Neural network compensation of gear backlash hysteresis in position-controlled mechanisms

This paper demonstrates that artificial neural networks can be used to identify and compensate for hysteresis caused by gear backlash in precision position-controlled mechanisms. A major contribution of this research is that physical analysis of the system nonlinearities and optimal control are used to design the neural network structure. Network sizing and initializing problems are thus eliminated. This physically meaningful, modular approach facilitates the integration of this neural network with existing controllers; thus, initial performance matches that of existing control approaches and then is improved by refining the parameter estimates via further learning. The neural network operates by recognizing backlash and switching to a control which moves smoothly through the backlash when the torque transmitted to the output shaft must be reversed.