The Study Of The Identification Of Critical Quality Characteristics Of Semiconductor Products Based On RF-RFE Algorithm

With the increasing number of electronic information products,users are placing higher demands on the performance of electronic products and this is placing more stringent standards on the reliability and lifetime of semiconductors and other basic quality characteristics,as well as placing higher demands on quality control.Product quality has always been regarded as the lifeblood of a product,and quality control and improvement exist at all stages of the product life cycle,but with the development of new technologies,traditional quality control methods are becoming less and less effective when dealing with relatively complex products such as semiconductors.In the face of a large number of quality characteristics,the selection of a subset of key quality characteristics is one of the problems facing complex quality control.This thesis addresses the problem of identifying critical quality characteristics for semiconductor products,with the following main research elements:(1)In view of the specificity of the quality data of semiconductor products,this thesis summarizes the characteristics of the quality characteristics set of semiconductor products,i.e.high dimensionality,small batch size,data imbalance,correlation between quality characteristics,etc.The miceforest algorithm is used to improve the random forest algorithm by using the grid search method after multiple interpolation of missing data values in this thesis;The RF-RFE algorithm is used to identify the key quality characteristics of the semiconductor quality dataset.To address the problem of data imbalance,the ADASYN method was used for data balancing operation,and the experimental results verified the effectiveness of the ADASYN method in data balancing;finally,the correlation test was carried out on the subset of candidate key quality characteristics using the Spearman correlation coefficient,and the quality characteristics with strong correlation were removed.The experimental results show that the feature subset of key quality feature identification conducted in this thesis can effectively reduce the dimensionality of the product quality dataset and has good performance in improving the performance of the model.(2)After identifying a subset of key quality features for this semiconductor product,this thesis proposes a final product quality level prediction model based on Stacking integrated learning.In this thesis,for the characteristics of the Stacking model,a total of five prediction models,K-nearest neighbour prediction model,random forest prediction model,Gradient Boosting prediction model,light GBM prediction model and decision tree prediction model,are used as base learners to construct the Stacking model,and after5-fold cross-validation,the random forest prediction model is used as a meta-learner to predict the final quality of the product.The experimental results show that the product final quality level prediction model proposed in this thesis outperforms the prediction effect of a single model in terms of performance,which verifies the effectiveness of this thesis for the product final quality level prediction method.