| dc.description.abstract | A networked control system (NCS) is an interconnected control system in which sensors, actuators, and controllers communicate with each other through a shared network. Although NCSs are beneficial thanks to easy maintenance, architectural flexibility, decreased wiring weights, and tele-operating possibilities, NCSs also have some challenges such as disturbance, noise, bandwidth limitation, delay, and packet dropout. The popularization of smartphones and the drastically increasing number of internet of things (IoT) devices require not only a high-speed internet such as 5G, but also a wise strategy for optimal bandwidth allocation. In this dissertation, optimal bandwidth allocations for NCSs with disturbance and noise are proposed based on performance index function (PIF), artificial neural network (ANN), and Q-learning algorithms. A ball magnetic-levitation (maglev) system, four DC motor speed-control systems, and a wireless autonomous robotic wheelchair are implemented as test beds. The relationship between system performance, sampling frequency, and the standard deviation of white Gaussian disturbance are approximated using a 6 th-degree polynomial. The PIF and ANN methods can estimate the standard deviation of disturbance when current a sampling frequency and an error variance are provided. Dynamic bandwidth allocation using PIF, ANN, and Q-learning is proposed and verified by experimental results for a single-server and single-client DC motor system.
The proposed methods show integral absolute errors (IAE) of 166 615, 16 773, and 16 945 and bandwidth utilizations (BU) of each method are 13.15%, 13.38%, and 13.98%, respectively, after 15 000 iterations, when various standard deviations of disturbance are injected. These results present a better performance and a reasonable average BU compared to fixed sampling frequencies. When information of the estimated standard deviation of disturbance, BU margin of safety, weight of each system, and total time delay is given, the optimal sampling frequency for a multi-server and multi-client system can be determined based on the PIF, ANN, and Q-learning, respectively. They are validated by experiments in two cases. The first case is conducted with a ±0.8-V disturbance, 10% safety margin of BU, 1.25-ms total time delay, and various weights for four DC motor systems. The second has the conditions of a ±0.8-V disturbance, 10% safety margin of BU, 1.25-ms total time delay, various weights for four DC motor systems with a maglev and a wheelchair robot system, of which BUs are 44% and 1% respectively. Experimental results prove that all three methods can be used to find the optimal sampling frequencies for each system when an NCS has limited bandwidth as well as sufficient bandwidth. | en |
