Rotation Control Performance of a Friction Welding Repair System Using a Neural Network

This paper presents the rotation control performance of a friction welding repair system. The control parameters are adjusted by a neural network (NN). This control is capable of supporting the wide variations in torque that occur in this type of repair process. The rotation system consists of a hydraulic motor whose speed is controlled by a proportional directional valve. For this, a NN was developed with various structures composed of one, three, four and six neurons having activation functions of the following type: linear, quadratic, sigmoid and step. A theoretical plant was developed that simulates the behavior of the rotation system of a friction drive unity. This plant was used in the adjustment of the weights of the NNs by means of a genetic algorithm (GA). This plant was also used to evaluate the performance of NNs in the rotation control after their adjustment. The choice of the most suitable NN was made taking into consideration not only their control performance, but also the number of parameters to be adjusted, because the higher the number of these parameters, the greater the difficulty faced by the GA in the adjustment. Having selected the NN, its performance was compared to that of a PID controller. It was observed that the NN that reached the best performance was of one neuron with a linear activation function and its performance was higher than that of the PID controller.