Research On Machine Learning Methods For Welders’ Operational Skills For Horizontal Welding Pool Control

Welding is used in many areas of industrial production,automated welding has become the development trend of the welding industry,the current automated welding can still only carry out assembly line operations and simple welding tasks,its intelligence level is low,welding robots can only achieve in a specific environment to control the molten pool.Industrial production is developing towards high precision,intelligence and efficiency,and it is difficult for welding robots to meet the gradually increasing industrial requirements,and the lack of welders and the harsh welding environment has forced welding robots to engage in more areas of production,but their welding quality does not meet the requirements and cannot replace manual welding operations.The level of intelligence of welding robots severely restricts the development of the automated welding sector.The control of the melt pool is the key to ensuring the quality of the weld.As the real-time state of the melt pool during the welding process is uncertain,it is difficult to establish a mathematical model to analyse the flow process due to numerous factors.The study of welders’ control of the melt pool in the cross-welding process,the welders observe the real time melt pool state,obtain melt pool information and adjust the torch attitude based on welding experience to achieve control of the melt pool.In order to fully explore and utilise the welders’ experience-based behavioural mechanisms,i.e.observing the molten pool and analysing the molten pool information to adjust the welding parameters in real time,this paper uses a machine learning approach to analyse the welders’ real time transverse welding process and establish a model of intelligent welding for human welders,analysing how the welders use their experience to control the transverse weld pool,and investigating the welders’ operational behaviour towards transverse weld pool control from the following aspects.(1)To find senior welders to perform cross-welding experiments,the manual welding process image acquisition was set up to capture molten pool images through vision sensors and used as the raw data set,this process is the basis for analysing the operating behaviour of cross-welders.(2)In order to study the welders’ operational behaviour in cross-welding and to find out how the welding gun changes according to the changes in the melt pool,it is necessary to extract the characteristics of the welders’ operations.To do this,a welders’ transverse welding process image acquisition platform is built,and the captured melt pool images are processed to extract the features of the melt pool and the features of the welding gun.The original image was affected by arc light interference and noise,so the image was pre-processed using median filtering to remove pretzel noise and extract the region of interest(ROI),then image enhancement to highlight the melt pool contours and edge detection operators to extract the melt pool edges.The welding gun is more seriously disturbed by the arc light,so the contour of the melt pool and the contour of the welding gun need to be extracted separately,and finally the centre point of the melt pool and the centre point of the welding gun are obtained,i.e.the position of the welding gun and the position of the melt pool.(3)As the position of the welding gun changes according to the position of the molten pool at the previous moment,there is a temporal relationship between the two.The powerful sequential learning capability of the Long and Short Term Memory(LSTM)network of machine learning is used to build an intelligent welding model for welders based on LSTM.The model takes the melt pool position information extracted from the cross-weld melt pool image as the input to the model,and the welding gun position information as the output of the model.As the LSTM model built is a multiple input to output,the network is more complex and the Grey Wolf Optimisation algorithm(GWO)is designed to optimise the network and improve the accuracy of the model.(4)The GWO was improved to build an IGWO-LSTM intelligent welding model for welders.The improved model outperformed the GWO-LSTM model and also compared the error of back propagation neural network(BP)and gated recurrent neural network(GRU)in the test set,the IGWO-LSTM model had the lowest root mean square error(RMSE)in the test set.The IGWO-LSTM model was able to resolve the change in torch position with melt pool position and predict the torch trajectory during the welders’ transverse welding process,initially achieving intelligent welding and laying the foundation for intelligent welding.