Visual And Spectroscopic Based Quality Monitoring Of CMT Fusion Welding Processes

Arc welding process contains a large number of complex information.At the same time,there are many factors such as high temperature,arc light,spatter,soot and electromagnetic interference.The state of welding process determines the forming quality after welding,so the supervision of welding process is an important and complex task.With the development of intelligent manufacturing,online quality monitoring of fusion welding has become an important research content.This thesis has studied the cold metal transfer(CMT)weld process,and an online quality monitoring system is designed based on visual and spectral information to achieve efficient monitoring the welding state.The main contributions are listed as follows:(1)Weld pool image acquisition and contour extraction technology: a passive weld pool vision imaging system has been established,combined with trigger settings,to capture high-quality weld pool images.Aiming at the imaging characteristics of the weld pool image,a set of superpixel-based region merging and adaptive matting methods are designed to extract the weld pool contour.Superpixel segmentation and maximum similarity combination are combined to obtain the initial weld pool contour.The closed-form matting algorithm is used for fine edge segmentation to extract the accurate contour of the weld pool,which lays the foundation for obtaining the morphological parameters of the weld pool.By comparing the results with other algorithms,the effectiveness of the proposed method is verified.(2)Multi-source information fusion technology for quantitative monitoring of welding wire components: the camera and the spectrometer are triggered to collect arc information through the FPGA at the current peak time,and the image and spectral data are synchronized.A convolutional neural network is introduced into this task,and a multimodal network based on image and spectral feature fusion(ISFNet)is proposed.According to the characteristics of arc spectrum and image,improved Alex Net is used to extract spectral features and Res Net is improved to extract image features.Then the features are fused and the results are predicted.Combines with the improved normalized mean square error(NMSE)loss function training model,the quantitative monitoring of welding wire composition is realized.The predicted result of ISFNet is 0.4951,which is smaller than that of single source information model and other models,the effectiveness of the proposed method is verified.(3)Using the LabVIEW platform to design and develop a CMT welding process quality monitoring system,which has the function of parameter setting,real-time display of weld pool and arc information,multi task synchronous processing and analysis,and defect waning.