摘要
焊缝熔透状态是定量评价激光焊接质量最重要的指标之一,实时准确识别焊缝的熔透状态是动态激光焊接过程监测和控制的关键。针对铝合金薄壁件的脉冲激光焊接,本文提出了基于声波时频特性和深度学习的焊缝熔透定量评估新方法。首先,搭建视觉-声发射多信息实时同步传感系统平台,获取反映匙孔动态行为的视觉图像和声波信号,并对声波信号进行分帧和小波包阈值去噪预处理;然后,使用平滑伪魏格纳维利分布(SPWVD)提取各帧声波信号的时频域图像,同时引入灰度共生矩阵(GLCM)提取时频域纹理特征,并将提取的纹理特征送入反向传播神经网络(BPNN)进行预测;最终,以SPWVD声波时频图为原始输入,构建基于卷积神经网络(CNN)的焊缝熔透分类模型。结果表明:声波时频图与匙孔动态行为、焊缝熔透状态具有高度相关性;相比于准确率为85%的传统BPNN分类模型,基于SPWVD时频图的CNN分类模型有着更高的准确率(98.8%)。所提定量评估新方法为铝合金薄壁件脉冲激光焊接熔透的在线智能诊断与自适应控制提供了参考。
Objective Laser welding processes with high energy density,precision,and efficiency are widely employed in the automotive,aerospace,and medical industries.They have significant advantages in joining materials,such as aluminum alloys,stainless steel,magnesium alloys,and dissimilar metals.The welding penetration state is one of the most critical indicators for the quantitative evaluation of laser welding quality.The precise identification of the weld penetration state in real-time is a key bottleneck in monitoring and controlling dynamic laser welding processes.The complex physical-chemical metallurgical interaction between the laser beam and the metal workpiece releases intense optical,thermal,and acoustic radiation.The acoustic information is derived from the thermal vibration under a high heat input,and the pressure shock wave generated when the keyhole is internally stressed.Its acoustic characteristics(sound pressure amplitude and frequency characteristics)are closely related to the state of the keyhole.In this paper,we present a new method for quantitatively assessing weld penetration state based on acoustic time-frequency characteristics and deep learning for pulsed laser welding of thinwalled aluminum alloys.This study will contribute to research on the high correlation between acoustic information and weld penetration state in laser welding.Methods First,as shown in Fig.1,a visual-acoustic-emission multi-information real-time synchronous sensing system was developed to acquire visual images and acoustic signals reflecting the dynamic behavior of the keyhole,and the acoustic signals were preprocessed using frame splitting and wavelet-packet threshold denoising methods.Second,a smoothed pseudo-Wigner-Ville distribution(SPWVD)was used to extract time-frequency domain images from each frame of acoustic signal,and a gray-level cooccurrence matrix(GLCM)was introduced to extract the time-frequency domain texture features and feed them into the backpropagation neural network(BPNN)for prediction.Finally,a convolutional neural network(CNN)-based weld penetration state classification model was established using the SPWVD acoustic time-frequency map as the original input.Results and Discussions During the preprocessing of the acoustic signal,wavelet-packet threshold denoising effectively filters out some burrs in the signal,and the denoised signal is framed in one pulse period,as shown in Fig 2.Second,the time-frequency maps of the acoustic signals extracted using the SPWVD method exhibit significant differences in the texture features with different penetration states,as shown in Fig 6.Here,the four texture features of the SPWVD time-frequency maps extracted from the GLCM show a clear trend as they change from non penetration to partial penetration and then to full penetration states.Finally,we constructed GLCM-BPNN and SPWVD-CNN classification models and compared the advantages and disadvantages of both classification models.Despite the high correlation between the acoustic time-frequency map and the dynamic behavior of the keyhole and welding penetration,the CNN classification model based on the SPWVD time-frequency map shows a higher accuracy(98.8%)than the traditional BPNN classification model(85%).This indicates that the deep learning model based on the SPWVD timefrequency map as a direct input to the CNN model yields improved recognition results.Conclusions(1)The preprocessing method of acoustic signals using frame splitting and wavelet packet thresholding can effectively intercept useful segments and obtain a signal with good denoising results.(2)The SPWVD method extracts a time-frequency map of the pulsed laser welding acoustic signal.The texture features of SPWVD time-frequency map is highly correlated with the dynamic behavior of the laser-welded keyhole and the weld penetration state.(3)The SPWVD-CNN deep learning weld penetration state recognition model has a classification accuracy exceeding 98%.The proposed model provides a new approach and technical path for reliable monitoring of the thin-walled aluminum alloys laser welding process.
作者
罗钟毅
吴頔
王润
董金枋
杨方毅
张培磊
于治水
Luo Zhongyi;Wu Di;Wang Run;Dong Jinfang;Yang Fangyi;Zhang Peilei;Yu Zhishui(School of Materials Science and Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;Shanghai Collaborative Innovation Center of Laser Advanced Manufacturing Technology,Shanghai 201620,China;School of Materials Science and Engineering,Shanghai Jiaotong University,Shanghai 200240,China;Han’s Laser Technology Industry Group Corporation Ltd.,Shenzhen 518052,Guangdong,China)
出处
《中国激光》
EI
CAS
CSCD
北大核心
2023年第8期31-40,共10页
Chinese Journal of Lasers
基金
国家自然科学基金(51905333)
中国博士后科学基金(2021M692039)
上海市青年科技英才扬帆计划(19YF1418100)。
关键词
激光技术
脉冲激光焊
声波时频特性
熔透评估
平滑伪魏格纳维利分布
卷积神经网络
laser technique
pulsed laser welding
acoustic time-frequency characteristics
penetration assessment
smoothed pseudo Wigner-Ville distribution
convolutional neural network
