Polymer Melt Stability Monitoring in Injection Moulding Using LSTM-Based Time-Series Models

This work presents a data-driven framework for early detection of polymer melt instability in industrial injection moulding using Long Short-Term Memory (LSTM) time-series models. The study uses six months of continuous production data comprising approximately 280,000 injection cycles collected from a fully operational thermoplastic injection line. Because melt behaviour evolves gradually and conventional threshold-based monitoring often fails to capture these transitions, the proposed approach models temporal patterns in torque, pressure, temperature, and rheology to identify drift conditions that precede quality degradation. A physically informed labelling strategy enables supervised learning even with sparse defect annotations by defining volatile zones as short time windows preceding operator-identified non-conforming parts, allowing the model to recognise instability windows minutes before defects emerge. The framework is designed for deployment on standard machine signals without requiring additional sensors, supporting proactive process adjustments, improved stability, and reduced scrap in injection moulding environments. These findings demonstrate the potential of temporal deep-learning models to enhance real-time monitoring and contribute to more robust and adaptive manufacturing operations.