Real time implementation of NARMA-L2 Control of a single link manipulator

American Journal of Applied Sciences, Dec, 2008 by S.S. Mokri, H. Husain, W. Martono, A. Shafie

INTRODUCTION

Industrial manipulator robots play an important role in the field of flexible automation. A single link manipulator is the most basic manipulator, which is operated to perform tasks such as moving payloads or painting objects. To obtain a high performance single link manipulator, position controllers are necessary in order that the manipulator follows a preselected positional trajectory specified either as point-to-point or continuous path tracking motion with minimal deviation.

Control of DC driven single link manipulator is a nonlinear control problem due to gravitational force, mass of the payload, posture of the manipulator and viscous friction coefficient. Besides, uncertainties arise due to changes of the DC motor (actuator) parameters such as change of rotor resistance caused by temperature change as well as random variation of friction while operating. As a consequence, classical control algorithm, which is developed based on linear system assumption such as PID controller (13) is inadequate to deal with this problem.

In addition, an overwhelming majority of the available controllers are designed based on the assumption that the actuator dynamics are negligible (12), (2), (10), (25). This assumption reduces the dynamic model of the robot and facilitates the design of controllers. As a result of this simplification, unmodeled disturbances exist in the robot control systems, which affect the tracking and positioning of the robot. Although there are several methods to make a controller robust, with respect to the unmodeled dynamics, the performance of the controller is not as expected, meaning that the tracking errors are bounded but do not converge to zero (23).

To deal with unknown nonlinearities, various control strategies have been proposed in the forms of variable structure controller (25), robust control (6) and adaptive controller (20). However, the essential characteristic of these controllers is the model dependence, i.e., the requirement for explicit a priori specified model structure is still a necessity. In case of the manipulator robot, it is difficult to obtain some parameters such as the inertia matrix and mass centers at any joint with sufficient accuracy. It is then considered that these controllers are pertinent in the sense that an accurate model of the manipulator is demanded prior to controller design stage.

Therefore, a viable alternative to achieve an efficient control scheme is through the appliance of intelligent control. Intelligent control approaches such as neural network and fuzzy inference system do not require mathematical model of the system under controlled and have the ability to approximate nonlinear system. The real time applications of fuzzy logic control specifically for single link manipulator were reported in (8), (11).

Many researches have been attempting to use neural network intelligent controls for the trajectory control of robot manipulators. Among recent works carried out in the field of control of robot manipulators using neural network based controllers is (9). They proposed a controller for robust backstepping control of a general nonlinear system using neural networks. One of the nonlinear systems considered is robot manipulator. Here, the inertia matrix is considered to be known. Ahmad et al. (1) proposed a neural network controller based on modified Kohonen's Self Organizing Map (SOM) which controls the joint of the manipulator. Basically, the proposed controller consists of two neural network schemes; the neural network controller and the robotic emulator. The neural network controller will determine the joint torque after introducing the desired end effector's coordinate. The torque value will then become the input to the robotic emulator; which output is feedback to the controller to form a closed loop control.

In addition (19), also reported a direct inverse neural network based control scheme to solve the tracking control problem for the robot arm. Attempts are made in the literature to combine sliding mode and neural network to achieve the desired position control (17). They presented a sliding model neural network control scheme. Here, a neural network controller is developed to estimate the equivalent control in the sliding mode control.

As a matter of fact, these newly devised neural networks based controllers effectiveness is verified through simulation studies. On the other hand, actual real time applications as well as comparative experimental results study are rarely established. Illustrating one example is the work of (5), in which the author reported an indirect neural network real time control for an industrial robot arm. Also (21), described experimental results for position control of real manipulator using a kind of neural controller that operates in parallel with a conventional controller based on the feedback error learning architecture. They stated that the advantage of this controller is that it does not require any modification of the previous conventional controller.