Neural Network Learning Control for Friction Compensation with Enhanced Generalizability

Friction nonlinearity poses a significant challenge to the high-precision, high-speed machining of Computer Numerical Control (CNC) machine tools. Traditional model-based system identification and compensation methods heavily depend on model assumptions, resulting in limited flexibility, while neural network-based approaches generally exhibit limited generalization capabilities. To achieve effective and generalizable friction compensation in practical CNC systems, this paper proposes a feature selection method that enhances the generalization ability of neural network-based models. Building on this, a Radial Basis Function Neural Network (RBFNN) learning control framework is developed to implement friction compensation by modifying the system's feedforward input. Theoretical analysis confirms the stability of the proposed framework. Experimental results on a three-axis CNC machine demonstrate that the proposed method, after a single training session, effectively compensates for friction across multiple trajectories not included in the training dataset.