Bachelor of Science (B.Sc.)

Istanbul, 2000-2004

University:

Istanbul Technical University

Faculty / Department:

Electrical and Electronics Engineering Faculty / Electrical Engineering Department

GPA:

3.39 (third-highest rank in the department)

Thesis Title:

Phase Lead, Phase Lag and Phase Lead-Lag Types Controller Design in the Time-Domain by Using Root Locus Technique

Thesis Summary:

PD, PI and PID (Proportional-Integral-Derivative) type controllers are simple, yet versatile, feedback compensator structures widely used in industrial control systems. They are universally accepted in industry due to the model-independency and wide applicability. In this thesis, generalized variants of them, namely Phase Lead, Phase Lag and Phase Lead-Lag type controllers are studied and the design procedures are investigated.

In introductory sections, a detailed classification of the open-loop/closed-loop, continuous/discrete, linear/nonlinear, time-invariant/time-varying, instantaneous (memoryless)/dynamic (with memory), deterministic/stochastic, stable/unstable, causal/non-causal systems is given. Time domain criterions are familiarized by using first-order, second-order and higher-order system responses. The concept of the dominant pole is explained in terms of transient and steady state responses. The Root Locus technique is examined.

In the design section, the parameters of Phase Lead, Phase Lag and Phase Lead-Lag type controllers, which stabilize the various higher order unstable systems, are calculated by using Root Locus technique.

Master of Science (M.Sc.)

Istanbul, 2004-2006

University:

Istanbul Technical University

Institute:

Graduate School of Science, Engineering and Technology

Faculty / Department:

Electrical and Electronics Engineering Faculty / Control and Automation Engineering Department

GPA:

3.80

Thesis Title:

System Identification with Artificial Neural Networks

Thesis Summary:

In the age of supercomputers, as machines are humanized, the metals and plastics are evolving into the bones and skin, softwares are evolving into the nervous systems. Eventually, many of capabilities that human have, can be mimicked by machines, as is learning. The subject of this thesis, that is artificial neural networks, can be interpreted as a subsection of artificial intelligence and closely related to the machine learning. So, first and foremost, artificial intelligence approaches and related learning strategies are introduced. A short history is given and the importance of the logic XOR function is clarified with regards to the historical development process of neural networks.

Following a brief introduction, Perceptron and Adaline models, the simplest and smallest unit of an artificial neural network, are studied. Linear mapping and classification performances of these two neurons are compared. The differences between their learning rules, Hebb and Delta methods, are revealed. A single layer neural network is obtained by combining multiple Perceptron models. Decision surface is defined and its geometrical interpretation is represented for a special situation. A multilayer neural network is constituted by joining multiple single layer networks. Backpropogation algorithm and update formulas are derived by using generalized delta rule.

After derivation of the mathematical equations and implementation of the iterative codes, an input-output model is obtained for a linear time-delay system. Besides that, two other input-output models are developed by using parametric linear auto-regressive with exogenous input (ARX) and non-parametric non-linear fuzzy logic approaches for comparison purposes. The simulation results are compared in terms of modeling errors. For reducing the error to acceptable values obtained by using linear ARX approach, the order of the transfer function has to be chosen very high, which results in numeric instabilities. On the other hand, if the system order is known, then the fuzzy logic approach leads to smallest modeling error. However if the system order is not known, which is a situation that can often be encountered in real applications, while the fuzzy logic approach is not applicable, still satisfactory results can be achieved by using artificial neural network model.

Doctor of Philosophy (Ph.D.)

Ankara, 2007-2014

University:

Middle East Technical University

Institute:

Graduate School of Natural and Applied Sciences

Faculty / Department:

Engineering Faculty / Electrical and Electronics Engineering Department

Major-Minor Fields:

Control and Robotics Divisions

GPA:

3.25

Thesis Title:

Modeling and Real-Time Control of the Three Degrees-of-Freedom Parallel Manipulated Robotic Sensor Head

Thesis Summary:

From small scale on-board cameras that placed on the head of the humanoid robots to large scale military purposed cameras placed on unmanned aerial vehicles, vision sensors everywhere. Undoubtedly, each application has its own challenges, but working with cameras may be more troublesome and effortful than any other sensor. Even in a fully controlled indoor application with exactly known lighting conditions, small differences in illumination can cause problems. Moreover, equipments may become inoperable at unmanageable environmental conditions such as in glistening situation. Besides unfavorable lighting originated problems, vision sensors are considerably sensitive to vibrations and impacts. So, process may become even more complicated depending on whether the camera and/or the object of interest is stationary or not. Additionally, mobility requirement at high speed is another factor that fuzzifies the features and reduces the information that is contained in the image.

At our laboratory, an experimental study conducted on a fixed-mounted camera to the highly mobile, dynamically stable six-legged robot revealed that the performance of the feature detectors (i.e., Canny edge detection and Harris corner detection algorithms) degrades with increasing platform velocities. Obtained results showed that at various normalized velocity values which are measured as 0.1, 0.4, 0.8 and 1.0, respectively, every 2, 16, 28 and 60 of 100 frames can not be included to feature detection algorithms due to the intolerable motion blur effect caused by visual disturbances. Even though deblurring methods are examined by many researchers and studied at various fields of engineering, the performance of these methods is limited and results may not satisfy the requirements of all kind of applications. Particularly, the proposed methods are not suitable for images taken from a camera placed on a mobile robot platform moving in an absolutely uncontrollable environment. At that point, an important question arises. Is it possible to degrade the effects of visual disturbances during capturing process, instead of after capturing process? Fortunately, this question leads us to the stabilization problem which is the subject of this thesis. Minimizing the effects of disturbances on-the-fly is considered to be a more robust approach than deblurring the motion imposed images by postprocessing after they are captured. This prescience is the basis of our motivation source for studies about the camera stabilization mechanism to be used on our six-legged robot platform.

It is possible to examine this mechanism in two parts; mechanical and electrical setups. Mechanical setup, which is inspired by well-known Stewart Platform, can be roughly described as a closed-chain parallel manipulator. The translational degrees-of-freedom are excluded from the 6-DOF Stewart Platform by removing the three actuators, that leads to the reduced 3-DOF structure with rotational degrees-of-freedom only. All three links connecting the paralel plates to each others, include a linear actuator, two universial (or spherical) joints and other auxiliary structural connectors. On the other side, electrical setup is composed of three-phase brushless linear DC servomotors and their drivers for actuating the top plate, Hall effect sensors for measuring the shaft positions, inertial measurement units for measuring the plate orientations, camera for capturing the images, a data acquisition card for gathering signals from measurement sources and a PC104 form factor based series processing hardware for running a real-time operating system that calculates all the algorithms.

Hardware-in-the-loop design and optimization strategy is not an option due to the constraints imposed by the experimental setup. Therefore, to choose the controller structure, as well as to design and optimize it's parameters, a realistic simulation model is derived that takes into account both mechanical and electrical parts. Derivation of the dynamic equations for the mechanical part is beyond the scope of this thesis, so a multi-body modeling tool is used for this purpose. A precise mathematical model that considers the linear motor, motor driver, Hall effect sensor, velocity profiler is constituted for electrical part. As a last step of the modeling works, a virtual camera model is obtained not only to calculate a motion blur related performance metric that will be used for comparison, but also to generate synthetic images that will be used for visualization. By using this virtual camera model that takes into consideration the intrinsic & extrinsic parameters, as well as the distortion (lens imperfectness) model, it is possible to directly compute the numerical value of the motion blur for a predefined scene according to the position and velocity of the camera.

The overall nonlinear simulation model that is obtained by combining all these submodels is used to design various controller structures including model-free Proportional-Integral-Derivative Controller (PID), Sliding Mode Controller (SMC), Proxy-Based Sliding Mode Controller (PBSMC), Fuzzy-Logic Controller (FLC), Artificial Neural Network Controller (ANNC), as well as model-based Linear Quadratic Tracker (LQT), Model Predictive Controller (MPC). Controller parameters are optimized to allow both overdamped and underdamped operation modes. The simulation results are compared in terms of time domain specifications (i.e., rise time, maximum overshoot), as well as vision system degradation metrics (i.e., motion blur).

It is recognized that the source of motion blur is more closely related to the error in velocity reference rather than position reference. For this reason, overdamped operation modes which leads to smoother variations in velocity states give smaller motion blur values. In light of these results, the best controller structure and parameter set are determined and realized on physical system. It is observed that, without any fine-tuning in controller parameters, experimental results completely match up with simulation results.