Temperature Prediction of Friction Stir Welding Based on Bayesian Neural Network

This paper established a neural network based on Bayesian regularization, which was used for temperature prediction of feature points in friction stir welding metal. Sample data included rotation speed, welding speed, distance, depth, time and temperature of feature points. The representative sample data was obtained through the orthogonal experiments. Neural network was improved by Bayesian algorithm, and the constraint term representing complexity of network was introduced into the objective function. Then generalization ability of neural networks has been enhanced. Prediction accuracy, training time and stability of network are better than in the traditional neural network. Test proved that this smaller network model operates faster, and it can accurately predict the temperature of feature points of friction stir welding.