Predicting ceramic tile quality using LSTM and CatBoost: a case study of industrial roller kiln operations

Abstract

This paper investigates an application-oriented artificial intelligence framework for predicting the first-grade production percentage of ceramic tiles in an industrial roller kiln. The objective is to relate kiln operating conditions and biscuit physicochemical properties, including kiln-zone temperatures, firing cycle, biscuit flexural strength, biscuit moisture content, and final-product water absorption, to the quality outcome observed in production. Because data acquisition in industrial kiln operation is costly and limited, the study uses an expert-validated, uncertainty-aware training strategy rather than treating synthetic records as new independent plant measurements. After outlier removal, an independent real test set was separated and kept unchanged. Controlled perturbations within the normal industrial measurement uncertainty were applied only to the training subset: non-constant kiln-temperature variables were locally varied within the process tolerance confirmed by plant experts, while the reported uncertainty of the first-grade production percentage was considered within +/−0.7 percentage points for training robustness. Several baseline models, including FNN, CNN, LSTM, CatBoost, and the proposed LSTM-CatBoost framework, were evaluated using multiple regression metrics. The LSTM-CatBoost model achieved the strongest overall performance, with an R ² of 0.946 on the test evaluation. The main contribution of this work is not the proposal of a new general-purpose machine-learning architecture, but the adaptation and evaluation of a practical data-driven quality-prediction framework for real roller-kiln operation under industrial data constraints.