Cost-imbalanced hyper parameter learning framework for quality classification. (1st January 2020)

• Record Type: Journal Article
• Authors: Zhang, Yunchao; Li, Yu; Sun, Zeyi; Xiong, Haoyi; Qin, Ruwen; Li, Chen
• Is Part Of: Journal of cleaner production. Volume 242(2020)
• Publication Date: 2020-01-01
• Subjects: Quality classification -- Cost-imbalanced -- Hyper-parameter learning -- Machine learning -- Decision tree -- Particle swarm optimization; Factory and trade waste -- Management -- Periodicals; Manufactures -- Environmental aspects -- Periodicals; Déchets industriels -- Gestion -- Périodiques; Usines -- Aspect de l'environnement -- Périodiques; 628.5
• DOI: 10.1016/j.jclepro.2019.118481
• Languages: English
• ISSNs: 0959-6526

Abstract

Abstract: A quality control system is an indispensable section in various manufacturing and service industries. It plays a critical role in reducing process flaws, optimizing process parameters, improving production quality and productivity, as well as enhancing customer satisfaction. In this paper, we propose an intelligent data-driven quality classification platform by leveraging a novel integrated hyper learning framework to further strengthen the cost-effectiveness in quality control by reducing the economic loss due to misclassification. The misclassification-dependent weights are proposed and used for training the classifier with an emphasis on cost-effectiveness. The proposed integrated hyper learning framework is used to optimally identify such weights. Specifically, the framework consists of two nested layers, where the inner-layer addresses the optimal classifier training with a given set of misclassification weights, while the out-layer updates such weights iteratively according to the performance in terms of the economic loss due to misclassification by the classifier identified by the inner-layer towards optimality. The case studies are implemented using five different datasets in different manufacturing and service industries, including food, auto, steel, and glass. The economic loss, as well as additional carbon emission due to misclassification when using the quality classifier identified through the proposed framework, is compared to three other algorithms under different settings of penalty costs due to misclassification. The results illustrate that the proposed intelligent data-driven quality classification platform outperforms the other ones in terms of the reduction of the economic loss due to misclassification and demonstrate the robustness of the performance with respect to various misclassification penalty costs. As for the carbon emission reduction, the proposed model can outperform, in most cases, the three other algorithms. While the consistency of this superiority cannot be guaranteed since the environmental concern is not modeled in the objective function. Highlights: A framework for optimizing hyper parameters in machine learning model is proposed. The framework is built by a nested model using PSO and decision tree. The effectiveness of the framework is validated using real industrial data. The monetary loss due to misclassification in quality classification is improved. The emission due to misclassification in quality classification are reduced.