SAMORAZVIJAJOČI SE SISTEMI PRI SPREMLJANJU IN VODENJU PROCESOV

Andonovski, Goran (author), Škrjanc, Igor (mentor), Klančar, Gregor (comentor)

V pričujoči doktorski disertaciji smo raziskovali različna področja uporabe samorazvijajočih se sistemov, kot so vodenje, identikacija in spremljanje procesov. Samorazvijajoči se sistemi, ki jih obravnavamo v disertaciji, temeljijo na posplošenem mehkem modelu AnYa (po priimkih avtorjev Angelov in Yager), ki se razlikuje od klasičnih mehkih modelov (Mamdani in Takagi-Sugeno) v načinu deniranja strukture modela. Namesto vnaprej deniranih mehkih pravil (Gaussovo, trikotno, lingvistično itn.), model AnYa na podlagi sprotno sprejetih podatkov tvori mehka pravila v obliki oblakov podatkov (ang. data clouds). Z vsakim novim prejetim podatkom se struktura in parametri modela prilagajajo novim spremenjenim stanjem v procesu. To nam omogoča implementacijo različnih algoritmov sprotnega vodenja, identikacije ali spremljanja dinamičnih procesov. Poleg implementacije teh algoritmov smo v doktorski disertaciji obravnavali tudi mehanizme samorazvijanja modelov za dodajanje novih in odstranjevanje nepomembnih oblakov. Prav tako smo iskali načine, kako preprečiti dodajanje oblakov na osnovi osamelcev. V disertaciji smo predstavili robusten samorazvijajoči se adaptivni mehki regulator (ang. robust evolving cloud-based controller, RECCo). Regulator je sestavljen iz dveh glavnih delov: samorazvijajoča se struktura modela (z mehanizmom za dodajanje novih oblakov na podlagi lokalne gostote podatkov) in sprotna adaptacija parametrov lokalnih regulatorjev (na podlagi gradienta kriterijske funkcije). Mehanizem samorazvijanja modela skrbi za zaznavanje nelinearnih področij v procesu, kar pomeni, da se parametri lokalnih regulatorjev prilagajajo delovni točki procesa. Z normiranjem podatkovnega prostora smo dosegli enostavnejše nastavljanje začetnih parametrov regulatorja. Podali smo tudi smernice, kako nastaviti oziroma izračunati začetne vrednosti parametrov regulatorja. V doktorski disertaciji smo prikazali nekaj primerov uporabe RECCo-regulatorja na simuliranih in realnih napravah. Vodenje na simuliranih procesih smo izvedli na modelu toplotnega izmenjevalnika in na modelu distribuiranega sistema sončnih kolektorjev. Uporaba vodenja na realnih napravah pa je bila izvedena pri regulaciji temperature na toplotnem izmenjevalniku in regulaciji nivoja na sistemu dveh povezanih tankov. Mehanizem samorazvijanja na osnovi oblakov podatkov smo vpeljali tudi v mehki prediktivno funkcijski regulator (ang. fuzzy cloud-based predictive func- tional controller, FCPFC). Za delovanje tega regulatorja potrebujemo model procesa, ki ga želimo voditi. S tem namenom smo združili samorazvijajoči se model z rekurzivno metodo najmanjših kvadratov. Na ta način lahko identiciramo dinamičen model procesa, ki je potem del prediktivnega regulatorja. Model uporabimo za predikcijo reguliranega signala na vnaprej določenem horizontu in nato določimo še regulirni signal, ki minimizira razliko med izhodnim in referen čnim/želenim signalom. Taksen pristop je primeren za regulacijo nelinearnih dinamičnih procesov. Delovanje predlaganega prediktivnega regulatorja FCPFC smo preizkusili na modelu reaktorja z neprekinjenim mešanjem (ang. continu- ous stirred tank reaktor, CSTR). Dobljene rezultate smo primerjali z RECCoregulatorjem. V nadaljevanju smo predlagali in preizkusili samorazvijajoči se model na osnovi oblakov za identikacijo dinamičnih sistemov. V tem primeru smo raziskali različne mehanizme dodajanja in odstranjevanja oblakov in njihov vpliv na učinkovitost celotne metode. Predlagano metodo smo preizkusili na dveh različnih primerih. Prvi primer je model kemičnega procesa Tennessee Eastman, ki ima zelo kompleksno strukturo in dinamiko. Iz tega modela smo pridobili simulirane podatke ter poskušali pridobiti modele kazalnikov proizvodnje učinkovitosti. Rezultate smo primerjali z metodo eFuMo ter z nevronskimi mrežami. Drugi primer je bil realen sistem hladilne postaje, ki obratuje v enem od podjetij v Sloveniji. Pridobljene podatke smo prav tako uporabili za identi- kacijo dinamičnih modelov nekaj ključnih kazalnikov proizvodnje. Te modele smo naknadno uporabili za nadzorovano in prediktivno preklapljanje hladilnih agregatov, ki so ključni elementi celotnega sistema. Izkazalo se je, da z uporabo modelov lahko preprečimo nepotrebna preklapljanja agregatov in s tem omogočimo boljšo učinkovitost celotnega sistema. V zadnjem delu smo samorazvijajoči se model uporabili kot orodje za spremljanje procesov. Z uporabo mehanizmov za dodajanje novih oblakov lahko deniramo področje procesa, ki opisuje normalno stanje delovanja, in področje, ki označuje napako na procesu. Nato lahko z izračunom lokalnih gostot za vsak podatek posebej določimo ali predstavlja napako oziroma normalno delovanje. Predlagali smo tudi izračun delnih lokalnih gostot z upoštevanjem najbolj vplivnih komponent. Delovanje metode smo preizkusili na področju zaznavanja napak na sistemu Tennessee Eastman. Rezultate smo primerjali z nekaj znanimi metodami za zaznavanje napak na procesih, kot so PCA (ang. principal compo- nent analysis), ICA (ang. independent component analysis) in FDA (ang. sher discriminate analysis). Rezultati metode so primerljivi in dosegajo podobno natan čnost, kot že uveljavljene metode na tem področju. Na koncu smo samorazvijajoči se model razdelili na več hierarhičnih nivojev z namenom zaznavanja manevrov pri voznikih osebnih avtomobilov, kot so prehitevanje, zaviranje, ustavljanje in podobno. Metoda uporablja le osnovne senzorje (in ne naprednih, kot so kamere, laserji itd.), ki so del standardne opreme osebnih avtomobilov. Izkazalo se je, da predlagani hierarhični koncept od spodaj navzgor (od manj do bolj kompleksnih akcij) v kombinaciji s samorazvijajočim se modelom uspešno zaznava in ločuje med različnimi manevri pri voznikih.

• Language: Slovenian
• Keywords: model AnYa, samorazvijajoči se adaptivni mehki regulator
• Work type: Doctoral dissertation
• Organization: FE - Faculty of Electrical Engineering
• Year: 2018
• PID: 20.500.12556/RUL-105879
• Publication date in RUL: 21.12.2018

Secondary language

• Language: English
• Title: EVOLVING SYSTEMS IN PROCESS MONITORING AND CONTROL

Abstract

In the doctoral thesis the usage of evolving systems on dierent elds was investigated, namely the eld of control, system identication and process monitoring. The evolving systems in the current thesis are based on the simplied fuzzy model AnYa which diers from the classical ones (such as Mamdani and Takagi-Sugeno) in the way how the model structure is dened. The AnYa model uses data clouds in the antecedent part of the fuzzy system instead of predened fuzzy rules (Gaussian, triangular, trapezoid etc.) in the classical models. Therefore, the structure of the AnYa could be adapted with each data point received in online manner. This leads to easy implementation of advanced on-line methods for control, identication or monitoring of dynamic processes. Beside this we have investigated dierent evolving mechanisms for adding new rules/clouds and for removing less important ones. Firstly we have presented a robust evolving cloud-based controller RECCo for nonlinear processes. The controller consists of two main parts: evolving mechanism (adding new clouds according to the local density of the data) and adaptive law of the local controllers' parameters (based on gradient descent method). The evolving part takes care of the controlled process nonlinearity and the adaptive law adjusts to the current operating point. We proposed a normalization of the data space which simplies the process of setting the initial parameters. Also some instructions for setting/calculating the initial parameters are given. In the dissertation we show some practical examples of using the controller on simulated and real processes. For the simulated examples we used a model of heat exchanger and a model of distributed solar collector eld. Moreover a real plant of heat exchanger and two tank plant were used for temperature and level control, respectively. Based on the evolving model we also introduce a fuzzy cloud-based predictive functional controller (FCPFC). This controller requires a model of the controlled process. With this purpose we joined the evolving mechanisms with the recursive weighted least square method. Actually we developed a tool for identication of dynamic process model. The model is further used for the controlled signal prediction in a certain horizon and then the control signal is chosen to minimize the error between the predicted and the desired value of the signal. This approach is suitable for controlling nonlinear dynamic processes. The eciency of the proposed controller FCPFC was tested on Continuous Stirred Tank Reactor, CSTR. The results were also compared with the RECCo controller. In the second part of the doctoral thesis we proposed an evolving method for dynamic process identication. Actually the recursive identication method was already tackled in the previous part when FCPFC controller was proposed. In this part we additionally investigated dierent mechanisms for adding and removing data clouds (fuzzy rules) and their impact on the overall identi cation method eciency. The proposed method was tested on two dierent examples. The rst example was Tennessee Eastman process which has really complex structure with dynamic behavior. We acquired the necessary data from the model and then we identied the models of the production Performance Indicators (pPI). The obtained results were compared to the established methods eFuMo and neural networks. The second example was a real process of water chiller plant (WCP). The data were acquired directly from the plant and were used for model identication of the key indicators. The models were further used for process monitoring and predictive operating with the chillers, which are the key elements of the whole system. It has been shown that using the acquired models we can prevent unnecessary switching of the chillers which leads to more ecient operation of the system. In the last part of the thesis we investigated the usage of the evolving system for process monitoring and fault detection purposes. Using the evolving model we can dene the part of data space which describe normal process operation and faults. According to the local density we can easily determine if each data sample is a fault or not. Moreover we proposed a partial data density calculation which takes into account only the most in uential components. The eciency of the method was tested on the Tennessee Eastman process. The results were compared with well established methods on the eld such as: PCA (principal component analysis), ICA (independent component analysis) and FDA (sher discriminate analysis). With the comparison analysis of the results we show that the proposed evolving method is comparable with the established ones. At the end we developed a hierarchical evolving model for car-driver behavior detection. The method uses the basic sensors (no cameras, radars etc.) in the car which are part of standard equipment of a modern car. We show that with the proposed hierarchical concept and the evolving model we can eciently detect dierent maneuvers (such as overtaking, stopping, breaking, etc.) and can dierentiate between them.

• Keywords: AnYa model, robust evolving cloud-based controller