Deep learning prediction of cold-side quality characteristics from hot-side infrared sensing in glass bottle manufacturing

Quality inspection in glass bottle manufacturing traditionally occurs on the cold side of the production line, after the containers have cooled sufficiently for contact-based or optical measurements. While this ensures reliable defect identification, it introduces a substantial delay between forming and feedback, limiting opportunities for real-time process control and waste reduction. In this study, we present an industrial Artificial Intelligence (AI) framework that predicts cold-side quality characteristics directly from hot-side infrared measurements acquired immediately after mold release. Our methodology relies on correlation-based modeling rather than one-to-one traceability. We develop a supervised deep-learning model that maps hot-side infrared emission patterns to aggregated cold-side quality metrics. Experiments performed on data collected from a production-scale glass line show that the model reliably predicts thickness-related cold-side characteristics within operating regimes sufficiently represented in the training data, providing early-warning information roughly one hour before conventional cold-end inspection becomes available. The analysis also delimits the scope of the approach. Under temporal distribution shift, neck-related variables retain relatively low prediction errors but exhibit reduced coefficients of determination because of their narrow dynamic range, whereas ovalization and pressure-related variables show poor generalization. These limitations are consistent with the physical constraints of the single-view infrared sensing setup. The proposed framework should therefore be understood as a variable-selective early-warning tool rather than as a general predictor of all cold-side quality characteristics. Its predictions can support operators in detecting emerging deviations, adjusting forming parameters proactively, and reducing quality drift, thereby laying the groundwork for future integration into closed-loop process control.