Improvement of Data Accuracy on Backpropagation Neural Network-based Automatic Control System for Wheeled Robot

摘要

One of the problems in developing a neural network-based automatic control system is data accuracy. This study aims to improve the accuracy of the training data utilized in a backpropagation neural network-based controller system under the direct inverse control scheme. One of the main strategies in improving the data accuracy was by recalculating the sampling time to accommodate the delay time of GPS signals. After obtaining accurate training data, backpropagation training was then carried out with several variations in the number of neurons to get the best neural network configuration. The simulation results showed that the best control performance was obtained with 13 input neurons, 10 hidden neurons and 2 output neurons with normalized training Mean Square Error (MSE) of 1.22 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-2</sup> . This neural network controller was then implemented to a developed wheeled robot, and the trajectory generated by the robot was compared to a certain test data trajectory. The experiment proved that the wheeled robot is able to follow the desired test trajectory with an acceptable normalized MSE value of 0.164. This result indicates that a backpropagation neural network-based control system can be implemented in a wheeled robot with sufficient and accurate training data.