Abstract: To improve the accuracy of tool wear monitoring for small samples in the machining process,a Tent—Atom Search Algorithm—Back Propagation neural network (Tent-ASO-BP) tool wears prediction model based on Pearson+Gradient Boosting Decision Tree (Pearson+GBDT) feature extraction was proposed.Aiming at the problem of feature selection and parameter selection of the BP neural network,a two-layer filtering feature screening method based on Pearson+GBDT was proposed to obtain the network input features,and the optimal weights and thresholds of the BP neural network were solved by using Tent chaotic map to improve ASO.The experimental verification showed that ASO was improved by Tent chaotic mapping and was prevented from falling into local extreme value and premature convergence.The cross-validation proved that the training model of Tent-ASO optimized BP neural network had higher accuracy than ASO.At the same time,it verified that GBDT could screen a group of features used for tool wear value mapping,and the feature screening ability was stronger than the similar algorithms LightGBM,Catboost,decision tree,and random forest.

Key words: tool wear, Pearson correlation coefficient, gradient boosting decision tree, tent—atom search algorithm—back propagation neural network

摘要： 为了提高机械加工过程中小样本刀具磨损量监测的准确性,提出一种基于Pearson+GBDT特征提取、Tent混沌映射和原子搜索算法（ASO）优化BP神经网络（Tent-ASO-BP）的刀具磨损量预测模型。针对BP神经网络特征选择及参数选择难题,提出了基于Pearson+GBDT的双层过滤式特征筛选方式求取网络输入特征,并使用Tent混沌映射改进原子搜索算法（ASO）对BP神经网络最优权值和阈值进行求解。通过实验证明：Tent混沌映射改善了ASO,避免ASO陷入局部极值和过早收敛,即通过交叉验证证明Tent-ASO优化BP神经网络训练模型精度较ASO高。同时,验证了梯度提升决策树（GBDT）能够筛选出用于刀具磨损值映射的一组特征,且特征筛选能力强于同类算法LightGBM、Catboost、决策树、随机森林。

关键词: 刀具磨损量, Pearson相关系数, 梯度提升决策树, Tent-ASO-BP网络