STID-Mixer Model for Polymer Melt Index Prediction from Extrusion Process Sensor Streams

Source abstract

Online estimation of polymer melt index is difficult because laboratory assays become available only after the corresponding material has progressed beyond the critical extrusion stages, whereas plant historians continuously record thermal, pressure, torque, speed, and feeding behavior. This study developed STID-Mixer, a compact soft sensor that converts multirate extrusion streams into a residence-aware representation for continuous melt-index prediction. Historian signals were synchronized with laboratory samples, screened by production state, normalized within product grade, and transformed into sensor-time tokens containing dynamic descriptors and material-residence weights. The network then separated cross-sensor interaction learning from temporal-delay mixing and used a robust regression objective. Evaluation used 18,420 one-minute windows derived from 17 raw sensor variables, 11 engineered descriptors, and six polymer grades under a chronological campaign-based train-validation-test split. STID-Mixer achieved a test RMSE of 0.118 g/10 min, an MAE of 0.083 g/10 min, an R² of 0.936, and 91.7% coverage within ±0.20 g/10 min. Its RMSE was 21.9% lower than that of CatBoost, while single-window inference required 3.4 ms. Ablation, residual, drift, and replay analyses indicated that grade normalization, residence weighting, and separated sensor-time mixing contributed complementary benefits. Errors increased during early grade transitions, upper-range operation, and progressive die-pressure drift. The results support STID-Mixer as a computationally light advisory estimator for extrusion-quality surveillance, while prospective cross-line validation and uncertainty-aware alarms remain necessary before closed-loop use.