Vision Sensing And Intelligent Control Of Narrow Gap Hot Wire TIG Priming Full Fusion Welding

Welding penetration state is an important part of the welding quality constraints on the degree of welding automation.The ability to achieve the welding process penetration state of sensing,detection,analysis and feedback control is one of the key difficulties in determining the degree of automation and intelligent welding.Narrow gap hot wire TIG welding is widely used in the construction process of large pre ssure vessel storage tanks in the energy equipment industry,but due to the constraints of site construction conditions,the bevels to be welded often have large assembly errors and harsh working conditions.For a long time,skilled welders have been used to monitor and adjust the welding process in real time to ensure the melt-through state,but there are problems of high labor costs and low productivity.To address the above challenges and combined with the actual welding conditions of narrow gap hot wire TIG welding,a vision sensing device was built to sense,model and control the information of the back weld pool of narrow gap hot wire TIG vertical welding,and the following work was carried out by measuring the correspondence between the visual charact eristics signal of the back weld pool and the melt-through state under different penetration states.For existing narrow gap hot wire TIG vertical welding device with large frontal gun head,mechanism action on the welding pool obscured,frontal observation camera posture tilt,AC welding current zero point resulting in large changes in arc brightness,vertical welding conditions,gravity affects the shape of the pool,and other characteristics,taking into account the traditional frontal pool information sensing methods mostly require more complex sensing system,the device is larger and has more stringent requirements for the sensed pool shape and other restrictions,the use of back pool information sensing for penetration state information collection.In this way,the dorsal melt pool characteristic photoelectric sensing signal system and the dorsal melt pool characteristic visual signal sensing system were built,and their sensing effectiveness was verified through experiments,and the dorsal melt pool v isual signal sensing method was finally chosen to collect penetration state information.And by image size feature extraction algorithm,the image feature database and pre-processed image database are established.On the basis of the effective sensing melt pool feature information of the sensing system,two structured feature data models and two unstructured feature data models based on the different types of data features extracted by the sensing system were developed.First,the image features were extracted manually by image processing of melt pool images in different penetration states,and a binomial tree model was established to classify the extracted features into penetration states.After selecting the classification features for each node of the model and calculating the optimal classification threshold for each node,the final classification accuracy of the model was calculated and the classification accuracy was calculated to be 74.73%,and the model was poorly classified.An artificial neural network-based classification model was built based on the manual extraction of features from penetration images in different penetration states.Five layers were identified as the best layer depth in the network,and the network classification accuracy reached 85.1%,and the model classification performance was greatly improved compared with the binomial tree model.Later,two more types of unstructured data models were built.After image processing of the back melt pool images in different penetration states,a convolutional neural network-based penetration state classification model was constructed to model the end-to-end extraction of the back melt pool penetration image features,and the classification accuracy of the network reached 89.3% on the test set with good network generalization performance.To ensure the validity of the image features extracted by the convolutional neural network,a visual interpretability analysis of the network classification concern features was performed using the feature saliency map.The analysis test proves that the back melt pool picture features extracted by the network in different penetration states are all typical features in the current penetration state and correspond to the typical physical features in the penetration state.On the basis of the constructed convolutional neural network,a recurrent neural network was built by adding a recurrent layer before the output layer,and the temporal features of the penetration change process were extracted,and the network was trained by using the sequence pictures of the penetration state change,and the network classification accuracy reached 95.3%,and the model generalization performance was optimal.Finally,control tests were designed to verify the performance of the developed model.Built a fuzzy controller to enhance the binary tree model temporal feature extraction capability,and offline debugging of its performance using image datasets to improve the control system robustness.And the four controllers into the welding test system on the variable gap test plate for online welding test.The experimental results show that the recurrent neural network control system based on the extraction of unstructured image data and temporal features is the most effective for this welding process control,and obtains a fully fused controlled and stably formed weld seam.