Top

Causally Constrained Temporal Fusion Framework for Recycled Plastic Quality Grading with Near-Infrared Spectra and Process Data

Authors

Aleksandar Ilić
Faculty of Process Engineering, International University of Belgrade, Belgrade, 11000, Serbia
Saša Milenković
Faculty of Process Engineering, International University of Belgrade, Belgrade, 11000, Serbia
Bogdan Kostić
Faculty of Process Engineering, International University of Belgrade, Belgrade, 11000, Serbia

Files

Abstract

Recycled plastic quality grading requires material-composition evidence and processing histories that differ in sampling frequency, transport delay, and measurement availability. This paper proposes a causally constrained temporal fusion framework that combines near-infrared spectra with extrusion process sequences while excluding information unavailable at the grading decision. The architecture comprises spectral compression, residence-time alignment, graph-constrained variable selection, historical attention, and an ordinal output head. A consistency objective compares material-matched observations from different presumed nuisance environments, while an order-preserving calibration procedure and confidence threshold provide a route to additional inspection. Evaluation is illustrated with 2,400 constructed batch records covering five ordered grades, including 600 test cases; no industrial measurements or completed model-training results are reported. The specified confusion matrix yields 92.0% accuracy, approximately 92.0% macro-averaged F1, and a mean absolute grade error of 0.083. Separate comparison scenarios specify 88.0% accuracy for conventional temporal fusion and 89.0% after removal of the admissibility mask. Prescribed missing-input and operating-shift scenarios assign accuracies of 88.0% at 20% missing inputs and 87.2% under sensor drift, respectively. These values demonstrate an auditable reporting structure rather than empirical superiority. Calibration curves and inspection costs are independently specified examples and cannot be reconstructed from the confusion counts alone. The framework makes explicit the measurement, graph, and label-policy assumptions needed for subsequent implementation. Its industrial utility and any intervention interpretation require validation using laboratory-qualified batches, synchronized acquisition records, and independently operated production lines.

Details