Abstract
In modern industry, an industrial system is composed of lots of components or subsystems. Because the manufacturers produce these subsystems according to a generic standard or design guideline without full knowledge about the target system, the control model and system dynamics will only follow a generic pattern with typical parameters, which will cause parameter changes on the real system. Also, the system dynamics may be affected by thermal effects and mechanicaland electrical ageing, which cannot be directly measured and is challenging to account for in the generic control model. These factors will result in the performance deterioration of the control system under different scenarios. Therefore, it is necessary to provide a method for tuning and adapting the subsystems at runtime to improve the system performance.
In this report, we propose an online model identification and adaptation of a control system at runtime. The NARMA-L2 controller, a kind of neural network controller, is used for identification and controller design. We show that the NARMA-L2 controller can achieve high training accuracy and realize the control requirement in both linear and nonlinear systems. We validate our approach using the cruise control system designed in Webots. Based on the NARMA-L2 controller, online model identification method and an adaptive controller are designed to adapt to parameters' changes. Also, for validation of the controller, we develop a MATLAB toolbox for SiL and HiL
simulation to validate the controller, which supports code generation for different parameters, integration of widgets like SDF3 dataflow analysis, and switch between different implementation platforms and controllers. Our evaluation shows that adapting the controller at runtime has better performance than the traditional PID controller.
| Date of Award | 12 Jan 2021 |
|---|---|
| Original language | English |
| Supervisor | Dip Goswami (Supervisor 1) & Sajid Mohamed (Supervisor 2) |