A Novel Approach To Product Quality Control In Industry Based On Ensemble Learning

Product quality control in industry is an important means to obtain the stable quality of industrial products.Most of traditional methods on industrial product quality control are proposed to analyze the influence of data with potential information and value on quality indicators independently.However,the characteristics of data involved in the process of manufacturing production are dirty,mixed variants(e.g.,data type and category type),unrepresentative and imbalanced.Independent analyses on these data are difficult to figure out the internal relations between data.In this paper,we convert the industrial product quality control problem to a machine learning problem.Our research is built on manufacturing product quality control,which is divided into on two aspects: key quality indicators predictions and optimized adjustable parameters recommendation in different stages of the process of production.We propose an effective ensemble learning based approach for manufacturing product quality control,and the main innovations of this paper can be summarized as follows:1)For the data imbalance problem,we propose to address this difficult problem on the data level and the algorithm level.On the data level,we apply up-sampling and modified down-sampling strategy to adjust the imbalance distribution of training data.On the algorithm level,we borrow the idea of cost-sensitive learning methods and Boosting ensemble learning algorithms,and propose a cost sensitive extreme gradient boosting(CSXGBoost)based algorithm.2)For the key quality indicators prediction problem,we take advantages of ensemble learning and present a simple yet effect multi-model ensemble based method,which includes Random Forest,XGBoost and DART models.3)As to optimized adjustable parameters recommendation problem,we search the best adjustable parameters combination based on maximizing the corresponding key quality indicators and the similarity measurement of the features of fixed parameters.Extensive experiments demonstrate that our proposed approach can achieve superior performance on predicting key quality indicators in different production stages of manufacturing under evaluation criterion of smaller RMSE,and can work well on recommending optimized adjustable parameters of manufacturing data.We believe that our proposed approach is effective and practical enough,which can be applied as useful guidelines in product quality control.