基于MCFS-R-Vine Copula的过程故障检测

Process Fault Detection Based on MCFS-R-Vine Copula

R-Vine Copula模型因其对复杂高维变量相互关系具有的良好的刻画能力,逐渐在化工过程监控领域得到重视。在以往的R-Vine Copula模型应用中,对于数据变量处理采用两种方法,降维和不降维,但这会破坏数据结构或者增加R-Vine Copula模型构建的成本。基于此,提出一种结合多类特征选择(MCFS)的方法MCFS-RVC,来达到保留数据结构以及降低R-Vine Copula模型构建的成本的目的,并将其应用在化工过程监控领域。对R-Vine Copula模型得到的概率密度函数值进行对数转化,依据核密度估计理论对转化的数值进行密度估计形成相应的概率密度指标poutliers,实现了对非线性非高斯过程故障的实时检测。通过对TE(Tennessee Eastman)过程的仿真,验证了所提出方法的有效性和优越性。

The R-Vine Copula model has progressively gained importance in the process monitoring of chemical industry because of its good ability to characterize complex high-dimensional variable interrelationships.In previous applications of R-Vine Copula model,two approaches of data variables processing were used:dimensionality reduction or no dimensionality reduction.However,it could destroy the data structure or increase the cost of R-Vine Copula modeling.Based on these problems,a method called MCFS-RVC,which combines Multi Class Feature Selection(MCFS),is proposed to preserve data structures and reduce the cost of constructing R-Vine Copula models,and it is applied in the field of chemical process monitoring.The values of probability density functions obtained from the R-Vine Copula model were log-transformed,and it produced the corresponding probability density index Pouliers through kernel density estimation theory to achieve a real time nonlinear non-gaussian process fault detection.The monitoring results of the Tennessee Eastman(TE)process shows that the proposed MCFS-RVC approach achieves good performance in chemical process fault monitoring.