Research On Real-time Algorithm Of Dynamic Identification Of Aluminum Alloy TIG Welding Pool Based On IOT Network Monitoring Platform

Internet of Things(IOT)technology is a typical technology that combines hardware with information technology.Industrial Internet of Things technology accelerates the intelligent process of manufacturing.Combining the Internet of Things technology is one of the ways for welding this traditional manufacturing technology to become intelligent.However,the general welding monitoring technology combined with the Internet of Things only collects information without specific real-time processing.Therefore,based on the IOT network monitoring platform to classify the penetration status,extract the visual information of welding,and establish accurate and low-latency models and control strategies for the back welding seam prediction,it is of great significance to realize the dynamic recognition of the molten pool during the welding process.As a widely used welding technology for aluminum alloy TIG welding,it is a very meaningful and promising subject to implement the dynamic identification of the welding pool based on the network monitoring platform to the aluminum alloy TIG welding.This paper participated in the construction of a robot welding platform based on IOT network monitoring.From the beginning of welding to the end of welding,various information obtained is transmitted to the network cloud in real time,and the model is called in the terminal software for information processing and modeling to realize the entire welding The process is monitored in real time to realize the whole welding process and lowlatency online monitoring.Based on the convolutional neural network algorithm for automatically extracting features,this paper builds a classic convolutional neural network model for the classification of aluminum alloy TIG welding pools.Through the process of classification accuracy change,it is found that it does not completely converge.The structure is optimized,and the network structure is deepened and optimized to accurately classify the weld pool.It is found that the loss of the convolutional neural network model decreases quickly,but it fluctuates greatly and is unstable.In the end,adding a more cuttingedge residual structure to compare and deepen the optimized model.It is found that the resnet18 model with the residual structure is better than the deepened and optimized model in terms of accuracy and convergence performance.The accuracy rate on the test set reaches 97.8%.Recall The rate reached 98.52%,and the F1 comprehensive score also reached 99.25%.However,visualizing the output features of the deep learning model found that the features automatically proposed by the deep model in the first few layers have a certain interpretability,but when reaching a deeper depth,the automatic features become abstract and their interpretability is general.In the final verification experiment,resnet18 has an overall accuracy of 91%on the verification experiment atlas.When the terminal software calls the trained model,the classification time is less than 30 ms per image,which is in line with the penetration classification of the IOT network monitoring platform The accuracy of the model is required for real-time.In view of the low interpretability of the convolutional neural network,for further research and control,the relevant parameters of the front molten pool must be extracted.First,the traditional image processing algorithm was tested,and it was found that it could only do one positioning,and its robustness was poor.An improvement to the cascade regression learning algorithm is proposed to extract the contour points of the molten pool and further extract the frontal molten pool parameters of the aluminum alloy TIG welding pool.Compared with traditional contour extraction algorithms,it is more accurate,more interpretable than complex algorithms,and has lower latency.The improved cascade regression algorithm proposed is based on the ESR algorithm to scientifically reduce the range of contour points to improve the timeliness of the algorithm.The processing time for each image is only 11.2ms.The average relative errors of the final extracted molten pool area,molten pool width,and molten pool length were only 2.07%,-8.7% and 9.68% in the verification analysis.Meet the timeliness and accuracy requirements of IOT network monitoring in the visual information processing of terminal software.Based on the front-side molten pool parameters and welding parameters,XGBR,a variant of Xgboost,is used to perform regression prediction on the backside welding width,and compare and evaluate with other prediction models SVR and KNN.On the test set,the XGBR predicted RME is only 1.53%,the RMSE is only 0.44 mm,and the R-squared score is close to 1.In the verification experiment,the average relative error of XGBR was only 3.67%.The traditional PID control strategy used to control the back welding width is optimized,and the Alearning algorithm is added to reduce the steady-state error of the traditional PID and reduce the adjustment time.The most important thing is to use AClearning to achieve better automatic tuning of the three PID parameters.On the one hand,it realizes overall automation,and on the other hand,it realizes stable control under the entire IOT network monitoring platform.