中文摘要：

为降低冷滚打花键表面粗糙度，获得冷滚打加工最优参数组合，以滚打轮公转转速和工件进给量两个影响表面粗糙度的主要因素作为变量，设计了冷滚打花键及测量实验方案，采用白光共聚干涉显微镜测量冷滚打花键分度圆处表面粗糙度，依据实验数据通过试凑法建立了冷滚打花键表面粗糙度BP神经网络预测模型，最终确定的神经网络结构为2-6-2-1，对预测值与训练样本值及测试样本值进行了对比分析，结果表明：预测值与训练样本最大误差6．5％，与测试样本最大误差7．9％，预测值与训练样本之间的相关系数为0．996，与测试样本之间的相关系数为0．973，进一步说明了神经网络预测模型的有效性和精确性。

英文摘要：

In order to reduce the surface roughness of cold roll-beating spline and to get the optimal parameter combination, rota- ting speed of roller revolution and feeding rate of workpiece two main factors affecting surface roughness as variables, the cold roll-beat- ing spline and experimental project are designed. Surface roughness of cold roll-beating spline＇s pitch circle was measured through white light copolymerization interference microscope. The Back Propagation （BP） neural network prediction model for surface roughness of cold roll-beating spline was established based on the experimental data through trial and error method. The optimal neural network structure 2-6-2-1 was determined. The predicted values and the training samples and testing samples were contrasted and analyzed. The results show that the maximum error between the predicted values and the training sample is 6. 5% and the maximum error between the predicted values and the training sample is 7. 9%. The correlation coefficient between the predicted values and the training samples is 0. 996 and the correlation coefficient between the predicted values and the testing samples is 0. 973. The validity and accuracy of neural network prediction model are further illustrated.