基于神经网络的喷丸25CrMo合金疲劳寿命及残余应力松弛行为预测研究

首先，本文采用BP神经网络建立了喷丸25CrMo车轴钢疲劳寿命预测模型。然后，在此基础上采用遗传算法（GA）对BP神经网络的预测精度进行了优化。此外，还采用了径向基神经网络（RBF）进行建模分析，并与以上两种模型的预测结果进行对比，结果表明：遗传算法优化的BP神经网络（GA-BP）相比于BP和RBF神经网络具有更高的预测精度，其中训练集和测试集的平均预测精度分别为91.5%和85.4%。然后，基于GA-BP神经网络模型的连接权值矩阵和Garson方程进行了灵敏度分析，从而进一步量化了输入影响因素对喷丸25CrMo车轴钢疲劳寿命的相对影响比重；最后，还采用GA-BP神经网络预测了喷丸25CrMo车轴钢表面残余压应力的松弛行为，结果表明：测试集的平均预测误差仅为3.4%，表明了该神经网络预测性能良好。综上所述，本文采用神经网络建模分析了喷丸25CrMo车轴钢的疲劳性能和残余压应力松弛行为，显著降低了传统疲劳试验所需的成本，并且还保证了较高的准确性。

Prediction of fatigue life and residual stress relaxation behavior of shot-peened 25CrMo axle steel based on Neural Network

Firstly, the fatigue life prediction model of shot-peened 25CrMo axle steel was established by using BP neural network. Then, genetic algorithm (GA) was used to optimize the prediction accuracy of BP neural network. In addition, radial basis function neural network (RBF) was used for modeling and analysis, and compared with the prediction results of the above two models. The results showed that GA-BP had higher prediction accuracy than BP and RBF neural network, and the average prediction accuracy of training set and test set were 91.5% and 85.4% respectively. Then, sensitivity analysis was carried out based on the connection weight matrix of GA-BP neural network model and Garson equation, so as to further quantify the relative influence proportion of the input influencing factors on the fatigue life of shot-peened 25CrMo axle steel. Finally, GA-BP neural network was used to predict the relaxation behavior of compressive residual stress on the surface of shot-peened 25CrMo axle steel. The results showed that the average prediction error of the test set was only 3.4%, indicating that the network prediction performance was good. In conclusion, this paper used neural network modeling to analyze the fatigue performance and compressive residual stress relaxation behavior of shot-peened 25CrMo axle steel, which significantly reduced the cost of traditional fatigue test and ensured high accuracy.