激光熔化沉积过程缺陷识别方法

doi:10.12068/j.issn.1005-3026.2024.08.011

摘要/Abstract

摘要：

激光熔化沉积加工过程中的缺陷萌生是制约激光熔化沉积技术发展的关键性问题.实现对缺陷的精确自动识别是提高激光熔化沉积技术应用水平的重要途径.提出了熔池瞬态特征提取算法，分析了熔池瞬态特征对沉积层熔合不良缺陷的影响关系，建立了熔池瞬态特征数据集.对主流识别算法进行了模型训练测试，获取了相对最优模型ResNet 34.为解决ResNet 34训练损失拟合效果差、计算速度慢的问题，结合传统卷积网络和LSTM（long short?term memory）网络，建立了训练和测试精度高且计算速度快的LRCN 64模型，测试准确率达95.8%，实现了对熔合不良缺陷的识别，为实现沉积件在线无损检测提供了技术支撑.

关键词: 激光熔化沉积, 熔池瞬态特征, 熔合不良, 长期循环卷积神经网络（LRCN）, 残差神经网络（ResNet）

Abstract:

Defects in laser melting deposition are key problems restricting its development. Achieving precise automatic identification of defects is a crucial approach to enhance the application level of laser melting deposition technology. A novel algorithm for extracting the melt pool’s transient characteristics was presented， and the relationship between transient characteristics and lack of fusion defects of the deposition layers was found. Moreover， a dataset of the melt pool’s transient characteristics was established. The mainstream recognition algorithms were trained and tested， leading to the identification of the most effective model， ResNet 34. In order to solve the poor fitting training loss effect and slow calculating speed of ResNet 34， a hybrid LRCN 64 model was proposed combining the traditional convolutional networks and LSTM（long short?term memory） networks. It exhibited remarkable accuracy and significant calculating speed. The testing accuracy of the LRCN 64 model reaches 95.8%， thereby realizing the identification of lack of fusion defects， which provides valuable technical support to facilitate online non?destructive testing of deposited parts.

Key words: laser melting deposition, molten pool transient characteristics, lack of fusion, long?term recurrent convolutional neural network （LRCN）, residual neural network （ResNet）

中图分类号:

TU 621

刘伟嵬, 刘炳君, 刘焕强, 刘泽远. 激光熔化沉积过程缺陷识别方法[J]. 东北大学学报（自然科学版）, 2024, 45(8): 1150-1158.

Wei-wei LIU, Bing-jun LIU, Huan-qiang LIU, Ze-yuan LIU. Defect Identification Method for Laser Melting Deposition Process[J]. Journal of Northeastern University(Natural Science), 2024, 45(8): 1150-1158.

图/表 20

图1 同轴送粉式半导体激光熔化沉积加工系统

Fig. 1 Coaxial powder?delivering semiconductor

表1 同轴送粉式半导体激光熔化沉积加工系统设备

Table 1 Equipment of coaxial powder?delivering semiconductor laser melting deposition system

设备名称	型号
半导体激光器	LDF 4000-100 VGP
六轴机器人	KUKA KR30HA-KRC4
激光熔覆头	PRECITEC Cladding Head YC52
送粉器	煜宸RC-PGF-D双筒载气式送粉器
冷水机	MCWL-150T-01AK1S4精密冷水机

表2 316L不锈钢粉末化学组成（质量分数） (steel powder (mass fraction) %)

Table 2 Chemical composition of 316L stainless

C	Mo	Ni	B	Cr	Si	Fe
0.06	2.28	11.78	0.78	18.51	0.99	余量

表3 实验工艺参数

Table 3 Experimental process parameters

序号	激光沉积密度	激光功率	扫描速度	z轴提升量	激光光斑直径
序号	J·mm^-2	W	mm·s^-1	mm	mm
1	40	320	4	0.5	2
2	45	360	4	0.5	2
3	50	400	4	0.5	2
4	55	440	4	0.5	2
5	60	480	4	0.5	2
6	65	520	4	0.5	2
7	70	560	4	0.5	2
8	75	600	4	0.5	2
9	43	430	5	0.5	2
10	48	480	5	0.5	2
11	53	530	5	0.5	2
12	58	580	5	0.5	2
13	63	630	5	0.5	2
14	68	680	5	0.5	2

图2 熔合不良缺陷显微图像

Fig. 2 Micro image of lack of fusion defects

图3 第5组实验熔池图像（a）—第1层加工图像；（b）—第3层加工图像；（c）—第5层加工图像.

Fig. 3 The fifth set experiment melt pool images

图4 不同熔融状态下的熔池几何轮廓（a）—正常熔融；（b）—过度熔融；（c）—不完全熔融；（d）—不连续熔融.

Fig. 4 Geometric contour of molten pool in different molten states

图5 特征提取算法流程图

Fig. 5 Flowchart of the feature extraction algorithm

图6 LRCN模型

Fig. 6 LRCN model

图7 图像增强

Fig. 7 Image augmentation

图8 增强数据集分布

Fig. 8 Distribution of enhanced datasets

图9 模型训练与验证流程

Fig. 9 Model training and validating process

表4 传统模型参数与运行结果 (traditional models)

Table 4 Parameters and running results of the

模型	学习率	训练轮次	批数据大小	训练准确率/%	测试准确率/%
VGG 11	0.001	100	5	84.1	93.9
VGG 16	0.001	100	5	87.8	94.6
ResNet 18	0.001	100	5	88.4	96.3
ResNet 34	0.001	100	5	89.3	97.9

图10 VGG网络训练与测试结果（a）—VGG 11；（b）—VGG 16.

Fig.10 Training and testing results of VGG network

图11 ResNet网络训练与测试结果（a）—ResNet 18；（b）—ResNet 34.

Fig. 11 Training and testing results of ResNet network

表5 LRCN模型参数与运行结果

Table 5 Parameters and running results of LRCN models

模型	学习率	训练轮次	批数据大小	隐藏层大小	训练准确率/%	测试准确率/%
LRCN 32	0.001	100	5	32	98.1	93.7
LRCN 64	0.001	100	5	64	99.3	95.8
LRCN 128	0.001	100	5	128	99.1	94.4
LRCN 256	0.001	100	5	256	99.4	94.9

图12 LRCN模型训练与测试结果（a）—LRCN 32；（b）—LRCN 64；（c）—LRCN 128；（d）—LRCN 256.

Fig. 12 Training and testing results of LRCN model

图13 混淆矩阵图（a）—ResNet 34；（b）—LRCN 64.

Fig. 13 Confusion matrix diagram

表6 模型识别结果

Table 6 Model recognition results

熔合情况	ResNet 34			LRCN 64
熔合情况	准确率	灵敏度	特异度	准确率	灵敏度	特异度
近似圆形熔合不良	0.962	0.977	0.983	0.953	0.931	0.98
矩形熔合不良	0.98	0.951	0.994	0.949	0.912	0.985
不规则熔合不良	0.975	0.987	0.994	0.919	1.0	0.98
熔合良好	1.0	1.0	1.0	1.0	1.0	1.0

图14 模型计算时间对比

Fig. 14 Comparison of model calculation time

参考文献 20

1	Thompson S M， Bian L K， Shamsaei N，et al.An overview of direct laser deposition for additive manufacturing，part I：transport phenomena，modeling and diagnostics［J］.Additive Manufacturing，2015，8：36-62.
2	Shamsaei N， Yadollahi A， Bian L K，et al.An overview of direct laser deposition for additive manufacturing，part II：mechanical behavior，process parameter optimization and control［J］.Additive Manufacturing，2015，8：12-35.
3	彭谦，董世运，闫世兴，等.激光熔化沉积成形缺陷及其控制方法综述［J］.材料导报，2018，32（15）：2666-2671，2682.
	Peng Qian， Dong Shi‑yun， Yan Shi‑xing，et al.An overview of defects in laser melting deposition forming products and the corresponding controlling methods［J］.Materials Review，2018，32（15）：2666-2671，2682.
4	Svetlizky D， Das M， Zheng B L，et al.Directed energy deposition（DED） additive manufacturing：physical characteristics，defects，challenges and applications［J］.Materials Today，2021，49：271-295.
5	Guo Q， Chen S Y， Wei M W，et al.Formation and elimination mechanism of lack of fusion and cracks in direct laser deposition 24CrNiMoY alloy steel［J］.Journal of Materials Engineering and Performance，2020，29（10）：6439-6454.
6	Nursyifaulkhair D， Park N， Baek E R，et al.Effect of process parameters on the formation of lack of fusion in directed energy deposition of Ti‑6Al‑4V alloy［J］.Journal of Welding and Joining，2019，37（6）：579-584.
7	Bustillos J， Kim J， Moridi A.Lack of fusion in additive manufacturing：defect or asset？［EB/OL］.（2021-04-14）［2023-01-16］..
8	Bustillos J， Kim J， Moridi A.Exploiting lack of fusion defects for microstructural engineering in additive manufacturing［J］.Additive Manufacturing，2021，48：102399.
9	Wang W J， Ning J Q， Liang S Y.Prediction of lack‑of‑fusion porosity in laser powder‑bed fusion considering boundary conditions and sensitivity to laser power absorption［J］.The International Journal of Advanced Manufacturing Technology，2021，112（1）：61-70.
10	Seifi S H， Yadollahi A， Tian W M，et al.In situ nondestructive fatigue‑life prediction of additive manufactured parts by establishing a process‐defect‐property relationship［J］.Advanced Intelligent Systems，2021，3（12）：2000268.
11	Tang M， Pistorius P C， Beuth J L.Prediction of lack‑of‑fusion porosity for powder bed fusion［J］.Additive Manufacturing，2017，14：39-48.
12	唐梓珏.激光熔化沉积熔池动态特征演化行为及关键性质量预测研究［D］.大连：大连理工大学，2020.
	Tang Zi‑jue.Study on evolution of molten pool dynamic characteristics and prediction of key product qualities in laser melting deposition［D］.Dalian：Dalian University of Technology，2020.
13	唐梓珏，刘伟嵬，颜昭睿，等.基于熔池动态特征的金属激光熔化沉积形状精度演化行为研究［J］.机械工程学报，2019，55（15）：39-47.
	Tang Zi‑Jue， Liu Wei‑wei， Yan Zhao‑rui，et al.Study on evolution behavior of geometrical accuracy based on dynamic characteristics of molten pool in laser‑based direct energy deposition［J］.Journal of Mechanical Engineering，2019，55（15）：39-47.
14	Song L J， Bagavath‑Singh V， Dutta B，et al.Control of melt pool temperature and deposition height during direct metal deposition process［J］.International Journal of Advanced Manufacturing Technology，2012，58（1）：247-256.
15	Wang L， Zhang Y M， Chia H Y，et al.Mechanism of keyhole pore formation in metal additive manufacturing［J］.NPJ Computational Materials，2022，8（1）：143-153.
16	Gargalis L， Ye J C， Strantza M，et al.Determining processing behaviour of pure Cu in laser powder bed fusion using direct micro‑calorimetry［J］.Journal of Materials Processing Technology，2021，294：117130.
17	Donahue J， Hendricks L A， Guadarrama S，et al.Long‑term recurrent convolutional networks for visual recognition and description［J］.IEEE Transactions on Pattern Analysis and Machine Intelligence，2017，39（4）：677-691.
18	Lu J， Behbood V， Hao P，et al.Transfer learning using computational intelligence：a survey［J］.Knowledge-Based Systems，2015，80：14-23.
19	Wen L， Li X Y， Gao L.A transfer convolutional neural network for fault diagnosis based on ResNet-50［J］.Neural Computing & Applications，2020，32（10）：6111-6124.
20	Chaib S， Liu H， Gu Y F，et al.Deep feature fusion for VHR remote sensing scene classification［J］.IEEE Transactions on Geoscience and Remote Sensing，2017，55（8）：4775-4784.

[1]	刘伟嵬，邓剑洋，张靖文，牛东东. 基于深度学习的挖掘机工作阶段的分类与识别[J]. 东北大学学报（自然科学版）, 2023, 44(10): 1464-1474.
[2]	杨玉莹，巩亚东，屈硕硕，蔡明. 增/减材复合加工316L宏观形貌和残余应力试验[J]. 东北大学学报:自然科学版, 2020, 41(3): 380-386.