Topside Visual Sensing And Deep Learning Algorithm Based Prediction Of Keyhole Status/Penetration In Plasma Arc Welding

As a high energy-density welding method,keyhole plasma arc welding(K-PAW)can join medium-thickness plates in one pass without any groove.The plasma arc can penetrate through the molten pool and form a keyhole in the weld pool.The keyhole stability is sensitive to the changes of process parameters.Real-time detection of keyhole status is prerequisite for effective control of welding process.Both the indirect detection and backside visual sensing of the keyhole status have their own shortcomings.In this study,visual sensing was applied from the topside to capture the whole weld pool,plasma arc and keyhole entrance.Dynamic information of the weld pool and keyhole entrance was extracted.Based on the deep learning algorithm and obtained topside weld pool images,the keyhole status and penetration was predicted.This research would lay the foundation for real-time measuring and control of the keyhole status in PAW process.A novel visual sensing system was developed.Composite filtering system wasdesigned to overcome strong arc disturbance,Topside weld pool could be clearly observed from a small angle along the welding direction.Accordingly,to analysis of weld pool characteristic,image processing algorithm was applied to draw the outlines of weld pool with constrained experience parameters.The edge information of the three frame processed weld pool image,arc minimum circumscribed curve were used as guide information to improve the algorithm.Then the weld pool boundaries can be successfully tracked.It's shows that the proposed algorithm is robust and adaptable compared to the traditional edge search algorithms.Detection of the topside weld pool from the side of the workpiece perpendicular to the welding direction was put forward for capturing the undivided topside weld pool.The specially designed dimmer class and bandpass filter helped to distinguish the weld pool and welding arc boundaries the same time.Boundaries of the complete topside weld pool were acquired through image processing of the captured raw images.Optimized Canny operator with 5 constraints used for tracing the boundary lines could avoid the noise disturbances.A synchronous visual sensing system with dual CCD cameras was designed.The keyhole entrance and topside weld pool geometries were visually measured simultaneously.By image registration,topside weld pool and keyhole entrance images were successfully fused into a cohesive whole.The Topside weld pool and keyhole entrance dimensions were extracted.With the varied welding parameters,keyhole entrance geometries varied with larger amplitudes than the topside weld pool geometries.Clear Images of the topside weld pool and keyhole exit were acquired through a synchronous visual sensing system.The relationships between the weld arc,weld pool and keyhole were investigated under different welding processes.It was found that The reflex arc in the topside weld pool showed different forms whenever it was in blind keyhole,steady keyhole and growth of keyhole.This could be used to reflect keyhole and penetration status.Due to the complex nonlinear relationship between the topside weld pool features and keyhole status,the key features extraction from the images was the precondition for the keyhole and penetration status prediction.The deep learning convolution neural network was used to extract the feature information automatically from the topside weld pool images.The topside weld pool image was taken as the input of the model and is used for predicting the penetration after the full training.Experiments were carried out to establish the training database containing 12000 positive weld pool images and relative keyhole images.10000 images were taken as the training data,1000 images were used as the verified data,1000 images was taken as the test data.A convolution neural network model which contains five convolution layers and three full connection layers was established.After training,the model can automatically obtain the good characteristics of input weld pool image for classification and identification.The test results showed over 90%prediction accuracy of keyhole/penetration status.