Research On Welding Pool Detection Technology Of Pipeline All-position Welding Based On Machine Vision

Welding plays an important role in the field of petrochemical industry,arc welding process has complex physical information,in which the dynamic behavior of the weld pool directly affects the weld formation,determines the stability of the welding process and welding quality,and expensive professional weld pool camera is not conducive to the popularization of construction site.Therefore,this paper takes all-location pipeline as the research object to detect the dynamic behavior of welding pool,study the relationship between welding deviation and characteristics of welding pool,detect the surface morphology of welding pool in real time,establish neural network to predict weld quality,and dynamic tracking of welding pool,so as to optimize welding process,improve welding quality and reduce manual operation load.Firstly,the basic theories and technologies of enhanced denoising,semantic segmentation methods and evaluation indexes of deep learning models are analyzed,the experimental scheme of this paper is designed,the hardware selection of the dynamic behavior detection system of molten pool is carried out,reasonable optical lenses are selected to construct the filter system according to different light bands,and the molten pool shape is captured dynamically.The weld pool image was calibrated and preprocessed to provide the basis for the subsequent image processing.The characteristics of the weld pool were defined,the mechanical phenomenon of the weld pool at all positions was analyzed,the forming characteristics of the weld pool at all positions were studied,and the welding deviation measurement model was established based on the trailing Angle of the weld pool.Secondly,the measurement principle is tested.The processing flow is as follows:guided filter image enhancement,Otsu algorithm arc extraction,image subtraction,weld pool edge extraction,trailing-sampling point extraction,gradient descent fitting,welding deviation measurement and deviation warning.The constructed system collects molten pool images in real time,uses convolutional neural network to extract morphological features,realizes the prediction of porosity defects in the welding process,builds a deep training model,and uses real-time molten pool images as input to fully learn and classify and further establish the relationship model between molten pool images and porosity defects by the network,so as to predict welding defects.Adjust process parameters in time.Finally,the improved generation antagonistic network(GAN)was used to de-noise and enhance the molten pool image in the welding process,to solve the phenomenon of image blur caused by shielding gas and arc light,and to provide an image basis for molten pool tracking.Using the dynamic segmentation method of molten pool based on the improved PSPNet model,the dynamic tracking and accurate detection of molten pool are realized by constructing multi-layer information extraction network structure,obtaining global information by cavity convolution with large field of view,and modifying activation function.A set of upper computer molten pool visual detection software integrating image acquisition control,system calibration,image processing,defect prediction and molten pool dynamic tracking was developed.In order to verify the function and stability of the welding molten pool visual detection system,a detection system platform was built to conduct welding experiments and result analysis.The method proposed in this topic can detect the weld pool in real time.The key point is to realize the real-time tracking based on the dynamic behavior of the weld pool,welding deviation warning,welding pool behavior monitoring,welding pore defects prediction,etc.,so that the process parameters can be adjusted in time.The system controls the cost of automatic welding equipment,and has a good application prospect in the field of petrochemical industry.