Bayesian neural networks for predicting quality in reclaimed waste sand for foundry applications

Abstract

Although advancements in smart manufacturing technologies have profoundly transformed the manufacturing industry, their application in traditional industries remains challenging. In particular, the casting industry faces significant obstacles, such as limited quality data acquisition for quantifying tacit knowledge and insufficient adoption of smart manufacturing technologies. As a potential remedy, this study demonstrates the application of smart manufacturing technologies for predicting the quality of reclaimed sand, specifically tailored for the sand casting industry. The developed strategy integrates: 1) detailed measurements of the environmental conditions in the sand reclamation process, and 2) a deep-learning-based model for predicting the loss on ignition (LOI) of reclaimed sand as a quality measure. The model is constructed using feature extraction from time-series data and Bayesian neural networks to predict LOI with quantified uncertainty. We propose a normality score-based reclaimed sand management strategy, which was evaluated over one and a half years of production conditions and reclaimed sand quality monitoring experiments. The demonstration case exhibits an average accuracy of 96.83 % in detecting problematic sand quality. Notably, the method significantly improved failure detection accuracy, increasing test data results from 38.34 % without uncertainty consideration to 72.5 % when uncertainty was incorporated. The proposed approach has the potential to advance the casting industry by enabling quality-data-driven management of the sand reclamation process, ultimately reducing defect rates and optimizing production costs.