LEE Uçak ve Uzay Mühendisliği Lisansüstü Programı
Bu topluluk için Kalıcı Uri
Gözat
Konu "Automatic control" ile LEE Uçak ve Uzay Mühendisliği Lisansüstü Programı'a göz atma
Sayfa başına sonuç
Sıralama Seçenekleri

ÖgeA modified anfis system for aerial vehicles control(Lisansüstü Eğitim Enstitüsü, 2022) Öztürk, Muhammet ; Özkol, İbrahim ; 713564 ; Uçak ve Uzay MühendisliğiThis thesis presents fuzzy logic systems (FLS) and their control applications in aerial vehicles. In this context, firstly, type1 fuzzy logic systems and secondly type2 fuzzy logic systems are examined. Adaptive NeuroFuzzy Inference System (ANFIS) training models are examined and new type1 and type2 models are developed and tested. The new approaches are used for control problems as quadrotor control. Fuzzy logic system is a humanly structure that does not define any case precisely as 1 or 0. The Fuzzy logic systems define the case with membership functions. In literature, there are very much fuzzy logic applications as data processing, estimation, control, modeling, etc. Different Fuzzy Inference Systems (FIS) are proposed as Sugeno, Mamdani, Tsukamoto, and ¸Sen. The Sugeno and Mamdani FIS are the most widely used fuzzy logic systems. Mamdani antecedent and consequent parameters are composed of membership functions. Because of that, Mamdani FIS needs a defuzzification step to have a crisp output. Sugeno antecedent parameters are membership functions but consequent parameters are linear or constant and so, the Sugeno FIS does not need a defuzzification step. The Sugeno FIS needs less computational load and it is simpler than Mamdani FIS and so, it is more widely used than Mamdani FIS. Training of Mamdani parameters is more complicated and needs more calculation than Sugeno FIS. The Mamdani ANFIS approaches in the literature are examined and a new Mamdani ANFIS model (MANFIS) is proposed. Training performance of the proposed MANFIS model is tested for a nonlinear function and control performance is tested on a DC motor dynamic. Besides, ¸Sen FIS that was used for estimation of sunshine duration in 1998, is examined. This ¸SEN FIS antecedent and consequent parameters are membership functions as Mamdani FIS and needs to defuzzification step. However, because of the structure of the ¸Sen defuzzification structure, the ¸Sen FIS can be calculated with less computational load, and therefore ¸Sen ANFIS training model has been created. These three approaches are trained on a nonlinear function and used for online control. In this study, the neurofuzzy controller is used as online controller. Neurofuzzy controllers consist of simultaneous operation of two functions named fuzzy logic and ANFIS. The fuzzy logic function is the one that generates the control signal. It generates a control signal according to the controller inputs. The other function is the ANFIS function that trains the parameters of the fuzzy logic function. Neurofuzzy controllers are intelligent controllers, independent of the model, and constantly adapting their parameters. For this reason, these controllers' parameters values are constantly changing according to the changes in the system. There are studies on different neurofuzzy control systems in the literature. Each approach is tested on a DC motor model that is a singleinput and singleoutput system, and the neurofuzzy controllers' advantages and performances are examined. In this way, the approaches in the literature and the approaches added within the scope of the thesis are compared to each other. Selected neurofuzzy controllers are used in quadrotor control. Quadrotors have a twostage controller structure. In the first stage, position control is performed and the position control results are defined as angles. In the second stage, attitude control is performed over the calculated angle values. In this thesis, the neurofuzzy controller is shown to work perfectly well in single layer control structures, i.e., there was not any overshooting, and settling time was very short. But it is seen from quadrotor control results that the neurofuzzy controller can not give the desired performance in the twolayered control structure. Therefore, the feedback error learning control system, in which the fuzzy controller works together with conventional controllers, is examined. Fundamentally, there is an inverse dynamic model parallel to a classical controller in the feedback error learning structure. The inverse dynamic model aims to increase the performance by influencing the classical controller signal. In the literature, there are a lot of papers about the structure of feedback error learning control and there are different proposed approaches. In the structure used in this work, fuzzy logic parameters are trained using ANFIS with error input.The fuzzy logic control signal is obtained as a result of training. The fuzzy logic control signal is added to the conventional controller signal. This study has been tested on models such as DC motor and quadrotor. It is seen that the feedback error learning control with the ANFIS increases the control performances. Antecedent and consequent parameters of type1 fuzzy logic systems consist of certain membership functions. A type2 FLS is proposed to better define the uncertainties, because of that, type2 fuzzy inference membership functions are proposed to include uncertainties. The type2 FLS is operationally difficult because of uncertainties. In order to simplify type2 FLS operations, interval type2 FLS is proposed as a special case of generalized type2 FLS in the literature. Interval type2 membership functions are designed as a twodimensional projection of general type2 membership functions and represent the area between two type1 membership functions. The area between these two type1 membership functions is called Footprint of Uncertainty (FOU). This uncertainty also occurs in the weight values obtained from the antecedent membership functions. Consequent membership functions are also type2 and it is not possible to perform the defuzzification step directly because of uncertainty. Therefore, type reduction methods have been developed to reduce the type2 FLS to the type1 FLS. Type reduction methods try to find the highest and lowest values of the fuzzy logic model. Therefore, a switch point should be determined between the weights obtained from the antecedent membership functions. Type reduction methods find these switch points by iterations and this process causes too much computation, so many different methods have been proposed to minimize this computational load. In 2018, an iterativefree method called Direct Approach (DA) was proposed. This method performs the type reduction process faster than other iterative methods. In the literature, studies such as neural networks and genetic algorithms on the training for parameters of the type2 FLS still continue. These studies are also used in the interval type2 fuzzy logic control systems. There are proposed interval type2 ANFIS structures in literature, but they are not effective because of uncertainties of interval type2 membership functions. FLS parameters for ANFIS training should not contain uncertainties. However, the type2 FLS should inherently contain uncertainty. For this reason, KarnikMendel algorithm is modified, which is one of the typereduction methods, to apply the ANFIS on interval type2 FLS. The modified KarnikMendel algorithm gives the same results as the KarnikMendel algorithm. The modified KarnikMendel algorithm also gives exact parameter values for use in ANFIS. One can notice that the ANFIS training of the interval type2 FLS has been developed successfully and has been used for system control.