Allocating Orders to Printing Machines for Defect Minimization: A Comparative Machine Learning Approach

Abstract

Zero-Defect Manufacturing (ZDM) is continuously emerging as the most critical target of the Industry 4.0 era. Minimization of the defected products in the industrial production chain can contribute towards significant improvements of the operational costs, the production efficiency and speed, as well as the environmental footprint. This work proposes a Machine Learning (ML) based scheme for intelligently and proactively allocating orders to printing machines, so as to ensure minimization of the defected products. Based on a historical dataset extracted by a printing company, ten supervised learning regression models were trained to estimate the machine-specific defect ratio of new orders, given their multi-feature requirements. To optimize the machine selection policy, several widely-used ML schemes were compared in terms of their performance on unseen data samples. Extensive simulations were carried out to stabilize the hyperparameters of the ML models, including single-model, ensemble and Deep Learning based regressors. Results showed that Multi-Layer Perceptron (MLP) outperform the rest of the benchmarking regressors in accurately predicting the defect rate of each machine, with the ensemble methods presenting also enhanced accuracy. Finally, the misclassification ratio of the proposed algorithm was assessed by quantifying the number of optimally allocated orders to printing machines, exhibiting that the majority of orders are correctly classified.