Font Size: a A A

Data Driven-based Monitoring And Prediction Of Laser Welding Process State

Posted on:2024-05-26Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z H ZhangFull Text:PDF
GTID:1521307334976369Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
With the continuous increase of non-standard and customized products in China’s manufacturing industry,how to quickly deploy welding tasks and ensure welding quality has become a bottleneck problem in the welding field.This problem drives welding mode towards flexible,refined,and intelligent advancement.Laser welding technology is an emerging material connection technology with advantages such as concentrated heat,small heat-affected zone,and non-contact processing.It is easy to achieve automation and integration of direct manufacturing on the product ion line.However,the complexity of laser welding technology is due to the combined effect of parameters such as materials,welding position,laser power,and optical path,which restricts the promotion of t his technology.Therefore,exploring the mechanism of the laser welding process,monitoring the process status,predicting future state,and ultimately providing theoretical and data support for the regulation of welding parameters is the basis and inevit able requirement of this technology towards intelligent manufacturing.This thesis focuses on the monitoring and prediction of the process state in laser welding.Based on the theory of laser welding and combined with digital twin theory and data-driven methods,a systematic study is conducted on the mon itoring and prediction of the penetration state,identification and prediction tracking of the weld position from the four aspects of "mechanism-data-characteristic-model".The main objective is to provide theoretical guidance for ensuring welding quality and promoting the development and application of intelligent laser welding equipment.The main research work of this paper is as follows:(1)In response to the demand for intelligent laser welding equipment,t his paper provides a comprehensive analysis of the basic theory and structure of digital twin from the perspective of the process layer.A five-dimensional model of laser welding digital twin is proposed,including laser welding system,virtual welding,tw in data,services,and connections.To meet the requirements of online monitoring in laser welding,a laser welding system is developed.The physical model of virtual welding is based on the laser welding molten pool gas-liquid interface tracking model,taking into account the influence of welding power and scanning speed on the heat source model.The sources and acquisition methods of twin data are analyzed,providing a theoretical basis and data source for data-driven laser welding state monitoring and prediction.(2)In response to the difficulty of online monitoring of the melt-pool state during laser welding,this study determined the key features and judgment criteria for the melt-pool state based on the behavior of the small holes in the melt pool,th e morphology of the melt pool,small holes,an d through-holes,as well as the postwelding weld morphology.The melt-pool state was divided into four categories:partial penetration,moderate penetration,fully penetration,and excusive penetration.A training set was established for training the convolutional neural network(CNN).The CNN was trained to recognize welding images and achieve online monitoring of the melt-pool state.The effects of the CNN’s structure and hyperparameters on its latency,accuracy,and generalization ability were comprehensively considered.Experimental results showed that the CNN not only recognizes the melt-pool state but also identifies the welding direction and the direction of the auxiliary light source.From the perspective of accuracy and melt-pool dynamics,dividing the meltpool state into three categories is more suitable for online monitoring and prediction than dividing it into two or four categories.The accuracy of the melt-pool state divided into three categories is 97%.(3)To address the issue of the lag effect of monitoring on the status of laser welding and its impact on online monitoring,a data-driven modeling method for predicting the melting state is proposed.By analyzing the dynamic behavior of the molten pool,it is found that the melt ing state conforms to the Markov chain property.Simulated data and empirical data are fused to enhance the sample set and drive the initialization of the prediction model.A melting state prediction model is proposed based on a Gaussian process Markov chain.Experiments are conducted on different plate thicknesses,power levels,and scanning speeds to verify the feasibility of the melting state prediction model.The results show that the combination of virtual and real data can be used for the initializati on of the melting state prediction model,and after online learning of the welding process in the physical world,the mean error and root mean square error of the prediction model are reduced.The computational cost of the developed prediction model is much lower than that of traditional numerical simulations,meeting the requirements of online prediction.(4)A method for online monitoring and parameter calibration of laser welding seams is proposed to address the problem of unstable and indeterminate cent er position of welding seams due to material defects and strong light interference in actual workpieces.Firstly,a welding seam information extraction method based on image processing algorithm is designed to extract the welding seam position and width.Secondly,an artificial neural network-Kalman filter is proposed to eliminate monitoring errors and reduce measurement noise.The parameter calibration in the welding seam information extraction is a tedious process.Therefore,the calibration problem is transformed into an optimization problem.A genetic algorithm-based parameter calibration method is proposed to automatically determine the optimal parameters in the welding seam monitoring.Finally,laser welding seam monitoring tests are conducted in both virtual welding environments and actual systems.Experimental results show that the proposed welding seam monitoring method has smaller errors than the method based on the Hough transform.The automatic calibration method can ensure the stability of weldin g seam monitoring under complex and changing welding conditions.(5)To address the problem of small field of view and unstable tracking in active coaxial visual monitoring,a welding seam position prediction and tracking method is proposed.An online Gaussian process is used to predict the future welding seam position and its uncertainty based on historical and current positions.The predicted position is used in the model predictive control method,and the proposed welding seam prediction and tracking method is evaluated and tested in a virtual welding environment.The effectiveness,accuracy,and robustness of the proposed method are verified through welding experiments.The model predictive control can calculate the optimal control quantity online when t racking the welding seam.Experimental results show that the control quantity calculated online by the proposed prediction and tracking method is lower than that of the PID algorithm,and the welding seam tracking process is stable and accurate.
Keywords/Search Tags:Laser welding, data driven, penetration state monitoring, penetration state prediction, seam monitoring, seam predictive tracking
PDF Full Text Request
Related items