Gmaw Welding Quality Monitoring Based On CNN And Molten Pool Image

"Made in China 2025" puts forward the key development areas of offshore engineering equipment and high-tech ship key manufacturing equipment.With the gradual advancement of "Industry 3.0" to "Industry 4.0",new demands are placed on the online inspection technology of welding automation and welding process.The traditional weld quality inspection methods are mostly post-weld inspections,and the real-time performance is bad.It is impossible to obtain the molten pool information when defects occur,and it is difficult to make up for defects.Based on convolutional neural network algorithm and welding pool image visual inspection method,studied the on-line monitoring of quality parameters such as penetration state,surface porosity and surface slag during GMAW process,which lays a foundation for online quality monitoring of GMAW process.A molten pool visual sensing test system was established.The influence of welding specifications on the image characteristics of MAG welding pool is analyzed.The welding test was carried out under different welding specifications,and about 15,000 molten pool images were obtained,and the influence of welding specifications on the characteristics of the molten pool image was extracted: the penetration state was greatly affected by the welding current,and the generation of surface pores and slag was related to the shielding gas flow.In order to obtain a clear image of the front molten pool,a filter system was designed to take the molten pool from the rear of the workpiece at an angle along the welding direction.The image processing methods such as de-equalization,normalization,and gamma transformation are used to reduce the problem of metal splash and strong arc interference,and improve the feature recognition effect of the molten pool.The classification model is trained by convolutional neural networks to eliminate strong arc interference.The camera calibration method is analyzed,the camera pose is solved to obtain the shooting angle,and after affine transformation,a clear front molten pool image is obtained.Based on the convolutional neural network algorithm,from the obtained positive molten pool image,the characteristics of the molten pool sensitive to the welding specification in the molten pool are analyzed.The visual convolution kernel output found that the convolution check was very interested in the black hole in the middle of the weld pool during the weld.When the weld produces surface vents,the convolution nucleus responds strongly to round black holes in the semi-solidified region of the tail of the weld pool.When a large amount of slag is generated on thesurface of the weld,the convolution core strongly responds to the boundary line between the molten pool metal and the slag.The convolution algorithm is studied,and different convolution algorithms are designed for various defect features.In the process of training,the automatic parameters and manual adjustment methods are used to continuously optimize the hyperparameters and model structure parameters.The prediction accuracy of the model is above 95%.Study the mixed programming scheme of C++ and Python.By integrating the camera's function functions,OpenCV image processing functions,and Python interface functions,functions including image acquisition,attribute setting,image saving,image processing,and predictive classification are realized.Predict the penetration state of the molten pool during the welding process,and predict whether welding defects such as weld penetration,surface pores,and slag are generated,and the test speed is about 20 frames every second.