A Neural Network-Based Friction Compensation Method for Machine Tool Feed Drives

Abstract This paper presents a novel neural network-based friction compensation strategy to greatly improve the tracking accuracy of precision motion stages. Compensation of friction disturbances is critical for enhancing accuracy of precision positioning systems. Modeling of friction is challenging due to its highly non-linear behavior, as it transitions through various regimes, depending on the stage location, displacement and velocity. For instance, in the pre-sliding regime, friction forces depend non-linearly on the stage displacement, exhibiting complex hysteresis behavior with non-local memory. In sliding regime, friction forces depend non-linearly on the stage velocity, following the Stribeck curve. Analytical models, such as the comprehensive Generalized Maxwell-Slip (GMS) model has been widely used for friction compensation. Although GMS can model the individual friction regimes, i) it is inefficient in accurately capturing the friction transition between the regimes, and ii) it requires lengthy parameter identification procedure, which is cumbersome and renders impractical in many industrial application. For instance, GMS parameters identified at low velocities fail to generalize the friction behavior to higher velocities, resulting in modeling inaccuracies. To address these deficiencies of GMS, this work proposes a novel GMS-Net, a neural network (NN)-based friction model, which can accurately model and compensate non-linear friction disturbances for precision motion systems. GMS-Net incorporates multiple sub-networks for estimating the pre-sliding and sliding friction, as well as an additional sub-network to handle smooth transitions between the regimes. It can be trained across various velocity ranges to improve generalization and estimation accuracy. Experimental results show that GMS-Net can significantly enhance positioning accuracy of feed drives, reducing the peak friction-induced tracking errors from 10 microns to 4 microns as compared to using the conventional GMS model.