Neural Network Based Controls

Dissertations and Theses

Publications

Presentations

Abstracts

Controller Design Techniques for Flexible Link Manipulators Including Back-Propagation Neural Networks

The control of a rotating single flexible link manipulator and/or a two “coupled flexible link manipulator arm is a highly nonlinear problem. Due to the distributed flexibility, the dynamics of the flexible link manipulator system are described by infinite-dimensional mathematical model. However, the problem is simplified by using finite-element modeling. The research presented in this dissertation addresses three controller design methodologies for the given finite-element modeling/analysis package ANSYS. The three parts are: 1) variable structure system control validation; 2) self-tuning control validation; and 3) on-line self -tuning adaptive control using back-propagation neural networks. All three methods employ the concept of pseudolink to represent the tip-position of the flexible link manipulator system.

In the first part, an extremely simple controller design methodology is derived by treating the effects of distributed flexibility as an uncertain term in torque about which no apriori knowledge is assumed other than a bound. The controller design methodology is validated for a single rotating flexible link manipulator modeled using ANSYS.

In the second part, a self-tuning control strategy for modeling, identification, and self-tuning control of a SISO discrete-time nonlinear system is validated. The classical recursive least squares algorithm for system identification and pole-placement technique for controller design is employed. The pseudolink concepts are used to determine on-line angular displacement of the end-effector of the flexible link manipulator modeled using ANSYS.

Finally, the third part addresses the topics of on-line system identification of an unknown nonlinear dynamical system using back-propagation neural network and the on-line self-tuning adaptive control of such systems using a separate neural network as a controller. An on-line self-tuning adaptive control output tracking architecture/law using three multi-layer back-propagation neural networks is proposed. The weight updating of the neural networks is generalized using gradient methods. The convergence of errors of the closed loop feedback system architecture is proved. The new OLSTAC scheme/architecture is applied to a single rotating flexible link manipulator as well as to a two “coupled” flexible link manipulator arm. It is demonstrated through illustrative simulations that the OLSTAC scheme/architecture is generic in the sense that it requires minimal knowledge of the unknown plant. While the proposed OLSTAC scheme is based on an assumption, Assumption SI-1, that make the sufficient condition that the unknown nonlinear/uncertain system can be separated into two nonlinear terms, one, the nonlinear term in the state, and the other, the nonlinear term representing control, it works admirably for the cases when the system is highly nonlinear and uncertain, e.g., the flexible link manipulator systems.

Implementation of an Adaptive Controller to Control a Single Flexible Link Manipulator

A neural network based On-Line Self-Tuning Adaptive Controller (OLSTAC) designed by Mahmood [1] is implemented on a nonlinear system. The nonlinear system used is a single flexible link manipulator, which uses a direct drive motor as an actuator. The controller is designed such that the end of the flexible link effectively tracks an arbitrary reference signal. A model of the flexible link is derived using the pseudolink concept [2]. Background of the operation of neural networks and their applications in adaptive controls is given. The operation of the OLSTAC is briefly explained and the OLSTAC is implemented using a digital signal processor. Results of the implementation are shown, along with brief descriptions of and comparisons with previous work done on the link.

Chatter Control for Boring Bars Using a Neural Network Based Control

Most control applications require an accurate system model for controller design. Although it is fairly easy for simple setups, it can be very cumbersome for more complicated systems. The system is put through a lot of tests to determine its transfer function which is essential for the design of the control system. This design is then valid for a strict set of operational limits. This work looks at overcoming some of the classical system modeling problems by using neural networks to learn the system parameters and its transfer function. The network is then used to control the system. The system that was chosen for this work is that of a boring bar with a piezoelectric pusher as the active element ( active dynamic absorber ). The neural network algorithm runs on a digital signal processing board using the DSP 56000 providing on-line control.

The signal from the accelerometer at the tip of the boring bar was used as the input signal for the neural network. The network generates an appropriate output to the peizoelectric pusher in the boring bar, providing enough damping to reduce the vibrations at the tip of the bar. The results from the open loop and closed loop systems are discussed along with the two methods used for learning in the neural network. An introduction to neural networks, especially CMAC networks is an important part of this work as is explained in the project definition and scope of this project.

Neural Network Based Adaptive Control of a Flexible Link Manipulator