基于大数据的力学性能预测与工艺参数筛选

神经网络数据样本的质量影响着模型的预测精度；选择有效的工艺参数进行建模可以提高模型的训练速度，节省训练时间。针对数据样本质量和输入工艺参数的选择问题进行了研究，采用了计算马氏距离的方法剔除异常点，改善数据样本的质量。基于采集到的热轧汽车大梁板的17个工艺参数的生产数据，采用贝叶斯神经网络建立力学性能预测模型。通过采用平均影响值筛选出对力学性能影响较大的工艺参数进行建模，以简化模型。结果显示：简化后的模型取得了较高的预测精度，对于抗拉强度和屈服强度，分别有96·64％和94·96％的数据预测值和实际值相对误差在±6％以内；对于伸长率，有96·64％的数据预测值和实际值的绝对误差在±4％以内。

Prediction of mechanical properties and process parameters selection based on big data

Predicted accuracy of the neural network model is mainly influenced by the quality of data samples.It is important to construct the model with effective process parameters,which can improve the prediction accuracy and the training speed.Sample quality and the method for selecting process parameters were investigated by calculating the Mahalanobis distance to remove abnormal points and improve the quality of data samples.According to 1 7 pa-rameters collected from the hot rolling automobile beam plate production,Bayesian neural network model was used to construct model to predict the mechanical properties.In order to simplify the model,mean impact value was used to choose the effective process parameters which have great influence on the mechanical properties.The results show that the simplified model has a high predicted accuracy.There are 96·64% and 94·96% data of which the relative error between predicted values and measured values are less than ±6% for tensile strength and yield strength,respectively.For elongation rate,there are 96·64% data of which the absolute error between predictive values and measured values was less than ±4%.