Weld Shape Prediction And Parameters Optimization Of Double Wire MIG Rapid Prototyping System Based On Neural Network

With the development of market economy and the intensification of competition in product manufacturing,rapid prototyping technology has attracted much attention due to its advantages of high flexibility,short production cycle and excellent performance of forming parts.As a kind of rapid prototyping,arc rapid prototyping uses arc hot melt welding wire to accumulate metal parts according to the trajectory.It is flexible in manufacturing and has high density in forming parts,and it has been widely used in the field of mechanical manufacturing.The developed double wire MIG rapid prototyping system not only has the advantages of high depositing rate of coarse wire,but also has the characteristics of high forming precision of fine wire.In order to guide the weld forming and optimize the welding process during welding,so as to reduce the test and save the welding cost.In this research,neural network models are established to visually predict the weld morphology and optimize the welding parameters.Firstly,sample data are collected.The research uses the orthogonal design method to design the test plan.The welding speed,wire diameter,wire feed speed,and welding voltage are used as test factors.The interaction between the various factors is not considered in the test.Because the number of factors and level is not the same and there is no orthogonal table for direct selection,so redesigned the mixed orthogonal table,at the same time,defined the test scheme.Experiments were carried out using double wire MIG rapid prototyping equipment and obtained 80 samples.After the samples are processed,the weld images which can observe the weld morphology is obtained under the microscope.The geometric parameters of the weld,such as weld width,residual height and penetration depth,are collected and recorded in the table for subsequent training.Then the influence of four test parameters in the orthogonal table on the weld formation in the double-wire MIG rapid prototyping system is analyzed.The rule of the influence of test parameters on the weld formation is preliminarily explored,which provides a theoretical basis for the subsequent prediction data,and further verifies the stability of the designed welding system.In order to achieve the purpose of direct output weld shape,the research used the key point coordinates and curve equation categories to determine the upper and lower contour curves of welding.This article uses the Python language and Tensorflow machine learning framework to write programs,and uses the Tensorboard visualization tool to build three types of neural network models.The first type of network model predicts the forming parameters of the weld and converts it to key point coordinates.The second and third types of neural network models were established to predict the upper and lower contour curves of the weld.Before training the second and third types of network models,the weld seam images data collected in chapter two were extracted and classified to make one-hot labels for training.The results show that the loss curves of the three types of network models are stable and the accuracy of the second and third types of network models are 80.7% and 83.8%,respectively.Finally,the prediction model of reverse welding process is established.And the welding parameters are optimized in combination with the first type of forward network model.The results show that the prediction error can be less than 5%.For the convenience of using,a visual interface was created using Matlab software.The interface has three modules: network hyper-parameter information,prediction interface and data management.Finally,the interface was used to predict the weld shape of fifteen sets of test data,and the data of weld width,residual height and penetration depth in the prediction process are recorded.The prediction performance in the first type of network model is better.The accuracy of upper and lower types of weld seam in test set are 0.8318 and 0.8780 respectively,which is better than that of training set.The results shows that the generalization performance of the second and third types of network is stronger.The image area is obtained by the pixel point calculation method.Compared with the actual weld area and the predicted weld area,the average relative error of the fifteen sets of data is 5.16%,it shows that the prediction models established in this paper have high reliability.