20.500.12556/DKUM-12214 NELINEARNE MULTIVARIATNE STATISTIČNE METODE ZA SPROTNO ODKRIVANJE NAPAK V INDUSTRIJSKIH PROCESIH NON-LINEAR MULTIVARIATE STATISTICAL METHODS FOR ON-LINE FAULT DETECTION IN INDUSTRIAL PROCESSES Disertacija obravnava odkrivanje napak v tehničnih sistemih z uporabo multivariatnih statističnih metod, prilagojenih realnim (nelinearnim) industrijskim procesom. Z vidika napak v tehničnih sistemih, ki so neizbežne in s tem povezanih ekonomskih izgub zaradi izpadov proizvodnje, je za povečanje kakovosti in učinkovitosti proizvodnje potrebno proizvodne procese sprotno nadzorovati in nadgrajevati s sistemi za odkrivanje napak. Le-ti temeljijo na primerjavi delovanja procesov z matematičnimi modeli, pri katerih pa je ključnega pomena točnost modela procesa, saj se tako izogne večjemu številu lažnih alarmov. Prispevki disertacije so osredotočeni na razvoj in optimizacijo sistema za odkrivanje napak, ki vsebuje modele procesa pridobljene na osnovi procesnih podatkov, ter razširitve multivariatne statistične metode glavnih komponent na nelinearne procese. Nelinearna metoda glavnih komponent je realizirana z avto-asociativno strukturo umetne nevronske mreže, kjer nevronski model predstavlja nelinearni model procesa. Delovanje takšnih algoritmov je podrejeno specifičnim pogojem industrijskega okolja in zmogljivosti opreme, zato sta optimizacija delovanja algoritmov in izvedba v realnem okolju bistvenega pomena. Predstavljen je razvoj avto-asociativne umetne nevronske mreže in optimizacija njene strukture in delovanja po Taguchi postopku načrtovanja s poskusi, ki se je v svetu industrijske proizvodnje že večkrat dokazala. Na podlagi sprotnega spremljanja oblike in obnašanja izluščenih nelinearnih glavnih komponent sistema je mogoče sklepati na nepravilnosti v delovanju procesov in potencialne napake ter izvajati preventivne ukrepe v proizvodnji. S kvalitetnim modelom procesa, ustreznim znanjem o procesu in vključenimi pod-modeli napak, se izvaja izolacija ter diagnostika sistemov. Predlagane rešitve so preučene in potrjene na simulacijskem modelu in realnem objektu. Dissertation deals with fault detection in technical systems using multivariate statistical methods adapted to real-time (non-linear) industrial processes. Faults in technical systems are unavoidable and therefore associated with economic loss due to production down-times. To increase quality and efficiency it is necessary to monitor real-time production processes and implement fault detection systems to detect abnormalities. Operation of such system is based on comparison between measured process data and mathematical model, where the model’s accuracy is crucial to avoid a large number of false alarms. Contribution of the dissertation is towards development and optimization of systems for fault detection, including process models based on process data, and extension of multivariate statistical method of principal components to fit nonlinear processes. The process model is presented by nonlinear method of principal components, realized by auto-associative structure of artificial neural network. Implementation of advanced algorithms depends on industrial environment specifics and equipment capabilities, therefore performance optimization is needed. Development and optimization of the auto-associative artificial neural network, its structure and operation is presented, where the artificial neural model is designed by Taguchi design of experiments methodology. FDI on-line monitoring algorithms are processing nonlinear principal components behaviour, extracted from neural network layer, and produce information about process abnormalities, potential faults and serve as precaution measure to avoid process down-time. Furthermore the model quality, a-priory knowledge about the process and implemented sub-models of process faults, are used to achieve fault isolation and process diagnosis. Proposed solutions are tested and confirmed in simulation and on real laboratory test plant. odkrivanje napak izolacija napak residuum multivariatne statistične metode metoda glavnih komponent nevronske mreže načrtovanje s poskusi fault detection fault isolation residual multivariate statistical methods principal components artificial neural networks design of experiments true true false [B. Bratina] Slovenski jezik Angleški jezik Doktorska disertacija 2009-10-15 08:47:02 2009-10-16 08:02:38 2022-04-12 07:18:49 2022-04-19 03:16:45 2009 0 Maribor 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 [519.237:519.71]:658.51(043.3) 13510678 URN:SI:UM:DK:OXCK6LLA DR_Bratina_Bozidar_2009.pdf DR_Bratina_Bozidar_2009.pdf 1 BD439682B0BF95356FB1A35EFBD7AF47 09418cf06c94e117980899733a514fb71fa4a1c70e2b995e46faf8b02768c913 58df720e-7c03-11eb-bb7a-00155d0001ca 20.500.12556/dkum/a5d08849-f573-433f-b768-91bd1a2c775a https://dk.um.si/Dokument.php?lang=slv&id=11436 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0