Akash, R., Haddad, Manu  ORCID: https://orcid.org/0000-0003-4153-6146 and Sarathi, R.
2026.
Physics-informed spectral fusion and attention-based deep learning for chalking severity classification in polymeric insulators.
IEEE Transactions on Dielectrics and Electrical Insulation
10.1109/tdei.2026.3724775
|
Preview |
PDF
- Accepted Post-Print Version
Available under License Creative Commons Attribution. Download (1MB) | Preview |
Abstract
Chalking is one of the most complex degradation phenomena that occurs in polymeric insulators. Conventional surface diagnostic methods, such as contact-angle and roughness measurements, are insufficient for reliable classification of chalking severity because the degraded layer is heterogeneous and powdery. This paper proposes an integrated spectroscopic approach combining ATR-FTIR and LIBS through a physics-derived Chalking Severity Index (CSI) and a Dual-Branch Attention Mechanism Convolutional Neural Network (DBAM-CNN). The CSI provides a continuous measure of degradation by combining FTIR chemical markers with LIBS elemental ratios, thereby providing consistent, physics-informed severity labels without manual visual scoring. The proposed DBAM-CNN achieved an overall accuracy of 97.4% and a fold-mean macro-F1 of 0.973 under batch-grouped five-fold cross-validation, exceeding conventional fusion baselines by approximately 6–16 percentage points in accuracy. An occlusion-sensitivity analysis further indicates that the model predictions are sensitive to degradation-relevant spectral regions rather than to chemically irrelevant features.
| Item Type: | Article |
|---|---|
| Date Type: | Published Online |
| Status: | In Press |
| Schools: | Schools > Engineering |
| Additional Information: | RRS applied |
| Publisher: | Institute of Electrical and Electronics Engineers |
| ISSN: | 1070-9878 |
| Date of First Compliant Deposit: | 10 September 2026 |
| Last Modified: | 10 Sep 2026 14:15 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/189151 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions