| With the arrival of intelligent era,all walks of life from artificial to intelligent direction to change.In the traditional welding field,welding quality detection,weld size measurement is its important content,the traditional manual detection method is no longer applicable,the use of machines to replace has become an inevitable trend,so research on a high efficiency,high precision welding quality detection method is very important.In this paper,aiming at the quality detection problem of ultrasonic assisted MIG welding,a welding and visual image acquisition platform was built,and combined with the laser assisted method,machine learning and deep learning methods were used to successfully achieve weld size prediction,weld size visual measurement and weld surface defect detection,providing technical guidance and application value for the size control and quality detection of the welding process.The following three parts are mainly realized:(1)According to the material characteristics and welding conditions of galvanized steel plate,an ultrasonic assisted MIG welding system and a laser-assisted welding image acquisition system were established.25 groups of data including the relationship between welding process parameters and weld size were obtained according to the orthogonal test design.Welding parameters were selected as welding current,welding voltage and welding speed.Weld size is the residual height,welding width of the two dimensional parameters.The model of mapping relationship between welding parameters and weld size was established by BP neural network,and the network structure was determined as 3-7-1.The weight and threshold of BP neural network are optimized by genetic algorithm for BP neural network which is prone to local minimization.Through the analysis of the training results of the model,it can be concluded that the BP neural network optimized based on genetic algorithm has better prediction effect and meets the requirements of weld size prediction.(2)This paper calibrated the visual image acquisition system,mainly camera and laser calibration,and carried out the operation of ROI region extraction,5×5 median filtering,threshold segmentation and closing operation on the collected laser-assisted weld images.The center line of laser fringe was successfully extracted by using Zhang’s fast parallel algorithm,and the weld feature points were extracted by using slope analysis method for laser fringe,which successfully obtained the transformation between the image weld geometric size information and coordinates.The maximum absolute error of the measurement of residual height and fusion width was only 0.06 mm and 0.09 mm.(3)In this study,the target detection model Retina Net is selected for detection of weld surface defects.In view of the shortcomings of welding process of galvanized steel plate,attention module SENet is introduced and embedded into backbone network Res Net-50 to form a new feature extraction module SE＿Res Net-50 to further extract useful information and features.After the top-down structure of pyramid structure,a bottom-up feature fusion path M module is added to effectively enhance the whole pyramid feature layer.The improved model was used to train the images of the collected welds,realizing the detection of five surface types of good welds,pores,edge bites,burnings and dents,and the m AP value reached 96.2%.In order to verify the effectiveness of the improved feature fusion module and attention mechanism SENet,we set three models: Retina Net original algorithm,retinanet-1(single improved feature fusion module)and retinanet-2(single added SENet attention module),and train them in the same environment.The experimental results show that the overall m AP of the two models combined retinanet-1 and Retinanet-2 is higher than that of the combined Retina Net model,further demonstrating the effectiveness of the improved module. |
PDF Full Text Request