Fault identification utilizing hybrid modelling based feature extraction models

Date
2022-04
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
Fault detection and identification models are critical in process monitoring and control as the models are essential in maintaining normal operating conditions. Fault identification models identifies the types of fault which occur in a process and the cause of the fault which allows corrective measures to be applied. Many fault identification models operate by identifying a process fault once a fault detection model has detected the presence of a fault. Fault identification is posed as a multiclass classification problem with each class corresponding to a fault case with a normal operation class introduced to account for the fault detection aspect. A one-vs-one multiclass support vector machine (SVM) classifier is proposed as the fault identification model. A model parameter estimation method was proposed to improve the performance of the fault identification model. The parameter estimation behaves as a feature extraction method. Hybrid modelling is identified as a possible model parameter estimation method. Hybrid modelling combines first-principle models and data-based models. The data-based models are trained to estimate the model parameters based on incoming process data. The data-based models considered are partial least squares regression (PLS), dynamic PLS, and recursive PLS models. A non-isothermal jacketed continuous stirred tank reactor model is developed as a test case model with a catalyst deactivation fault, inlet concentration fault and heat transfer fault applied to the model. The fault identification models are trained using process data corresponding to a catalyst deactivation-inlet concentration fault pair and catalyst deactivation–heat transfer fault pair. The performance of the fault identification models is compared using the sensitivity and specificity measures. The performance of fault identification models using a standard SVM and kernel SVM with a radial basis function kernel were compared. The kernel SVM showed similar performance to the SVM for the catalyst deactivation fault and heat transfer fault with sensitivity values of 0.684±0.044 and 0.752±0.067, and a shorter training time than the standard SVM model. When the performance of the classifiers incorporating non-linearly regressed model parameters were evaluated by identifying the catalyst deactivation fault and heat transfer fault it was found that the standard SVM model using the regressed parameters had higher sensitivities (0.686±0.042, 0.811±0.031) and specificities (0.989±0.005, 0.968±0.027) than the kernel SVM with sensitivities (0.633±0.058, 0.734±0.033) and specificities (0.974±0.005, 0.924±0.038) using the regressed parameters. When the performance of the hybrid fault identification models was evaluated, the standard SVM using dynamic PLS showed better performance than the other models with higher sensitivities (0.695±0.041, 0.761±0.056) and specificities (0.9800±0.004, 0.949±0.049). When the performance of all the models were compared it was found that the standard SVM using non-linearly regressed parameters was the best performing model. The multiclass SVM approach has been shown a viable fault identification method. Implementing the model-based feature extraction method was shown to improve the performance of fault identification models. The SVM model using non-linearly regressed parameter estimates was found to be the best performing model. It is recommended that in future work the PLS models are replaced with another data-based model such as artificial neural networks.
Foutopsporing- en -identifikasiemodelle is krities in prosesmonitering en -beheer omdat die modelle essensieel is vir handhawing van normale bedryfskondisies. Foutidentifikasiemodelle identifiseer die tipe fout wat plaasvind in ’n proses en die oorsaak van die fout, wat die korrekte maatreëls toelaat om toegepas te word. Baie foutidentifikasiemodelle word bedryf deur ’n prosesfout te identifiseer sodra ’n foutopsporingsmodel die teenwoordigheid van ’n fout opgespoor het. Foutidentifisering word voorgestel as ’n multiklas-klassifiseringsprobleem met elke klas wat korrespondeer met ’n foutgeval, met ’n normale bedryfklas voorgestel om die foutopsporingsaspek in berekening te bring. ’n Een-vs.-een multiklas ondersteuningvektormasjien (SVM) klassifiseerder word voorgestel as die foutidentifikasiemodel. ’n Modelparameterberamingsmetode is voorgestel om die doeltreffendheid van die foutidentifikasiemodel te verbeter. Die parameterberaming tree op as ’n kenmerkekstraksiemetode. Hibried modellering is geïdentifiseer as ’n moontlike modelparameterberamingsmetode. Hibried modellering kombineer eerste-benaderingmodelle en datagebaseerde modelle. Die datagebaseerde modelle word opgelei om die modelparameters te beraam gebaseer op inkomende prosesdata. Die datagebaseerde modelle oorweeg is gedeeltelike kleinste kwadraatregressie (PLS), dinamiese gedeeltelike kleinste kwadraat-, en rekursiewe gedeeltelike kleinste kwadraatmodelle. ’n Nie-isotermiese ommantelde kontinue geroerde reaktormodel is ontwikkel as ’n toetsgevalmodel met ’n katalisator deaktiveringsfout, inlaatkonsentrasiefout en hitte-oordragfout wat toegepas is op die model. Die foutidentifikasiemodelle is opgelei deur prosesdata wat met ’n katalisator deaktiveringsinlaatkonsentrasiefoutpaar en katalisator deaktiveringshitte-oordragfoutpaar korrespondeer, te gebruik. Die doeltreffendheid van die foutidentifikasiemodelle was vergelyk deur die sensitiwiteit- en spesifisiteitmaatreëls te gebruik. Die doeltreffendheid van foutidentifikasiemodelle is vergelyk deur ’n standaard SVM en kern SVM met ’n radiale basiesfunksie te gebruik. Die kern SVM het soortgelyke doeltreffendheid getoon as die SVM vir die katalisator deaktiveringsfout en hitte-oordragsfout met sensitiwiteitswaardes van 0.684±0.044 en 0.752±0.067, en ’n korter opleidingstyd as die standaard SVM-model. Toe die doeltreffendheid van die klassifiseerders wat nie-liniêre regressie modelparameters inkorporeer geëvalueer is deur die katalisator deaktiveringsfout en hitte-oordragsfout te identifiseer, is dit gevind dat die standaard SVM-model wat die parameters gebruik, hoër sensitiwiteite (0.686±0.042, 0.811±0.031) en spesifisiteit het (0.989±0.005, 0.968±0.027) as die kern SVM met sensitiwiteite (0.633±0.058, 0.734±0.033) en spesifisiteit (0.974±0.005, 0.924±0.038) wat die parameters gebruik. Toe die doeltreffendheid van die hibriede foutidentifiseringmodelle geëvalueer is, het die standaard SVM wat dinamiese PLS gebruik beter doeltreffendheid getoon as die ander modelle met hoër sensitiwiteite (0.695±0.041, 0.761±0.056) en spesifisiteit (0.9800±0.004, 0.949±0.049). Toe die doeltreffendheid van al die modelle vergelyk is, is dit gevind dat die standaard SVM wat nie-liniêre regressie parameters gebruik, die beste presteer het. Die multiklas SVM-benadering is getoon as ’n uitvoerbare foutidentifikasiemetode. Implementering van die model-gebaseerde kenmerkekstraksiemetode is bewys om die doeltreffendheid van foutidentifikasiemodelle te verbeter. Die SVM-model wat nie-liniêre regressie parameter beraminge gebruik het, is gevind om die beste presterende model te wees. Dit word aanbeveel dat toekomstige werk die PLS-modelle vervang met ’n ander datagebaseerde model soos kunsmatige neurale netwerke.
Description
Thesis (MEng)--Stellenbosch University, 2022.
Keywords
Support vector machines, Fault location (Engineering), Feature extraction, Hybrid modelling, Partial least squares, UCTD
Citation