Residual learning of the dynamics model for feeding system modelling based on dynamic nonlinear correlate factor analysis

Feeding system modelling is the foundation for control strategy optimization, contour error compensation, etc., to improve the productivity and quality of a part. This paper proposes a novel residual learning approach for fitting the simulation error of the dynamics model of a machine tool feeding system. Then, the feeding system model consisting of the dynamics model and the residual model is constructed by integrating prior knowledge with statistical learning knowledge. The residual model is trained by using the training dataset generated from the dynamics model instead of using only the input and output data (i.e., the end-to-end data). In addition, a dynamic nonlinear correlate factor extraction method is proposed to extract the training dataset from the dynamics model and the reference data. Compared to the end-to-end data, the training dataset knows very well about the system’s nonlinear features owing to the internal prior knowledge of the dynamics model. Experiments conducted on a vertical milling centre confirm the effectiveness of the feeding system model in dynamic response prediction. Compared to the existing dynamics or data-driven modelling method, the proposed method can achieve higher prediction accuracy in nonlinear motion processes, such as the reverse process, and can obtain stable performance with respect to different feedrates owing to the residual learning approach.