关键词: 数据驱动, 进给系统, 机器学习, 神经网络, 跟踪误差

Abstract: Existing data-driven feed drive system modeling methods use end-to-end model, which employs machine learning algorithms to construct the model between the reference trajectory information (displacement, velocity, acceleration etc.) and the servo error directly, reducing the complexity of modeling. However, this kind of modeling method ignores the non-linear disturbance in motion control system reflected in the control signal, limiting the accuracy of constructed model. To solve the problem, a data-driven motion control error modeling method using control signal as intermediate value is proposed. Proposed method samples the reference trajectory information (velocity, acceleration, jerk etc.), control signal, tracking error, and constructs a reverse feature. Then, the control signal prediction network based on reference trajectory information and tracking error prediction network using control signal and reference trajectory information are constructed and trained, setting up the proposed model. Utilizing the feature that control signal can effectively reflect the nonlinear disturb, the accuracy of data-driven motion control error model is improved. In the validation and testing, the reference trajectory information is input into control signal prediction network to get the predicted control signal. And, the predicted control signal and reference trajectory information are input into tracking error prediction network, attaining the utilization of control signal as intermediate value in the data-driven modeling of feed drive system. Proposed modeling method is validated through experiments. Compared with traditional end-to-end modeling method without control signal, the proposed method shows better prediction accuracy. The prediction accuracy improves 16.33% and 20.42% respectively in X axis and Y axis, and the motion control contour accuracy improves 85.59% after error compensation, proving the effectiveness and feasibility of proposed method.

Key words: data-driven, feed drive system, machine learning, neural network, tracking error