摘要
焊缝熔宽是评估焊接质量和焊接稳定性的重要指标。针对强噪声环境下的激光-MIG复合焊接过程,本文研究了基于反向传播(BP)神经网络补偿色噪声卡尔曼滤波算法的熔宽检测方法。首先建立激光-MIG复合焊缝熔宽检测系统的状态方程和测量方程,通过视觉传感和色噪声卡尔曼滤波算法对焊缝熔宽进行估计;然后采用高精度激光扫描仪对焊缝的三维轮廓进行扫描,根据二阶差分法获得焊缝轮廓宽度,并将其作为熔宽的真实值;接着将卡尔曼滤波增益、新息值和预测值与卡尔曼滤波最优估计之差作为输入,利用BP神经网络对熔宽的卡尔曼滤波最优估计进行补偿。结果表明:BP神经网络补偿测量色噪声卡尔曼滤波算法能够有效降低焊缝熔宽检测的误差。与单独使用卡尔曼滤波算法相比,BP神经网络补偿卡尔曼滤波算法具有更好的非线性映射能力,可以提高卡尔曼滤波焊缝熔宽检测的准确度。
Objective For decades,laser-are hybrid welding has gained remarkable attention as a reliable technology for material joint processing.It has been applied to various fields of the manufacturing industry due to several characteristics,such as deep penetration,high welding speed,and high-quality shaping.In the laser-arc hybrid welding process,the change in parameters may deeply influence the weld formation.To detect weld defects or monitor the quality of welding beads,several scholars have studied and explored the correlation between welding features and quality.Thus,numerous studies have investigated the relationship between metallic vapor features and molten pools.Among these features,weld width is a crucial evaluation criterion for welding quality and stability.It is commonly acknowledged that high-speed cameras are widely used to capture all types of features during laser-arc hybrid welding.This study presents an online detection of weld width based on the Kalman filter algorithm(BP-KF),which is compensated by a back-propagation neural network and can detect accurate weld width in a strong noise laser-MIG hybrid-welding environment.We assume that our innovative approach can provide the basis for online detection of a laser-arc hybrid-welding process.Methods The laser-MIG hybrid-welding detection system was established using a high-speed camera,arc welding machine,power fiber laser,and an image processing personal computer.During laser-arc hybrid welding,a high-speed camera was used to collect an image of a molten pool outline.Note that image processing is crucial for obtaining the width of the molten pool from an image.First,a molten pool area was defined and extracted by processing the sequential images emerging from the camera.Next,the end of the molten pool was identified by the difference of gray value in the image,and the keyhole was used to mark the position of the molten pool.After segmenting the image using a watershed algorithm,the width of the molten pool can be measured using the conversion from the pixel to the unit of distance.A high-precision laser scanner was used to scan the three-dimensional contour of the weld,and the width of the weld contour was obtained using the second-order difference method,which was used as the approximate true value of the weld width.According to the state and measurement equations of the laser-MIG hybrid welding width detection system,the weld width was estimated using visual sensing and colored measurement noise Kalman filter(KF)algorithm.Finally,the Kalman filter gain,new information,and difference between the predicted value and the Kalman optimal estimation were taken as the inputs.After obtaining the difference between Kalman optimal estimation and true weld width,the Kalman optimal estimation of the weld width was compensated by the BP neural network to improve the accuracy of weld-width detection.Results and Discussions Based on the comparison of the measured weld width and true values,both values with observable differences had the same variation tendency(Fig.6).To decrease the errors between the measured and true weld width values,the colored measurement noise Kalman filter algorithm was adopted to restrain errors from noise.However,the Kalman filter algorithm could not further eliminate these errors.Therefore,to enhance accuracy,the BP neural network was used to predict nonlinear errors caused by the fluctuation of the molten pool.After comparing the weld width values with true weld width,KF,and BP-KF,we observed that the values from BP-KF were generally closer to the true values than those from KF(Fig.7).The absolute errors between the true values and values from KF or BP-KF were calculated,BP-KF absolute errors were less than that of KF,which indicates that BP-KF can further decrease the errors between measurements and true values(Fig.8).Based on the abovementioned difference,the errors from the molten pool measurement,KF,and BP-KF were analyzed.It was crucial to demonstrate that weld width errors from BP-KF were less than others,such as max,mean-absolute,root-mean-square,and mean-absolute percentage errors(Table 1).Particularly,the weld width detected by BP-KF can satisfy the demands of manufacturers.Conclusions This study successfully adopts the watershed image segmentation approach to extract weld width in a strong noise laser-MIG hybrid-welding environmeit.Based on the relationship between the width of a molten pool and true weld width,colored measurement noise is recognized as a source of errors,and the Kalman filter algorithm is suitable for eliminating noise errors.In addition,complex fluxion of a molten pool leads to the nonlinear variation between the width of the molten pool and that of the welding bead,which is another source of errors during laser-MIG hybrid welding.Therefore,a BP neural network is chosen to predict the nonlinear difference between the width of Kalman optimal estimation and the true weld width,so that the errors caused by the fluctuation of the molten pool can be further restrained.Experimental results demonstrate that using a compensating colored measurement noise Kalman filter algorithm,which is compensated by the BP neural network,can reduce weld width detection errors better than other methods and can improve the detection accuracy.Compared with the Kalman filter algorithm,the BP neural network has a good nonlinear mapping ability,which can effectively improve the Kalman filter accuracy for weld width detection.
作者
刘秀航
黄宇辉
张艳喜
高向东
Liu Xiuhang;Huang Yuhui;Zhang Yanxi;Gao Xiangdong(Guangdong Provincial Welding Engineering Technology Research Center,Guangdong University of Technology,Guangzhou 510006,Guangdong,China)
出处
《中国激光》
EI
CAS
CSCD
北大核心
2022年第16期100-106,共7页
Chinese Journal of Lasers
基金
广州市科技计划(202002020068,202002030147)。