基于KPCA-SVM的相关和独立变量故障检测方法

沈阳化工大学信息工程学院,辽宁沈阳110142

自动控制技术;核主成分分析;支持向量机;故障检测;相关变量;独立变量

Related and independent variable fault detection method based on KPCA-SVM
GUO Jinyu,YU Huan,and LI Yuan

College of Information Engineering, Shenyang University of Chemical Technology, Shenyang110142, Liaoning Province, P. R. China

automatic control technology; kernel principal component analysis; support vector machine; fault detection; related variables; independent variables

DOI: 10.3724/SP.J.1249.2023.01014

备注

实际工业过程中有一些独立于其他变量的过程变量,为能够分别检测独立变量与相关变量,提出一种基于核主成分分析(kernelprincipalcomponentanalysis,KPCA)和支持向量机(supportvectormachine,SVM)相结合的相关和独立变量故障检测方法,记为KPCA-SVM.首先,应用基于互信息的变量划分策略,通过计算变量之间的互信息将过程变量划分为相关变量和独立变量.然后,在相关变量空间和独立变量空间分别建立KPCA和SVM模型对测试数据进行监测.与传统KPCA和SVM算法相比,KPCA-SVM方法结合了KPCA在检测相关变量和SVM方法在检测独立变量上的优势,提高了KPCA和SVM方法的故障检测性能.最后,将KPCA-SVM方法应用于田纳西-伊斯曼(Tennessee-Eastman,TE)工业过程进行故障检测,并与KPCA、核熵成分分析(kernelentropycomponentanalysis,KECA)和SVM方法进行比较.仿真结果表明,KPCA-SVM方法具有较好的检测效果,对于多种故障的检测效果有所提升,其中对于微小故障5的检测效果有明显提升,进一步验证KPCA-SVM方法的有效性.
In the real industrial process, some process variables are independent of other variables, a fault detection method of related and independent variable based on kernel principal component analysis and support vector machine (KPCA-SVM) is proposed to detect these independent variables separately from related variables. Firstly, a variable division strategy based on mutual information is applied to divide the process variables into related variables and independent variables by calculating the mutual information between variables. Then, KPCA and SVM models are established in the related variable space and the independent variable space to monitor the test data. Compared with the traditional KPCA and SVM methods, the KPCA-SVM method combines the advantages of KPCA in detecting related variables and SVM methods in detecting independent variables, and improves the fault detection performance of KPCA and SVM methods. Finally, the KPCA-SVM method is applied to the Tennessee-Eastman (TE) industrial process for fault detection, and compared with KPCA, kernel entropy component analysis (KECA) and SVM methods. The results show that the proposed KPCA-SVM method has a good detection effect and improves the detection effect of multiple faults, among which the detection effect of minor fault 5 is significantly improved, which further verifies the effectiveness of the KPCA-SVM method.
·