基于Kriging模型和强化学习的激光直接能量沉积成形尺寸控制
Dimensional Control of Laser Direct Energy Deposition Forming Based on Kriging Model and Reinforcement Learning
胡楷雄 1李克 2周勇 2李卫东3
作者信息
- 1. 武汉理工大学交通与物流工程学院,湖北 武汉 430063;武汉理工大学襄阳示范区湖北隆中实验室,湖北 襄阳 441000
- 2. 武汉理工大学交通与物流工程学院,湖北 武汉 430063
- 3. 上海理工大学机械工程学院,上海 200093
- 折叠
摘要
针对传统的比例积分微分控制方法存在随工艺参数变化需重新整定控制器参数的缺陷,采用了一种基于Kriging预测模型的强化学习校正框架,用于预测和控制沉积层尺寸,而无需进行参数整定.该框架通过不断迭代学习工艺参数对沉积层尺寸的影响,对内置的Kriging预测模型进行校正,提高其预测性能,从而输出更优的工艺参数.实验结果表明,该框架能够抑制熔道回流效应,在减小成形件累计高度误差的同时,有效抑制宽度误差,从而有效提升激光直接能量沉积零件的尺寸精度.
Abstract
In view of the defect of the traditional proportional-integral-derivative control method,which needs to reset the controller parameters as the process parameters change,this study employs a reinforcement learning correction framework based on the Kriging model,where the framework is specifically designed to predict and control melt pool dimensions,thereby eliminating the need for parameter tuning.Through iterative learning of the effects of process parameters on melt pool dimensions,the framework corrects the embedded Kriging prediction model by enhancing its predictive performance and yielding more optimized process parameters.Experimental results demonstrate that this framework can mitigate the melt pool backflow effect,proficiently manage width errors,reduce cumulative height errors in formed components,and significantly enhance the dimensional accuracy of laser direct energy deposition components.
关键词
激光直接能量沉积/Kriging模型/强化学习/尺寸控制Key words
laser direct energy deposition/Kriging model/reinforcement learning/dimensional control引用本文复制引用
出版年
2024
NETL
NSTL
万方数据