Markov Chain Monte Carlo-based Manufacturing Process Control Algorithm: An Application to Steel Plate Production

Abstract

In steel plate production, the final product is said to be in good quality if three mechanical properties - yield strength, tensile strength, and elongation - are close to their prespecified target values. In this study, we propose a probabilistic manufacturing process control algorithm for the following inverse problem: the algorithm is designed to suggest field experts a set of values of process parameters to control in advance, in order to achieve prespecified target properties of the final product as closely as possible. We formulate this problem as the estimation of the conditional density of the process parameters to control given target properties and the other process parameters. Unlike many existing algorithms for the inverse problem, it allows us to quantify the uncertainty of the proposed process condition and compare it with those of alternatives. For its inference, Markov chain Monte Carlo (MCMC) algorithm is considered to let practitioners flexibly model the complex physical relationship between process variables and mechanical properties. Additional hyperparameters are introduced to take reliability of fitted prediction models, which may have been trained based on data corrupted with uncontrollable noises, into consideration. We apply the proposed method to the manfucaturing process data from one of the largest steel manufacturing companies in South Korea. Furthermore, an evaluation metric is introduced to evaluate the performance of the proposed algorithm.