Prediction Of Weld Penetration Based On Multi-sensor Information

During the fusion welding process,the penetration is the key indicator of welding quality.Many failures of the welding structure are due to insufficient or incomplete penetration of the weld.Therefore,it’s very important to detect the penetration of the weld.At present,the general detection method of weld penetration destructive test.Cutting the section of welding to observe the macroscopic metallography.It’s impossible to achieve in engineering.The adept welding worker can adjust the welding parameters in real time according to the sound of the electric arc and the shape of the weld pool to ensure the weld depth is qualified.Therefore,the prediction model of weld penetration can be built by collecting the sound of the electric arc,the images of molten pool and welding electrical signal.In this paper,the experimental platform collecting welding data is designed and built,achieved the synchronous acquisition and preservation of welding current,arc voltage,the sound of the electric arc and the images of molten pool.The relationship between the signal collected by the welding data acquisition platform and the weld penetration is analyzed.Moreover,the characteristic values of the signal are extracted.Based on the results predicted by the multivariate linear regression,the support vector regression and the BP neural network,the BP neural network model displayed minimum average absolute error compared with the other two models and the error value is only 1.1%.The resurfacing welding process is carried out Q235 base materials with a thickness of15 mm using the method of CO₂ gas shielded welding.The BP neural network is used to calculate the weld penetration of the flat surfacing and the value of the average relative error only up to 1.4%.With the increase of welding current,the penetration of the weld increases.With the increase of arc voltage,the penetration of the weld decreases.With the increase of welding speed,the penetration of the weld decreases.With the increase of gas flow rate,the penetration of the weld decreases.In addition to the above welding parameters,the space position of the welding gun also has a great influence on the penetration of the fillet weld.Therefore,the fillet weld experiment is carried out under the condition of the same material and welding method to investigate the relationship between the penetration of the fillet weld with the spatial position of the welding gun.Besides,the position of the welding gun in 3D space depends on the angle between the welding gun and the base plate,the centering position of the welding gun and the tilting angle of the welding gun.Therefore,the accuracy of the fillet weld penetration prediction model can be optimized by adding the space position of welding gun to the predict model as input node and combining with Dropout’s improved BP neural network model.The results show that the average relative error of the side plate penetration a,the root penetration b,and the bottom plate penetration c can be controlled within 3.8%.