基于神经网络的焊接接头力学性能预测平台

Platform of mechanical properties predication for welded joints based on neural network optimized by genetic algorithm

基于神经网络建立焊接接头力学性能预测平台,可实现合金钢、钛合金及铝合金焊接接头力学性能预测。预测内容包括抗拉强度、屈服强度、伸长率以及断面收缩率等。同时可以分析参数变化对接头力学性能的影响。算法核心主要是应用遗传算法优化Back-Propagation神经网络连接权,其具有很好的全局搜索特性以及不易陷入局部最优化,同时应用既有高斯-牛顿法的快速收敛特性,也有梯度下降法的局部搜索特性的LM算法,使预测结果具有很好的泛化性能和较高的预测精度。

The platform of mechanical properties predication for welded joints based on neural network was established in this paper. It could be achieved to predict mechanical properties of welded joints of alloy steel,titanium alloy and aluminum alloy. The predictions included tensile strength,yield strength,elongation and section shrinkage,and simultaneously it could analyze the influence of parameters on the mechanical properties of joints. The core of algorithm was the application of genetic algorithm optimizes BP neural network connection weights. This algorithm was good at global search as well as was not easy involved into a local optimization,so it could make the predictions own a good generalization function as well as high predictive accuracy. The system also provided an alternative algorithm LM. It was the combination of Gauss-Newton algorithm with gradient descent algorithm. It owned self-learning neural network characteristics,rapid convergence characteristics of Gauss-Newton algorithm and local search characteristics of gradient descent algorithm,therefore,it had good generalization function and high predictive accuracy.