mapping

针对化工过程数据强非线性和动态性的特点,提出了一种基于动态等距离映射(Dynamic Isometric Mapping,DISOMAP)流形学习的非线性过程故障检测方法.该方法首先采用DISOMAP算法提取训练样本的子流形特征,自适应学习近邻点参数,保留了采样数据的流形结构,然后运用线性回归方法得到原空间和降维子流形空间的投影映射,从而将观测数据从原高维空间映射到低维嵌入空间,最后在变换后的低维空间构造统计量T2和SPE进行监控.TE过程的仿真结果表明,所提出的DISOMAP故障检测方法可以比核主元分析(Kernel Principle Component Analysis,KPCA)更为有效地监控过程变化,检测到故障的发生.

The data collected from chemical process are strongly nonlinear and dynamic related.To solve this problem,a nonlinear dynamic fault detection method using dynamic isometric mapping（DISOMAP） manifold learning was proposed.It first extracts sub-manifold feature from original data set with adaptive neighbor parameters,which preserves geometric structure.Then linear regression projection mapping which maps the original high dimension space to a low dimension embedding space is used.Finally,T2 and SPE statistics are constructed in the process monitoring application.The simulation results of Tennessee Eastman process show that DISOMAP-based method is more effective than KPCA（kernel principal component analysis） for process monitoring and fault detection.