李治文,张志芬,张帅,等.基于空气传播声发射与深度迁移学习的激光粉末床熔融缺陷在线监测[J].精密成形工程,2023,15(11):76-88.
LI Zhi-wen,ZHANG Zhi-fen,ZHANG Shuai,et al.Online Monitoring for Laser Powder Bed Fusion Defects Based on Air-borne Acoustic Emission and Deep Transfer Learning[J].Journal of Netshape Forming Engineering,2023,15(11):76-88. |
| 基于空气传播声发射与深度迁移学习的激光粉末床熔融缺陷在线监测 |
| Online Monitoring for Laser Powder Bed Fusion Defects Based on Air-borne Acoustic Emission and Deep Transfer Learning |
| 投稿时间：2023-08-17 |
| DOI：10.3969/j.issn.1674-6457.2023.011.009 |
| 中文关键词: 增材制造 激光粉末床熔融 空气传播声发射 迁移学习 缺陷在线监测 |
| 英文关键词: additive manufacturing (AM) laser powder bed fusion (LPBF) air-borne acoustic emission (ABAE) transfer learning online monitoring of defects |
| 基金项目:国家重点研发计划(2022YFB4600803) |
|
| 摘要点击次数: 1480 |
| 全文下载次数: 546 |
| 中文摘要: |
| 目的 对激光粉末床熔融(LPBF)过程中的缺陷进行缺陷监测机理研究，并开发可靠的现场质量监测方法来指导零件生产过程。方法 使用空气传播声发射(ABAE)技术监测LPBF过程，获取多组变工况LPBF过程中的ABAE信号。通过对LPBF过程进行熔池动态分析及对比缺陷信号时频域特征，探究了LPBF缺陷声学监测机理。同时，提出了一种基于多源域知识融合的深度迁移学习(DTL-MDKF)方法以构建LPBF缺陷在线监测模型，并对模型的准确度和有效性进行了系统评估。结果 LPBF熔池状态与产生的缺陷密切相关，所对应的声源发生机理不同是导致其信号特征呈现差异性的主要原因。所提出的DTL-MDKF方法模型对5类LPBF缺陷的识别准确率可达98.2%，且t-SNE可视化分析结果证明了其良好的深度特征挖掘能力以及将模型中多源域自适应融合知识迁移用于目标域LPBF缺陷监测任务的有效性。结论 相比于传统单一源域迁移学习方法，所提出的方法模型具有更好的性能，能够实现对复杂多变工况的LPBF制造缺陷的精准实时监测，可为LPBF零件的生产提供一定的指导。 |
| 英文摘要: |
| The work aims to research the monitoring mechanisms of defects generated in the laser powder bed fusion (LPBF) process and develop reliable on-site quality monitoring methods to guide the part production process. In this paper, air-borne acoustic emission (ABAE) technique was used to monitor the LPBF process, and several sets of ABAE signals of LPBF process under variable operating conditions were obtained. The LPBF defect acoustic monitoring mechanism was explored by analyzing the molten pool dynamics of the LPBF process and comparing the time-frequency domain characteristics of defect signals. Meanwhile, a deep transfer learning method with multi-source domain knowledge fusion (DTL-MDKF) was proposed for constructing an online monitoring model of LPBF defects, and the accuracy and effectiveness of the model were systematically evaluated. The LPBF melt pool state was closely related to the defect generated. The different generation mechanisms of the corresponding sound sources were the main reasons for the variability of their signal characteristics. The proposed DTL-MDKF method model could identify 5 types of LPBF defects with an accuracy of 98.2%, and the t-SNE visualization analysis results demonstrated its good deep feature mining capability and the effectiveness of the model's multi-source-domain adaptive fusion knowledge transfer for the task of monitoring LPBF defects in the target domain. Compared with the traditional single-source domain transfer learning method, the proposed method model has better performance and can realize accurate real-time monitoring of LPBF manufacturing defects in complex and variable working conditions, which can provide certain guidance for the production process of LPBF parts. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
渝公网安备 50010702501719号