Yazar "Tufenkci, Sevilay" seçeneğine göre listele

Listeleniyor 1 - 5 / 5
  • Küçük Resim Yok
    Öğe
    A theoretical demonstration for reinforcement learning of PI control dynamics for optimal speed control of DC motors by using Twin Delay Deep Deterministic Policy Gradient Algorithm
    (Pergamon-Elsevier Science Ltd, 2023) Tufenkci, Sevilay; Alagoz, Baris Baykant; Kavuran, Gurkan; Yeroglu, Celaleddin; Herencsar, Norbert; Mahata, Shibendu
    To benefit from the advantages of Reinforcement Learning (RL) in industrial control applications, RL methods can be used for optimal tuning of the classical controllers based on the simulation scenarios of operating con-ditions. In this study, the Twin Delay Deep Deterministic (TD3) policy gradient method, which is an effective actor-critic RL strategy, is implemented to learn optimal Proportional Integral (PI) controller dynamics from a Direct Current (DC) motor speed control simulation environment. For this purpose, the PI controller dynamics are introduced to the actor-network by using the PI-based observer states from the control simulation envi-ronment. A suitable Simulink simulation environment is adapted to perform the training process of the TD3 algorithm. The actor-network learns the optimal PI controller dynamics by using the reward mechanism that implements the minimization of the optimal control objective function. A setpoint filter is used to describe the desired setpoint response, and step disturbance signals with random amplitude are incorporated in the simu-lation environment to improve disturbance rejection control skills with the help of experience based learning in the designed control simulation environment. When the training task is completed, the optimal PI controller coefficients are obtained from the weight coefficients of the actor-network. The performance of the optimal PI dynamics, which were learned by using the TD3 algorithm and Deep Deterministic Policy Gradient algorithm, are compared. Moreover, control performance improvement of this RL based PI controller tuning method (RL-PI) is demonstrated relative to performances of both integer and fractional order PI controllers that were tuned by using several popular metaheuristic optimization algorithms such as Genetic Algorithm, Particle Swarm Opti-mization, Grey Wolf Optimization and Differential Evolution.
  • Küçük Resim Yok
    Öğe
    An Approach for DC Motor Speed Control with Off-Policy Reinforcement Learning Method
    (2023) Tufenkci, Sevilay; Kavuran, Gürkan; Yeroglu, Celaleddin
    Integration of self-learning mechanisms with control systems is frequently encountered in the literature due to the development of autonomous systems. This paper proposes a tuning method of PI controllers using a deep reinforcement learning algorithm, which is known as self-learning structure. The coefficients of the PI controller, which is used to control a DC motors, are determined. The proposed method aims to adjust the voltage value applied to the input of the DC motor to reach the desired speed with the tuned PI controller using the twin- delayed deep deterministic policy gradient (TD3) reinforcement learning algorithm. The Kp and Ki coefficients of the PI controller are taken as the absolute values of the neural network weights, which are driven by Gradient descent optimization to positive values with a fully connected layer. The proposed tuning method has been shown to provide a higher gain margin and a more optimal solution.
  • Küçük Resim Yok
    Öğe
    An overview of FOPID controller design in v-domain: design methodologies and robust controller performance
    (Taylor & Francis Ltd, 2023) Tufenkci, Sevilay; Alagoz, Baris Baykant; Senol, Bilal; Matusu, Radek
    The complex v-plane is an emerging design domain for fractional order control system design. Recently, several works demonstrated the advantages of tuning FOPID controllers in v-plane. These approaches essentially perform the minimum angle pole placement to a target angle within the stability region of the v-plane and facilitate fractional order control system design tasks because of inherently guaranteed stabilisation of fractional order transfer functions. Accordingly, the optimal FOPID controller tuning problem can be simplified to placement of minimum angle system pole to a target point within the stability region of the v-plane. After reviewing previous v-domain design works, authors investigate prominent target points that can result in improved FOPID control performance for the v-domain design task. The consideration of target points in polar coordinates can provide two design parameters (angle and magnitude), which can convey essential system knowledge associated with the stability and control performance of FOPID control systems. In this perspective, effects of minimum angle pole positions on control performance indices are investigated in detail, and some prominent target points to manage FOPID design in v-domain have been reported. The v-domain design examples are illustrated to reveal the effects of the sampled pole positions on the robust control performance.
  • Küçük Resim Yok
    Öğe
    Implementations of TD3 and DDPG Reinforcement Learning Techniques for Tuning PID Controller of TRMS System
    (Springer Heidelberg, 2025) Tufenkci, Sevilay; Alagoz, Baris Baykant; Kavuran, Gurkan; Yeroglu, Celaleddin; Herencsar, Norbert; Mahata, Shibendu
    Reinforcement Learning (RL) is a learning method that utilizes interactions between agents and their environments, providing a valuable tool for controller design through simulations. However, traditional industrial systems such as PID control loops have yet to fully embrace the advantages of RL algorithms for effectively tuning controllers. This study presents an experimental initiative demonstrating the implementation of an RL-driven method for optimal PID controller tuning to address challenges in rotor control, explicitly focusing on the Twin-Rotor Multi-Input Multi-Output System (TRMS). Rotor control presents a complex challenge involving aerodynamics and external disturbances. The research implements two RL algorithms, namely the Deep Deterministic Policy Gradient (DDPG) and the Twin Delay Deep Deterministic Policy Gradient (TD3), in a tailored simulation environment to train RL agents to achieve optimal PID control dynamics. Results of simulation and experimental studies indicate that RL algorithms can be implemented for PID controller tuning when the simulation environment for training the RL algorithms well-represent the dominating dynamics and control complications of real-world systems. In this case, both the simulation and experimental results are in good-agreement.
  • Küçük Resim Yok
    Öğe
    Improved classification of star and galaxy from telescope by using a spatio-spectral feature ResNet model
    (Elsevier Sci Ltd, 2026) Tufenkci, Sevilay; Alagoz, Baris Baykant
    Given the vast number of galaxies and stellar constellations, automatic identification and morphological classification of stars and galaxies have become increasingly important for astronomical research. Astronomers rely on automated methods for distinguishing star and galaxy images in astronomical observations. The Convolutional Neural Networks (CNNs), which are powerful machine learning tools for the multi-class, closed-set image classification problems, have been effectively applied to the classification of astronomical images. However, accurate classification of astronomical images remains a challenging task because of a number of natural and technical difficulties, such as atmospheric seeing, instrumental noise, brightness variation and the low-image resolution. This study investigates use of spatio-spectral features in order to enhance the performance of ResNet-based classification models for distinguishing stars and galaxies in telescopic images. The spatial features in the pixel domain are combined with spectral features from the frequency domain to obtain a three channel spatio-spectral image representation. We demonstrate that combining spatio-spectral features improves the performance robustness of ResNet neural network classification model. Advantages of these features in the image classification problem come from properties that phase spectrum is nearly invariant to brightness variations, whereas the amplitude spectrum is relatively invariant to source position shifting in the image. To illustrate the effectiveness of spatio-spectral features in star-galaxy classification, the authors conducted experiments on low-resolution, noisy images that were captured by the 1.3-m telescope at the Devasthal Observatory. The results show that incorporating spatio-spectral features into ResNet-50 models can improve the classification accuracy by up to 12 % on this dataset. (c) 2025 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.