5th Parallel Kinematics Seminar, Parallel Kinematic Machines in Research and Practice, Chemnitz, April 25-26, 2006, Tagungsband S. 603-616
In this paper a method is proposed to identify the real geometry of an assembled hybrid- kinematic by means of radial based neural networks with positioning errors as input and geometric deviations as output data. The calibration is realized in 3 major steps. Firstly a set of parameters is defined to describe potential geometric defects. In simulations with an arbitrary variety of settings the reciprocal dependencies are carried out. Secondly simple test traces are defined to pre-calibrate an isolated fraction of the error values which are nearly unaffected by other defects along these traces. The parameter estimation is done by a neural network trained by simulated data. Thirdly the remaining parameters are estimated by a second neural network which is trained by simulations based on the pre-calibration. In this step the calibration traces are spirals that can be followed with a double bar measurement system. An iterative application of steps two and three increases the accuracy.