LuGre-Net: a hybrid neural network for friction modeling of feed systems in machine tools

Friction has a substantial impact on the performance of machine feed systems, and establishing a high-precision friction model is the main premise of model-based friction compensation. To address the low accuracy and difficult parameter identification of the mathematical friction model, as well as the generalization issue of the neural network model, a novel friction model called LuGre-Net that combines the framework of the mathematical LuGre model with a neural network is proposed in this paper. Inspired by the formulas of the LuGre model, the topology of LuGre-Net is developed, with the ranges of LuGre's parameters incorporated into the LuGre-Net network as a priori knowledge. The experimental results demonstrate that the proposed LuGre-Net achieves great prediction accuracy, with a root mean square error (RMSE) and maximum absolute error (MAE) of 0.04 and 0.13 Nm on the test set, respectively. The RMSE of LuGre-Net is 63.9 and 17.8% lower than that of LuGre and the back propagation neural network (BPNN), respectively, while the MAE of LuGre-Net is 53.7 and 38.1% lower. Additionally, LuGre-Net outperforms BPNN in terms of generalization and sample dependence.