A SWIR spectral measurement system for robust plastic sorting using AutoEncoder-gated XGBoost with out-of-distribution detection

Accurate material-specific sorting is essential for high-throughput plastic recycling, yet conventional visible-band vision methods are brittle to transparency, surface contamination, and oxidation. This work presents a mathematically grounded two-stage framework for classifying plastics from short-wave infrared (SWIR) 1-D spectra while reliably rejecting out-of-distribution (OOD) inputs. Stage 1 employs an AutoEncoder trained on in-distribution (ID) spectra to compute sample-wise reconstruction errors; an OOD gate is defined using both ID-only thresholding and validation-calibrated thresholding strategies, allowing comparison between unsupervised and semi-supervised operating modes. This distinction provides a more realistic assessment of the proposed system under both fully unsupervised deployment and limited calibration conditions. Stage 2 classifies the remaining ID-gated samples with XGBoost, and a softmax confidence threshold further suppresses overconfidence. The proposed framework was evaluated using 3,640 spectra acquired in the 900–1700 nm range, comprising six ID polymer classes (LDPE, HDPE, PETa, PETc, PET and PP) and OOD samples consisting of chemically distinct polymers (PS, PVC). Under the ID-only unsupervised threshold (Tp99), the proposed method achieves an OOD F1-score of 0.81. Under the validation-calibrated semi-supervised threshold (TF1), the OOD F1-score improves to 0.94 (precision = 0.92, recall = 0.95). The average inference time of the combined AE + XGBoost pipeline was approximately 2.1 ms per spectrum, corresponding to about 476 spectra/s. Under a 2 mm line-spacing condition, this corresponds to an estimated conveyor speed of approximately 57 m/min, supporting the practical feasibility of the proposed system for high-throughput recycling environments. Overall, the integrated design combining distribution-aware gating with gradient-boosted classification improves robustness to unseen materials and supports reliable, automation-oriented sorting in industrial recycling lines.