Prediction Of Mechanical Properties Of Welded Joints Based On Machine Learning

The mechanical property of welded joints is one of the important indicators for testing the quality of welded products.In order to ensure the quality of welded products meets the requirements of welding engineering,the traditional method is to test the mechanical properties of joints through a large number of welding procedure qualification tests,which has high cost and low efficiency.This paper proposes to use the machine learning method to predict the mechanical properties of welded joints,so as to reduce or even replace the welding procedure qualification test.First of all,through in-depth study of machine learning algorithm theory,this paper designs a set of model training algorithm based on BP neural network for the prediction of small sample dataset modeling of welded joint mechanical properties.And propose to use the K-fold cross-validation method and the grid search method to train and optimize the model.In addition,based on the dataset of TIG welding process and joint property tests for TA15 titanium alloy,the algorithm design of the application design is trained to obtain the joint tensile strength prediction model and the yield strength prediction model with excellent generalization performance,and the effectiveness of the algorithm is verified.Secondly,based on the neural network model training algorithm,this paper designs a set of sample incremental learning algorithm flow,which realizes the model dynamic learning new data and real-time update.Based on the titanium alloy GTAW welding data samples,70% of the dataset were randomly selected for basic model construction,and the test error of the optimal tensile strength prediction model was 7.51%.Then,the other 30% of the original data set is used as incremental learning supplementary data,and the original model is optimized for incremental learning.The model test error after incremental learning optimization is 5.59%,which significantly improves the model prediction accuracy.The designed incremental learning algorithm is practical and effective by the actual welding data verification.Finally,based on the Python programming language,this paper designs and develops a new type of welded joint mechanical performance prediction system with independent intellectual property rights.The system consists of five subsystems,namely database and model library management system,model training system,model prediction system,model incremental learning system and user rights management system.The system can not only realize the storage,management and sharing of enterprise welding data and models,but also assist the welding engineer to train and optimize the model to predict the mechanical properties of the welded joints.In order to further verify the reliability of the system,based on the TA15 titanium alloy TIG welded joint hardness test data set,the joint hardness prediction model was obtained in the system.