Abstract: Aiming at the problem of high sensitivity to iteration initial guess when using Newton-Raphson iteration method to compute the forward kinematic solution of orthogonal Stewart-6D- platform, a compensation method based on Levenberg-Marquardt (L-M) algorithm of back propagation model iteration initial guess was proposed. The adaptation, the output prediction and error performance of the proposed algorithm were analyzed in comparison with the back propagation model based BFGS quasi-Newton algorithm, back propagation model based scale conjugate gradient, back propagation model based gradient descent with adaptive Ir algorithm. With the help of inverse solution model based Simulink and SolidWorks, successful training samples were established. 1000 groups of data were randomly and comprehensively picked up. 900 groups were used for training the net and the other 100 groups were used for testing whether the trained net is qualified. Simulation results show that the initial guess modified L-M algorithm of back propagation model was significantly calibrated. The convergent speed of the iteration modified algorithm was improved. The angle and displacement calibrating error rate was less than 0.1%. The test results of the improved BP Neural net can meet the requirement.

Key words: railway transportation, forward kinematics, railway vehicles, pose measurement system, Levenberg-Marquardt(L-M) algorithm, Newton-Raphson iteration