الفهرس 
Chapter 1: IntroductionOnly 14 pages are availabe for public view
 |  | from | 101 |  |  |
| | | from | 101 | | |
Abstract
Inverted Pendulum Systems (IPSs) are considered one of the most famous laboratory examples used by many researchers in control and robotics fields around the world to prove the validity and effectiveness of several control theories. This is because these systems are non-linear, open loop, and unstable. These characteristics make it the focus of attention and study. One of these systems is the Two-Wheeled Inverted Pendulum System (TWIPS), which is studied in this thesis.
In the first part of the presented thesis, a Newtonian model of the TWIPS is explained and derived in terms of nonlinear differential equations. The state-feedback-control strategy is used with Kalman Filter (KF) to keep the system stable in a short time. A square wave signal is used to excite the system in order to operate at wide range of frequencies to investigate its dynamics. Sufficient sets of real-time input/output data are logged to execute closed-loop experiments of the TWIPS. This data is used to estimate unknown system parameters using the ”grey” box approach with the help of MATLAB parameter estimation toolbox and by applying the Pattern Search Optimization Algorithm (PSOA) method. The PSOA method achieves improved results compared to the trust-region method concerning cost function value and therefore, accurate parameters. These parameters achieve the best acceptable and accurate responses that match real-time system response.
In the second part of the presented thesis, two neural networks controllers are proposed. The first is the Supervised Feed-Forward Neural Network (SFFNN), which is trained offline by using the previously collected input/output datasets. The SFFNN controller is tested with the same number of neurons in the input and output layer but a different number of neurons in the hidden layer. The SFFNN controller with ten neurons in the hidden layer achieves the best response between tested paradigms. The second controller is the Adaptive Linear Neuron (ADALINE) with a state-feedback controller which is trained online in two sessions, one for network initialization and the second for direct angle control.
Finally, a hybrid adaptive neuro-fuzzy controller with a state-feedback control scheme is introduced and expressed. The Adaptive Neuro-Fuzzy Inference System (ANFIS) belongs to the category of a hybrid neuro-fuzzy neural network which combines the advantages of neural networks and fuzzy logic. The ANFIS schatter block (s-function) which belongs to MATLAB Simulink library called “ANFIS” is used to control TWIPS online. A comparative study is carried out between the ANFIS controller and other designed controllers (state-feedback, SFFNN, and ADALINE with state-feedback). Real-time results demonstrated the performant gain of controller compared to others in terms of steady-state error elimination, low overshoot, low settling time, and disturbance rejecting.