L. A. Rybak, V. V. Erzhukov, S. G. Pochekaev, A. V. Chichvarin

In this paper offered common structure of control unit for machine-tool with parallel kinematics, in which all coordinate position is coupled, and moving on one coordinate required simultaneous moving on al other. Control unit is responsible for definition of surface and form of trajectory. General feedback on results of surface processing is implemented to scheme. Through the sensor are obtaining signals about length change and based on these results are solving of direct forward kinematics, thereby are obtaining real position of working tool. Then we are calculated inaccuracy as compare between control force (expected position of working tool) and real position. After that we are calculating correcting of length legs. In this time are solved inverse kinematics problem, namely finding the length of legs through control force. Summing up correcting of length with previously value we are obtaining control force to the servo drive. Solving of inverse kinematic problem may be obtained by calculating lengths of legs using vector shape. But this approach doesn-t using for solving direct kinematic problem. For this problem using numerical methods, especially neural networks. As training method we used combination of back propagation of error and annealing methods. In this paper listed comparative characteristics of several types of neural networks. Recognized that optimal in our case is using of correcting neural networks, i.e. in comparison with multilayered neural network it have smaller errors and training time. It will be observed, that forward pass time in network with correcting structure is also smaller. This comes from the fact that each network (basic and correcting) have only one layer and easier than network with multiple layers.