Research On Nonlinear Process Monitoring Method Based On Principal Component Analysis

For modern industrial processes,the production procedure is complex,and the environment is harsh.Once a failure occurs,it will affect the smooth operation of the whole process,thereby reducing the product performance,causing property damage and even leading to unforeseen catastrophic accidents.To ensure process safety and reliability,the importance of process monitoring systems is particularly obvious.By establishing the fault detection and diagnosis model,the monitoring system can monitor the entire production process operation status in real-time,find and locate fault points in time,then reducing unplanned downtime and ensuring product quality.In the traditional process monitoring methods,multivariate statistical process monitoring(MSPM)method utilizes the historical data to describe process behaviors,thus detecting abnormal situations.Moreover,its theoretical system is relatively complete,therefore attracting more and more attention in industry and academia.Principal component analysis(PCA)is one of the most popular MSPM methods and has been widely used in industrial process monitoring.However,the actual industrial process has very complicated operational characteristics,making it difficult for the traditional PCA method to accurately describe its behavioral characteristics,resulting in a large number of missed detection rates and false alarm rates.To improve the process monitoring performance,many scholars have proposed improved methods based on PCA.Based on the existing work,this paper mainly solves the following problems:1)Among various nonlinear PCA methods,the kernel-based method subtly extracts the nonlinear structure of the data through the kernel function and has gradually become a research hotspot in nonlinear process monitoring.However,in the kernel PCA(KPCA)method,kernel principal components(KPCs)are linear combinations of all kernel functions,resulting in poor interpretability.To this end,this paper proposes a sparse KPCA via sequential approach(SSKPCA).The method induces KPCs sparseness through the elastic net regularization penalty term and solves the established optimization objective through a sequential algorithm,thus obtaining the sparse representation.The case study on the Tennessee Eastman(TE)process shows that the SSKPCA method has better fault detection and identification results than other methods.2)The traditional KPCA method needs to map the test samples into the high-dimensional(or infinite-dimensional)feature space,which will lead to large computational and poor real-time performance.This paper proposes a nonlinear process monitoring method based on local-global randomized PCA(LGRPCA).Through nonlinear function,the proposed method maps the data in the original input space to the random Fourier features(RFF)space.Compared with the kernel-induced high-dimensional feature space,this spatial dimension is smaller,which can greatly reduce the amount of calculation.Using local and global information,the LGRPCA method better represents the manifold structure in the data.Through numerical example and TE process,it can be found that this method not only reduces the computational and storage costs,but also the monitoring performance is better.3)The actual industrial process has complex characteristics such as nonlinearity and dynamics,however,the existing nonlinear PCA and dynamic PCA methods only focus on the single characteristic in the process.Therefore,this paper proposes a multi-feature extraction technique based on PCA(MFPCA)for nonlinear dynamic process monitoring.Using the feature extraction form of serial structure,this method extracts the dynamic,linear and nonlinear characteristics of the data sequentially by dynamic-inner PCA(DiPCA),PCA and KPCA modeling methods.Through numerical example and TE process,the proposed MFPCA method has a better monitoring performance for nonlinear dynamic processes.