Ahmet Onat

Wireless Model-Based Predictive Networked Control System Over Cooperative Wireless Network

Owing to their distributed architecture, networked control systems (NCSs) are proven to be feasible in scenarios where a spatially distributed feedback control system is required. Traditionally, such NCSs operate over real-time wired networks. Recently, in order to achieve the utmost flexibility, scalability, ease of deployment, and maintainability, wireless networks such as IEEE 802.11 wireless local area networks (LANs) are being preferred over dedicated wired networks. However, conventional NCSs with event-triggered controllers and actuators cannot operate over such general purpose wireless networks since the stability of the system is compromised due to unbounded delays and unpredictable packet losses that are typical in the wireless medium. Approaching the wireless networked control problem from two perspectives, this work introduces a practical wireless NCS and an implementation of a cooperative medium access control protocol that work jointly to achieve decent control under severe impairments, such as unbounded delay, bursts of packet loss and ambient wireless traffic. The proposed system is evaluated on a dedicated test platform under numerous scenarios and significant performance gains are observed, making cooperative communications a strong candidate for improving the reliability of industrial wireless networks.

Control Over Imperfect Networks: Model-Based Predictive Networked Control Systems

Networked control systems (NCSs) are digital control systems in which the functionality of the sensor, control, and actuator reside in physically different computer nodes communicating over a network. However, random delays and data loss of the communication network can endanger the stability of an NCS. We have proposed model-based predictive NCSs (MBPNCSs) that compensate for the aforementioned problems and avoid performance loss using a predictive control scheme based on a model of the plant. There are three main contributions of this paper to existing methods: an NCS that can work under random network delay and data loss with realistic structural assumptions, an explicit mechanism for reducing the effects of network delay and data loss on the deviation of plant state estimates from actual plant states, and an architecture where upstream nodes can work without receiving acknowledge information about the status of previously sent data packets from downstream nodes. In this paper, we describe MBPNCS and then introduce a stability criterion. This is followed by computer simulations and experiments involving the speed control of a dc motor. The results show that considerable improvement over performance is achieved with respect to an event-based NCS.

Adaptive gait pattern control of a quadruped locomotion robot

The authors have proposed a control system of a quadruped locomotion robot by using nonlinear oscillators. It is composed of a leg motion controller and a gait pattern controller. The leg motion controller drives the actuators of the legs by using local feedback control. The gait pattern controller involves nonlinear oscillators with mutual interactions. In the paper, capability of adaptation of the proposed control system to variance of the environment is verified through numerical simulations and hardware experiments. With the input signals from the touch sensors at the tips of the legs, the nonlinear oscillators tune the phase differences among them through mutual entertainments. As a result, a gait pattern corresponding to the states of the system or to the properties of the environment emerges. The robot changes its gait pattern adaptively to variance of the environment and establishes a stable locomotion while suppressing the energy consumption.

Design and Implementation of a Linear Motor for Multicar Elevators

The multicar elevator system is a revolutionary new technology for high-rise buildings, promising outstanding economic benefits, but also requiring new technology for propulsion, safety, and control. In this paper, we report on experimental results with new components for linear-motor-driven multicar elevators. We show that linear synchronous motors with optimized design and with our new safety and control system can be considered as core components of a new generation of elevator systems. The main new results concern the development of a safety system integrated into the propulsion system, the design methodology of a linear motor optimized for the multicar elevator task, and the motion control system that is expected to be usable for extra high-rise buildings.

Modeling of Air Core Permanent-Magnet Linear Motors With a Simplified Nonlinear Magnetic Analysis

This paper presents a new analysis method for air core permanent-magnet synchronous linear motor design optimization problems with a large search space. The aim is to reduce the total computational time by replacing most of the finite-element analysis (FEA) steps with an analytical model of the motor consisting of nonlinear equivalent magnetic circuit and flux density distribution models. A small number of FEA steps are used to determine some parameters of the analytical model. The result of the proposed method is compared with results of both a pure FEA design and experiments on a real motor. The proposed method has similar performance to FEA, although computationally it is at least two orders of magnitude faster.

SURALP: A new full-body humanoid robot platform

SURALP is a new walking humanoid robot platform designed at Sabanci University - Turkey. The kinematic arrangement of the robot consists of 29 independently driven axes, including legs, arms, waist and a neck. This paper presents the highlights of the design of this robot and experimental walking results. Mechanical design, actuation mechanisms, sensors, the control hardware and algorithms are introduced. The actuation is based on DC motors, belt and pulley systems and Harmonic Drive reduction gears. The sensory equipment consists of joint encoders, force/torque sensors, inertial measurement systems and cameras. The control hardware is based on a dSpace digital signal processor. A smooth walking trajectory is generated. A variety of controllers for landing impact reduction, body inclination and Zero Moment Point (ZMP) regulation, early landing trajectory modification, and foot-ground orientation compliance and independent joint position controllers are employed. A posture zeroing procedure is followed after manual zeroing of the robot joints. The experimental results indicate that the control algorithms presented are successful in improving the stability of the walk.

Model Based Predictive Networked Control Systems

Networked control systems where the sensors, controller and actuators of a digital control system reside on different computer nodes linked by a network, aim to overcome the disadvantages of conventional digital control systems at the application level, such as difficulty of modification, vulnerability to electrical noise, difficulty in maintenance and upgrades. However random communication delay and loss on the network may jeopardize stability since the communication delay decreases the phase margin of the control system and data loss can be considered as noise. In this project, we propose a novel networked control method where satisfactory control is possible even under random delay and data loss. We keep a model of the plant inside the controller node and use it to predict the plant states into the future to generate corresponding control outputs. At every sampling period the states of the model are reset to the actual or predicted states of the plant. The ambiguity of plant state during periods of total communication loss are also addressed. The proposed model based predictive networked control system architecture is first verified by simulation on the model of a DC motor under conditions of data loss and noise. Then experiments are repeated on a dedicated test platform using a physical DC motor. Results show that significant amounts of data loss and delay can be tolerated in the system before performance starts to degrade.

Decentralized autonomous control of a quadrupedal locomotion robot using oscillators

This article deals with the design of a control system for a quadrupedal locomotion robot. The proposed control system is composed of a leg motion controller and a gait pattern controller within a hierarchical architecture. The leg controller drives actuators at the joints of the legs using a high-gain local feedback control. It receives the command signal from the gait pattern controller. The gait pattern controller, on the other hand, involves nonlinear oscillators. These oscillators interact with each other through signals from the touch sensors located at the tips of the legs. Various gait patterns emerge through the mutual entrainment of these oscillators. As a result, the system walks stably in a wide velocity range by changing its gait patterns and limiting the increase in energy consumption of the actuators. The performance of the proposed control system is verified by numerical simulations.

Reinforcement learning of dynamic behavior by using recurrent neural networks

Reinforcement learning is a learning scheme for finding the optimal policy to control a system, based on a scalar signal representing a reward or a punishment. If the observation of the system by the controller is sufficiently rich to represent the internal state of the system, the controller can achieve the optimal policy simply by learning reactive behavior. However, if the state of the controlled system cannot be assessed completely using current sensory observations, the controller must learn a dynamic behavior to achieve the optimal policy.In this paper, we propose a dynamic controller scheme which utilizes memory to uncover hidden states by using information about past system outputs, and makes control decisions using memory. This scheme integrates Q-learning, as proposed by Watkins, and recurrent neural networks of several types. It performs favorably in simulations which involve a task with hidden states.

Recurrent neural networks for reinforcement learning: architecture, learning algorithms and internal representation

Reinforcement learning is a learning scheme for an autonomous agent that allows the agent to find the optimal policy of taking actions which maximize a scalar reinforcement signal in unknown environments. If the agent has access to the whole state of the environment, a reactive policy which maps the sensory input to the action is sufficient. However, if the state of the environment is partially observable, special methods for creating a dynamic policy that utilizes the past observations are necessary. To overcome this problem, the authors have proposed a method using recurrent neural networks with Q-learning, as a learning agent. The paper compares several types of network architecture and learning algorithms for this method through computer simulation. Further, the internal representation in the trained networks is examined using a clustering technique. It shows that the representation of the environmental state is developed well in the networks.