航空学报 > 2022, Vol. 43 Issue (9): 425855-425855   doi: 10.7527/S1000-6893.2021.25855

粉末床激光成形熔池辐射强度信号的机器学习

段玉聪1, 王学德1, 周鑫1, 张佩宇1, 郭西洋1, 成星1,2, 樊军伟2   

  1. 1. 空军工程大学 航空等离子体动力学重点实验室,西安 710038;
    2. 西安空天机电智能制造有限公司,西安 710100
  • 收稿日期:2021-05-24 修回日期:2021-06-10 出版日期:2022-09-15 发布日期:2021-08-03
  • 通讯作者: 周鑫,E-mail:dr_zhouxin@126.com E-mail:dr_zhouxin@126.com
  • 基金资助:
    国家自然科学基金(91860136);广东省重点研发计划(2018B090905001);陕西省重大专项(2018zdzx01-04-01)

Machine learning of emission intensity signal of laser powder bed fusion molten pool

DUAN Yucong1, WANG Xuede1, ZHOU Xin1, ZHANG Peiyu1, GUO Xiyang1, CHENG Xing1,2, FAN Junwei2   

  1. 1. Science and Technology on Plasma Dynamics Laboratory, Air Force Engineering University, Xi 'an 710038, China;
    2. Xi 'an Aerospace Mechatronics & Intelligent Manufacturing Co., Ltd, Xi 'an 710100, China
  • Received:2021-05-24 Revised:2021-06-10 Online:2022-09-15 Published:2021-08-03
  • Supported by:
    National Natural Science Foundation of China (91860136); Key Research and Development Plan of Guangdong Province (2018B090905001); Key Science and Technology Project of Shanxi Province (2018zdzx01-04-01)

摘要: 粉末床熔融成形(PBF)的过程监测和质量识别是保障其制造质量的关键技术。在监测信号中,熔池辐射强度信号蕴含了丰富的熔池特征信息,但监测信号与材料科学现象的直接联系尚不明确,适合利用机器学习算法开展深入研究。首先,通过设置不同的工艺参数(部分偏离工艺窗口)实施316L不锈钢成形实验,成形过程中采集熔池辐射强度信号。然后,通过数据分割、特征提取和特征选择对信号数据进行预处理,构建了用于机器学习的数据集。最后,使用21种不同的机器学习算法,一是将熔池辐射强度数据按照工艺参数(如激光功率高、中、低)进行分类,经过训练的算法可在实际生产中识别异常(辐射强度异常、代表激光功率和扫描速度偏离最佳状态);二是将熔池辐射强度数据按最终成形块体的质量(密度、表面粗糙度)进行分类,经过训练的算法可在实际生产中识别质量。结果表明:对于熔池辐射强度的异常识别,二次支持向量机算法的分类效果最好,准确度达到96.4%以上;对于密度和表面粗糙度预测,由于样件质量与数据样本之间的复杂关系,预测结果呈现不同的分布情况,但预测准确度均达96.0%以上。

关键词: 机器学习, 粉末床激光成形, 熔池辐射强度, 异常识别, 质量识别

Abstract: Process monitoring and quality identification of Powder Bed Fusion (PBF) are the key technologies to ensure its manufacturing quality. Among the monitoring signals, the molten pool radiation intensity signal contains rich molten pool characteristic information, but the direct connection between the monitoring signal and the material science phenomenon is not clear, and it is suitable to use machine learning algorithms to carry out in-depth research. Through 316L stainless steel forming experiments, different process parameters (partial deviation from the process window) are set, and the emission intensity signal of the molten pool is collected during the forming process. The signal data are preprocessed through data segmentation, feature extraction and feature selection, and a data set for machine learning is constructed. Twenty-one different machine learning algorithms are used. The molten pool emission intensity data are firstly classified according to the process parameters (such as high, medium, and low laser power). The trained algorithms can identify abnormalities (abnormal emission intensity, representative laser power and scanning speed deviate from the best state). Secondly, the emission intensity data of the molten pool are classified according to the quality (density, surface roughness) of the final formed block, and the quality can be identified in actual production through the trained algorithms. The results show that for the identification of abnormalities of the emission intensity of the molten pool, the classification effect of the secondary support vector machine algorithm is the best, with an accuracy of over 96.4%. For prediction of density and surface roughness, the prediction results show different distributions due to the complex relationship between the sample quality and the data sample, but the prediction accuracy has reached more than 96.0%.

Key words: machine learning, laser powder bed fusion, molten pool emission intensity, abnormality identification, quality identification

中图分类号: