Nonlinear Process Monitoring Method Based On Kernel Forecastable Component Analysis

Modern industrial process becomes intelligent and complicated with the development of information technology.To ensure a secure and reliable industrial process,process monitoring method becomes more and more important.Nowadays the method based on data-driven is the focused research direction in the field.The key of the method based on data-driven is to use the massive data in the industrial process more efficiently to monitor industrial processKernel forecastable component analysis(KForeCA)is a new feature extraction method that can process nonlinear data well and minimize the uncertainty of data based on linear transformation in the high dimensional space,where the forecastable principle component subspace and orthogonal white noise subspace are obtained.And the method takes the temporally dependent into account.Thus the features have dynamic temporally dependent and prediction characteristics compared with the classical method.Based on these advantages of KForeCA,this paper applies the method into the nonlinear fault detection and diagnosis field,and the new monitoring statistics based on probability is proposed.Also KForeCA with the SVR is applied to predict the slowly varying fault for its predictable features.These are useful attempts in the process monitoring.The main work of this article specifically includes the following aspects.Firstly,KForeCA is applied to nonlinear fault detection.In order to highlight the useful fault information and the sensitivity of fault detection,the new monitoring statistics based on probability is proposed.And the moving windows theory is used to improve the performance of process monitoring with the recent process history information taken account into the current process information.Secondly,a new method based on Fisher discriminant KForeCA is proposed for fault diagnosis.In order to reduce the influence of redundant information and the error classification,the forecastable component is projected to the optimal classification direction and classifies the fault based on the minimum distance of discriminant.Thirdly,based on the predictable ability of the forecastsable component,KForeCA with SVR is applied to predict the slowly varying fault to solve the problem that slowly varying fault is hard to detection,which can avoid the losses caused by such fault effectively and assure the security and reliability of industrial process.Finally,process monitoring method based on KForeCA is applied on the Antarctic Taishan Station Wind-power-Solar data monitoring system.By monitoring the battery performance,the fault battery can be detected and which can improve the security and reliability of the power system.