Prediction on the Distributions of the Strength and Toughness of Thick Steel Plates Based on Bayesian Neural Network

In the past decades, several artificial intelligence techniques have been applied on regression tasks during the steelmaking process, achieving accurate point estimations. However, models for predicting targets with their uncertainties have been rarely applied to the steelmaking domain due to a computational and a technical limit of Bayesian neural networks (BNN) in the past. Recently, many tractable BNN models have been proposed for predicting the uncertainties in steelmaking, thanks to advances in neural networks. The most convenient and computationally efficient one is a Monte‐Carlo dropout (MC dropout), which interprets the distribution of the neural networks created by a dropout from a Bayesian perspective. In this article, we propose to apply MC dropout on a steelmaking process data of Pohang Iron and Steel Co. and predict two targets and their uncertainties: the strength and the toughness of thick steel plates. This article is widely applicable to other process data, and it is expected to reduce costs for defective products in the future through an uncertainty estimation. The Monte‐Carlo dropout (MC dropout) model is applied to predict the physical properties of thick steel plates and their uncertainties. Experimental results on the Pohang Iron and Steel Co. data validate the ability of the MC dropout model. Based on the predicted target and the estimated uncertainties, it is expected to reduce costs for defective products.