Platform For Predicting Mechanical Properties Of Welded Joints

The process of traditional welding procedure qualification has several shortcomings such as high cost and low efficiency, and high requirements of personnel, materials and equipments. More and more computer technology is applied to welding field along with the rapid development of computer assistant technology. Especially, predicting and modeling mechanical properties of welded joints is paid much more attention. In this paper, Artificial Neural Network (ANN) method has been applied to simulating and predicting mechanical properties for welded joints.First of all, on the basis of studying deeply into neural network algorithm, the paper has identified the chemical composition of parent metal and filler material, and welding parameters as network input parameters. At the same time, mechanical properties of welded joints were identified as output parameters. Hyperbolic tangent function connects the input layer and hidden layer, furthermore, linear function connects the hidden layer and output layer. Besides, the prediction accuracy of models was improved by several methods.Secondly, the platform for predicting mechanical properties of welded joints was developed. The system of this platform has two parts, one is responsible for training models with artificial neural network and the other one is for predicting mechanical properties with the models. In order to help ordinary welding technologists to train artificial neural network models by themselves the training system can process the sample files automatically, set network parameters intelligently, and combine models simply. In addition, the predicting system can manage models and predict mechanical properties conveniently.Additionally, the paper has collected and ruled out data from plenty of welding procedure qualification tests and built the corresponding data base. Mechanical properties of welded joints experiments were also carried for a typical steel Q345. As the main parameters, the welding current, welding voltage, preheat temperature and heat treatment temperature were concerned and 80 sets of SMAW experimental data were got. Having obtained a great deal of data sample from experimental results and literatures, the mechanical properties prediction models with different structures were established by training them with the system developed in the paper. The welding methods involved are including SMAW, SAW, GMAW and so on. The models were established based on actual data and considered the impact of the composition of the steels and welding parameters. Its practicality has greatly improved compared with previous neural network model reference to results of heat simulation.At last, all models were verified by the actual experimental data and theoretical knowledge and the effect of parameters to mechanical property was analysed. Consequently, accuracy of all models is within acceptable range and they can be used in production practice. Thereby, it was proved that the model training system designed in this paper had high training accuracy.