基于强化学习的木工送料平台误差控制研究

Research on Error Control of Wood Feeding Platform Based on Reinforcement Learning

研究了木工带锯机送料平台误差产生的主要原因。通过深度确定性策略梯度算法,对各轴的跟踪误差进行控制,以实现对送料平台末端夹具及木料运动轨迹误差的控制。使用Adams运动学仿真平台对送料平台进行动力学分析,通过与MATLAB进行联合仿真完成强化学习过程。由于送料平台各轴间耦合关系复杂,所以使用最小安全距离限制和积分法加快仿真学习过程。最后进行数据迁移,使用Codesys实现对送料平台实体机的虚轴控制进行仿真。结果表明,加入约束条件和优化方法后,强化学习过程更加稳定且收敛速度更快,深度确定性策略梯度算法减小误差可达63.97%,为后续锯切加工奠定了基础。

The main reasons for the error of the feeding platform of woodworking band sawing machine were studied in this article.The tracking errors of each axis were controlled through the deep deterministic policy gradient algorithm,in order to achieve control of the errors of the end fixture of the feeding platform and the trajectory of wood movement.The Adams motion simulation platform was used to perform dynamic analysis on the feeding platform,and the reinforcement learning process was completed through joint simulation with MATLAB.Due to the complex coupling relationship between the axes of the feeding platform,the minimum safety distance constraint and the integration method were used to speed up the simulation learning process.Finally,data migration was performed,and Codesys was used to implement experiments on the control of the feeding platform's physical machine.The results showed that the reinforcement learning process with constraints and optimization methods added was more stable and the convergence speed was faster.The deep deterministic policy gradient algorithm can reduce the error by 63.97%o,which provided a foundation for further sawing process.