基于RFECV-LightGBM-TPE联合模型的在制品质量问题预测方法研究  

Research on Quality Problem Prediction Method of WIP Based on RFECV-LightGBM-TPE Joint Model

作　　者：陈浩威 战洪飞[1] 林颖俊 余军合[1] 王瑞 CHEN Haowei;ZHAN Hongfei;LIN Yingjun;YU Junhe;WANG Rui(Faculty of Mechanical Engineering and Mechanics,Ningbo University,Ningbo Zhejiang 315211,China;Zhongyin(Ningbo)Battery Limited Company,Ningbo Zhejiang 315040,China)

机构地区：[1]宁波大学机械工程与力学学院,浙江宁波315211 [2]中银(宁波)电池有限公司,浙江宁波315040

出　　处：《机械设计与研究》2023年第5期115-122,共8页Machine Design And Research

基　　金：国家重点研发计划资助项目(2019YFB1707101,2019YFB1707103);国家自然科学基金资助项目(71671097);浙江省公益技术应用研究计划资助项目(LGG20E050010,LGG18E050002);宁波市自然科学基金资助项目(2018A610131);健康智慧厨房浙江省工程研究中心。

摘　　要：为对在制品进行质量预测,提出一种基于RFECV-LightGBM-TPE联合模型的在制品质量问题预测方法。该方法通过RFE递归特征消除和CV交叉验证进行特征筛选,确定特征维度。针对实时采集的在制品工艺参数不充分的情况,通过结合特征重要性的方法改进KNN缺失值填补算法,补全未采集的工艺参数,解决在制品难以进行质量预测的问题。同时利用历史数据训练LightGBM算法构建质量预测模型,并通过TPE算法进行参数寻优提高预测精度。最后以博世生产线采集的工艺参数为例将本文所提模型与常见模型进行对比,结果表明本文所提模型可以更好的对质量问题进行预测,给智能工厂的建设提供了参考价值。A method of predicting the quality of a product in process based on the RFECV-LightGBM-TPE model is proposed.The feature dimension is determined by RFE recursive feature elimination and CV cross validation.Aiming at the situation of insufficient process parameters collected in real time,the KNN missing value filling algorithm is improved by the method of feature importance to complete the uncollected process parameters,so as to solve the problem of difficult quality prediction.At the same time,the LightCBM algorithm is trained with historical data to build a quality prediction model,and parameter optimization is carried out by the TPE algorithm to improve the prediction accuracy.Finally,taking the process parameters collected from the Bosch production line as an example,the model proposed in this paper is compared with the common models.The results show that the model proposed in this paper can better predict the quality problems and provide reference value for the construction of intelligent factory.

关 键 词：在制品 质量预测 RFECV LightGBM TPE 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]