Intelligent controllers based on adaptive dynamic programming

[V. S. Đorđević]

VI, 111 listova

Određivanje tačnog modela je veoma izazovan i dugotrajan posao u većini praktičnih slučajeva, što algoritme upravljanja koji su zasnovani na modelu čini neizvodljivim. Čak i ako se može dobiti aproksimirani model sistema, dinamička neodređenost, prouzrokovana neslaganjem aproksimacije i realnog sistema, će degradirati performanse upravljanja klasično projektovanih regulatora. Stoga je projektovanje inteligentnih regulatora vođenih podacima, koji su robusni, adaptivni i pouzdani u prisustvu razčičitih neodređenosti i nepoznate dinamike, od primarnog značaja. U ovoj tezi se najpre predlažu inteligentni regulatori zasnovani na ADP-u linearnih sistema. Kod klasičnih algoritama upravljanja zasnovanih na ADP-u, uobičajeno se pretpostavlja da su promenljive veličine stanja merljive. Međutim, ova pretpostavka je generalno restriktivna, jer u mnogim inženjerskim primenama često nedostaje potpuna informacija o stanju, i samo je izlaz dostupan za merenje. U uslovima nepoznate dinamike sistema i nemerljivih stanja, od velikog interesa je korišćenje tehnika upravljanja zasnovanih na ulazno/izlaznim merenjima, koje se još nazivaju i tehnike sa povratnom spregom po izlazu. Takođe, stabilnost dinamičkih sistema se postiže putem povratne sprege po izlazu, pri čemu se za rekonstrukciju nepoznatih veličina stanja koriste ulazna i izlazna merenja. Da bi se ispunili zahtevi perzistente pobude, potrebno je dodati neki istraživački šum upravljačkom ulazu. Primenom teorije planiranja eksperimenta, u tezi se primenjuje optimalni istraživački šum u vidu sume sinusoida, koji će omogućiti da izlaz sistema nosi maksimum informacije o sistemu. Time će biti skraćeno vreme učenja, odnosno ubrzan proces projektovanja inteligentnih regulatora. Implementacijom strategije upravljanja koji su pokrenuti događajima smanjuje se broj uključivanja predloženih inteligentnih regulatora, čime se štede energetski, računski i komunikacioni resursi. Predložena strategija upravljanja će unaprediti algoritme upravljanja tolerantne na otkaze, čime su poboljšane performanse sistema upravljanja u slučaju pojave otkaza. Praktična primenljivost predloženih inteligentnih regulatora je povećana činjenicom da se nelinearni sistemi mogu verno predstaviti linearnim modelima sa estimacijom dinamike u realnom vremenu. Projektovani regulatori su sposobni da upravljaju linearnim i nelinearnim sistemima sa potpuno nepoznatim matricama sistema, a eksponencijalna konvergencija se postiže dok je uslov perzistentne pobude zadovoljen. Pored teorijskih okvira, praktični aspekt dobijenih rezultata je verifikovan na komleksnim inženjerskim problemima, kao što je upravljanje hidrauličkog servo aktuatora i samobalansirajućeg robotа...

Determining an accurate model is very challenging and time-consuming in most practical cases, making model-based control algorithms infeasible. Even if an approximate model of the system can be obtained, a dynamic uncertainty caused by the mismatch between the approximation and the real system will degrade the control performance of classically designed controllers. Therefore, the design of data-driven intelligent controllers, which are robust, adaptive and reliable in the presence of various uncertainties and unknown dynamics, is of primary importance. This PhD thesis first proposes intelligent controllers based on ADP of linear systems. In classical ADP-based control algorithms, it is commonly assumed that the state variables are measurable. However, this assumption is generally restrictive since, in many engineering applications, complete state information is often lacking, and only the output is available for measurement. In conditions of unknown system dynamics and unmeasurable states, using control techniques based on input/output measurements, also called output feedback techniques, is of great interest. Also, the stability of dynamic systems is achieved through output feedback, whereby input and output measurements are used to reconstruct unknown state variables. To meet the persistent excitation requirements, an exploration noise needs to be added to the control input. By applying the theory of experiment design, this PhD thesis applies optimal exploration noise in the form of a sum of sinusoids, which will allow the system output to carry the maximum amount of information about the system. This will shorten the learning time, i.e. speed up the process of designing intelligent controllers. Implementing an event-triggered control strategy reduces the number of switching of the proposed intelligent controllers, thus saving energy computational and communication resources. The proposed control strategy will improve the fault-tolerant control algorithms, thus improving the performance of the control system in case of faults. The practical application of the proposed intelligent controllers is increased by the fact that linear models with real-time dynamics estimation can faithfully represent nonlinear systems. The designed controllers are capable of controlling linear and nonlinear systems with completely unknown system matrices, and exponential convergence is achieved while the persistent excitation condition is satisfied. In addition to the theoretical framework, the practical aspect of the obtained results was verified on complex engineering problems, such as the control of a hydraulic servo actuator and a self-balancing robot. The validity and effectiveness of the proposed intelligent controllers based on ADP are evaluated through intensive simulations.

inteligentni regulatori, adaptivno dinamičko programiranje, nepoznata dinamika, upravljanje pokrenuto događajima, upravljanje tolerantno na otkaze

intelligent controllers, adaptive dynamic programming, unknown dynamics, event-triggered control, fault-tolerant control

681.513.7(043.3)

Tekst