基于机器学习的激光定向能量沉积质量预测与工艺参数优化  

A review of machine learning-based quality prediction and process parameter optimization in laser directed energy deposition

作　　者：王振忠 刘超凡 李毅 吴俊波 陈丰 WANG Zhenzhong;LIU Chaofan;LI Yi;WU Junbo;CHEN Feng(School of Aerospace Engineering,Xiamen University,Xiamen 361100,China;College of Mechanical Engineering and Automation,Huaqiao University,Xiamen 361021,China)

机构地区：[1]厦门大学航空航天学院,福建厦门361102 [2]华侨大学机电及自动化学院,福建厦门361021

出　　处：《厦门大学学报(自然科学版)》2025年第2期225-240,共16页Journal of Xiamen University:Natural Science

摘　　要：[背景]激光定向能量沉积(LDED)技术是增材制造中备受关注的工艺之一,具有加工效率快、成形灵活度高等优点.然而由于沉积过程中多种复杂因素的相互作用,使得加工过程中熔池的形态特征与影响因素的相互关系难以仅通过传统手段监测得到,因此,较难准确预测工件的质量和选择合适的工艺参数.而机器学习(ML)算法具有强大的非线性处理能力和自适应学习能力,有望实现对加工过程缺陷的检测与预测、工件成形精度和性能预测及工艺参数的逆向求解.[进展]首先,本文总结了两种常见LDED形式、LDED加工过程中常见缺陷类型及缺陷产生的原因,并对四种基于光学相机、温度场监测、光谱分析和声学传感器的数据采集方式进行分类比较;然后对常用于LDED中的监督学习、无监督学习及半监督学习算法进行比较总结,分析各自适用场景及优缺点;最后,从ML在LDED缺陷检测与预测、成形精度预测、成形性能预测和工艺参数逆向求解四个方面进行综述,系统阐述如何利用各类传感器并结合ML算法对LDED工件加工质量进行优化.[展望]建立多ML算法融合的模型、探究信号和缺陷之间的定量关系和实现多传感器数据融合监测,有望厘清LDED工艺参数与沉积质量的关系,为提升LDED沉积质量提供理论指导.[Background]As one of most prominent processes in additive manufacturing,laser directed energy deposition(LDED)technology offers advantages such as high processing efficiency and exceptional flexibility in shaping.However,due to the interaction of multiple complex factors during the sedimentation process,it is difficult to accurately predict the quality of the workpiece and select appropriate process parameters by monitoring the relationship between the morphological characteristics of the molten pool and the influencing factors solely through traditional methods.Machine learning(ML)algorithms combined with LDED,with their powerful nonlinear processing capability and adaptive learning ability,can detect as well as predict defects in the process,forecast the forming accuracy as well as the performance of the workpiece,and solve process parameters inversely.[Progress]First in this paper,we summarize two common forms of LDED,namely the typical defect types encountered in LDED processing and the factors contributing to their occurrence.Four primary data acquisition methods based on optical cameras,thermal field monitoring,spectroscopy,and acoustic sensors are categorized and compared.Second,we provide a comparative analysis of supervised learning,unsupervised learning,and semi-supervised learning algorithms commonly used in LDED,and discuss their respective application scenarios,strengths,and limitations.Finally,we systematically review four aspects:(a)ML-based defect detection and prediction,(b)shape accuracy prediction,(c)mechanical property prediction,and(d)inverse solving of process parameters.Consequently,we are capable of illustrating how various sensors can be integrated with ML algorithms to optimize the manufacturing quality of LDED components.[Perspective]Establishing a multi-ML algorithm fusion model,exploring the quantitative relationship between signals and defects,and implementing multi-sensor data fusion monitoring are expected to clarify the relationship between LDED process parameters as well as sedi

关 键 词：增材制造 激光定向能量沉积(LDED) 机器学习 过程监测 

分 类 号：TG665[金属学及工艺—金属切削加工及机床]