Research On Monitoring Technology Of Welding Condition Of All Position Pipelines Based On Arc Acoustic Signal

Welding is one of the most important techniques in the manufacturing industry,which plays an important role in the rapid development of our economy.In the process of welding,it is an effective way to realize intelligent welding to use arc acoustic signal to monitor the welding status in real time.Therefore,the method of welding condition monitoring of arc sound at all positions of pipeline is proposed.First of all,the general requirements for the establishment of the welding status system are described,the technical route is planned,and the hardware platform and software interface of the welding status monitoring system are designed.The system is composed of arc sound acquisition system,motion control system and upper computer software.The arc sound acquisition system can realize the collection,display and storage of arc sound signals.The motion control system mainly carries out the welding and motion functions of welding robots,and uses upper computer software to process and analyze the collected signals and identify the welding status.The key techniques and algorithms used in this paper are introduced.Secondly,the arc acoustic signal noise reduction processing and characteristic analysis,welding gun swing signal plays an auxiliary role.The noise reduction effect of hard threshold,soft threshold and improved threshold algorithm is compared.The denoising evaluation index is used to compare the hard threshold and soft threshold.The improved threshold algorithm is improved by 0.04 and 0.07 respectively.The time domain,frequency domain and MFCC characteristics of the signal after arc acoustic noise reduction are analyzed.Different welding states show different rules.Then,a database of 6 welding states including excellent welding quality,porosity,penetration,biting edge,left skew and right skew was established.A neural network model based on Inception module was established,MFCC features were taken as model input,local feature extraction and learning training of multiple convolutional layers and pooling layers were carried out,and accuracy and cross entropy loss functions were used as model evaluation indexes.The accuracy of support vector machine regression(SVR),BP neural network and Inception module neural network in the training set is 83.22%,91.87%,96.84%,and the cross entropy loss function in the training set is 90.85%,96.90%,98.67%.The accuracy rate in the test set is 84.81%,89.12% and 97.16%,and the cross entropy loss function in the test set is 91.47%,96.83% and 98.33%.It is proved that Inception module neural network has a certain stability and generalization,and the accuracy of Inception module neural network is obviously better than SVR and BP neural network in welding state recognition.Finally,a Win Form based host machine arc sound state monitoring software is developed,which integrates the functions of signal acquisition,display,storage,noise reduction,signal analysis and welding status recognition and monitoring.Through the experimental verification of the welding status monitoring system,the practicality of the software is verified.Inception module neural network can achieve high precision identification.