Dynamic Monitoring Method of Polymer Injection Molding Product Quality Based on Operating Condition Drift Detection and Incremental Learning

Prediction models for polymer injection molding quality often degrade due to shifts in operating conditions caused by variations in melting temperature, cooling efficiency, or machine conditions. To address this challenge, this study proposes a drift-aware dynamic quality-monitoring framework that integrates hybrid-feature autoencoder (HFAE) drift detection, sliding-window reconstruction error analysis, and a mixed-feature artificial neural network (ANN) for online quality prediction. First, shifts in processing parameters are rigorously quantified to uncover continuous drifts in both input and conditional output distributions. A HFAE monitors reconstruction errors within a sliding window to promptly detect anomalous deviations. Once the drift index exceeds a predefined threshold, the system automatically triggers a drift-event response, including the collection and labeling of a small batch of new samples. In benchmark tests, this adaptive scheme outperforms static models, achieving a 35.4% increase in overall accuracy. After two incremental updates, the root-mean-squared error decreases by 42.3% across different production intervals. The anomaly detection rate falls from 0.86 to 0.09, effectively narrowing the distribution gap between training and testing sets. By tightly coupling drift detection with online model adaptation, the proposed method not only maintains high-fidelity quality predictions under dynamically evolving injection molding conditions but also demonstrates practical relevance for large-scale industrial production, enabling reduced rework, improved process stability, and lower sampling frequency.