Online Welding Stability Monitoring Based On Image And Spectral Feature Learning

Intelligent welding is a key research in the industrial field,the purpose of which is to let the machine learn the thinking of welding workers.Applying deep learning to the field of intelligent welding,although there has been relevant research at home and abroad,still lacks a certain degree of universality.According to this circumstance,this paper studies the application of weld pool image and arc spectrum,and applies deep learning network to welding field,and achieves good results.The main research contents of this paper are as follows:(1)Welding stability monitoring based on weld pool image and arc spectrum.Aiming at the defects of the existing weld pool image acquisition,which mainly focuses on the single band,this paper designs a dual optical path imaging system by analyzing the distribution of the weld pool radiation and arc spectral radiation,so as to capture clear weld pool image with small arc interference.The dual-band images are separated by adaptive segmentation,and the accurate weld pool contours are obtained after combination.Based on the contours,the modified Lenet network models are trained to monitor the welding speed under different current.A transient spectrometer based on Hadamard code is developed to obtain high spectral-to-noise ratio arc spectral data in a larger field of view.Based on the arc spectral,the modified Lenet model is trained to monitor flow rate of protection gas.(2)Welding stability monitoring based on multi-source feature fusion.Aiming at the defects that the existing weld monitoring network can only deal with a single source data and unable to deal with multi-source data,a network model that fuses the weld pool image and the arc spectrum is proposed.Through the evaluation criteria,this paper analyzes the advantages and disadvantages of the proposed two multi-input network configurations,and select the network configuration with better adaptability to image and spectral data.Furthermore,the basic network is further optimized to improve the overall performance of the network.Compared with the single-source network,the fusion network has great advantages in recognition rate and so on.For the real-time requirements of welding monitoring,the speed optimization is carried out by using TensorRT engine.