The prediction model of CVN for welded joint based on neural network and genetic algorithm

Based on the welding process parameters of high strength pipeline steel, the artificial neural network (ANN) model has been developed to predict Charpy_V notch (CVN) impact toughness of the welded joint. The model with back propagation (BP) algorithm is built in batch mode and optimized using momentum and adaptive learning rate. Futhermore, it is also optimized using genetic algorithm (GA). Five process parameters, namely the welding layer, wall thickness, the welding processes, the preheat temperature, the mean energy input, are used as input variables and the CVN of the welded joints is considered as the output variable. The training and testing of the ANN model have been done using 119 datasets which were obtained from practical welding. The number of the testing samples with error less than 20% is about 74% in total testing data. It is found that the CVN of pipeline welded joints can be effectively predicted using the model and the GA outperforms the commonly BP algorithm used as an neural network training technique. Based on this model, the influence of the preheat temperature and the mean energy input on the CVN is also analysed. The results show that the preheat temperature and the mean energy input have little effects on the automatic welding, but large on the semi-automatic welding and the manual welding for root welding. So a reasonable choice of the preheat temperature and the mean energy input is necessary.