Data Driven Monitoring Methods For Complex Continuous Process

As an important part of automatic control system,the process monitoring system is designed for evaluating operation condition of the process,conducting fault detection,fault identification and so on.With the development of modern industrial process,both the scale and complexity of the process increases,bringing higher requirements upon control system and process monitoring system.Most processes are featured with nonlinear,time-varying,dynamic and uncertain characteristics,making some classical multivariate stastistical analysis and machine learning methods less practical and effiective.The thesis focuses on solving three troublesome problems in process monitoring field.1.How to model nonlinear correlation among process variables?2.How to model dynamic character of process variables?How to prevent model from misleading by label noise?Accordingly,the thesis proposes four processing monitoring mthods,which are further subdivided as fault detection and fault classification methods.All four of them are derived from conventional machine learning and deep learning methods,also incorporating some multivariate stastistical analysis models.The brief summaries of these four methods are as follows.(1)As a typical linear process monitoring method,probabilistic principal component analysis(PPCA)is less effective in monitoring nonlinear process.To cope with this problem,a just-in-time-learning based PPCA(JITL-PPCA)method for online process monitoring is proposed in Chapter 2.In JITL-PPCA,a local least squares support vector regression(LSSVR)model is used for extracting nonlinear correlations,which is updated according to an improved JITL algorithm.Then,the residual vector is input into PPCA for final fault detection.LSSVR model is an ideal way of applying kernel functions,since it brings about no optimazation problem like other kernel generilizations of PPCA.A simulated numerical dataset and the TE process benchmark dataset are used to evaluate the performance and effectiveness of the proposed method.The monitoring results show the effectiveness of the proposed real time JITL-PPCA method.(2)Two linear dimensionality reduction algorithms PPCA,PLVR and three nonlinear feature extraction algorithms,GPLVR,GPAM and DAE are analyzed and compared in the beginning of Chapter 3.Based on five algorithms’ strength and weakness,a conclusion is drawn that DAE is the most practical one,because it is easier to optimize DAE than other methods.To further improve DAE,Chapter 3 proposes a contractive denoising autoencoder(CDAE).Different from those process monitoring methods relying on monitoring statistics,Chapter 3 proposes a supervised process monitoring scheme,where features extracted by CDAE are used for training SVM classifier.TE process benchmark dataset is used to test the moniroting performance of the proposed CDAE-SVC method and other methods.The experiment results also prove that CDAE outperforms DAE in extracting more suitable and robust features.(3)To further improve neural networks’ performance in modeling dynamics among process variables,two types of approaches are adopted by researchers.The first type is to add recurrent structures into the network,the second type is to replace the static sample with a extended sequential vector as the model input.Dynamic DAE(D-DAE)is beset with computational complexity problems when the window width of the extended sequential vector is big.Chapter 4 proposes an improved dynamice DAE(CVA-DAE)fault classification algorithm,applying a CVA method to reduce the dimension of extended sequential vectors while retaining dynamic information.Besides,Chapter 4 also proposes a Discriminant DAE(DisDAE),features extracted by which are supposed to be more suitable for classification.TE benchmark process dataset is used to verify fault classification performance of the proposed CVA-DAE and CVA-DisDAE method.The experiment results verify the effectiveness of the proposed CVA-DisDAE based fault classification method.(4)In real industrial processes,label noise imposes greater detriment on modeling than feature noise.To handle label noise problems,Chapter 5 proposes a K-fold cross validation based label noise cleansing algorithm(KCV-LNC).When applied with stable classifiers,KCV-LNC algorithm is supposed to detect mislabeled samples in dataset and correct their label into right ones.Based on KCV-LNC,Chapter 5 proposes a label noise robust stacked DAE(LNC-SDAE)fault classification algorithm.Its robustness upon label noise is derived from KCV-LNC algorithm and dropout strategy in SDAE.Two experiment dataset are generated based on TE process benchmark dataset.The experiment results show KCV-LNC is capable to generate a stable cleansing performance when applied with different classifiers,with the number of mislabeled samples lowered to the absolute minority.LNC-SDAEs trained by corrupted datasets shows similar classification performance with SDAEs trained by original datasets,which proves LNC-SDAE’s robustness upon label noise.