Digital Twin-Driven Machine Learning and Physics-Informed Data Augmentation for Clutter-Resilient Chipless RFID Decoding in Smart Logistics
IEEE ACCESS, cilt.14, ss.116597-116612, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3717950
- Dergi Adı: IEEE ACCESS
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.116597-116612
- Ondokuz Mayıs Üniversitesi Adresli: Evet
Özet
The deployment of ultra-low-cost, FR-4-based passive chipless RFID tags in smart logistics is strongly affected by multipath fading, non-line-of-sight (NLoS) shadowing, and location-dependent warehouse clutter. Conventional machine-learning decoders trained only with idealized Gaussian noise may therefore overestimate performance in cluttered industrial environments. To address this limitation, this study proposes a digital twin-driven machine-learning framework supported by physics-informed data augmentation for clutter-resilient chipless RFID decoding. A 40 & times; 24 & times; 8 m(3) logistics facility was modeled in CST Studio Suite using an Asymptotic Ray-Tracing/Shooting and Bouncing Rays (SBR) solver. At z=0.9 m on the warehouse evaluation plane, 1,188 observation-grid coordinates were exported, and 994 valid nonzero spatial clutter samples were extracted, with clutter levels ranging from -36.0540 dB to -2.1606 dB. These digital twin-derived clutter samples were superimposed with 100 pristine FR-4 V-slot tag responses to construct a simulation-supported augmented dataset for Random Forest regression. To prevent data leakage, all augmented samples derived from the same base tag geometry were kept within the same group. Accordingly, 80 tag groups were used for training/validation, while 20 completely held-out tag groups were reserved for blind testing. The proposed model achieved a Group 5-Fold Cross-Validation MAE of 0.2097 +/- 0.0062 mm and, on the group-separated blind-test set, an MAE of 0.2188 mm, RMSE of 0.4269 mm, and R-2 of 0.9777. In contrast, a conventional Gaussian-noise-only baseline evaluated on the same clutter test set produced an MAE of 1.9837 mm and R-2 of -0.0958, exceeding the 1.5 mm industrial tolerance threshold. The trained model also achieved an average inference time of 27.622 ms per tag. These results show that digital twin-derived clutter modeling and group-aware validation provide a reproducible route toward clutter-resilient chipless RFID decoding in smart logistics.