Backlash compensation by neural-network online learning

To eliminate the influences of backlash nonlinearity that generally exist in servo systems, a new neural-network online learning compensation method is presented. Not basing on the identification of backlash nonlinearity, the online learning of neural networks is used instead making the output error of the system approximate to zero so that the system output can accurately follow the given input. To cooperating with this new method, the self-organizing fuzzy CMAC with Gauss basis functions neural network is proposed on the basis of utilising the advantages of traditional CMAC neural networks, fuzzy logic, basis functions and the self-organizing feature map algorithm. Finally, an actual experimental platform of servo system with low power was built. The experimental results show that the method presented can effectively remove the limit cycle caused by the backlash nonlinearity, and greatly improve the system accuracy.