A multi-head neural network architecture for EIT inverse problem resolution in wall moisture imaging
Artykuł w czasopiśmie
MNiSW
0
brak dyscyplin
| Status: | |
| Autorzy: | Kłosowski Grzegorz, Hoła Anna, Kulisz Monika, Rymarczyk Tomasz, Niderla Konrad, Sikora Jan |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 283 |
| Numer artykułu: | 122107 |
| Strony: | 1 - 16 |
| Impact Factor: | 5,6 |
| Web of Science® Times Cited: | 0 |
| Bazy: | Web of Science | Ei Compendex |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | TAK |
| Licencja: | |
| Sposób udostępnienia: | Witryna wydawcy |
| Wersja tekstu: | Ostateczna wersja opublikowana |
| Czas opublikowania: | W momencie opublikowania |
| Data opublikowania w OA: | 2 czerwca 2026 |
| Abstrakty: | angielski |
| Moisture accumulation in historic masonry accelerates material degradation and requires non-destructive techniques capable of mapping its spatial variability. Electrical impedance tomography (EIT) can provide con-ductivity images from boundary voltage measurements, but the inverse problem is ill-posed and conventional dense regressors often yield over-smoothed reconstructions and artifacts. This study introduces a data-driven EIT reconstruction pipeline based on a multi-head fully connected neural network that partitions the 3050-element flattened FEM image vector into 122 consecutive 25-element segments predicted in parallel from a shared latent representation of the 448-voltage measurement sequence. Synthetic training and validation data are generated with a finite-element forward model for randomly sampled moisture-dependent conductivity fields (49,000 training and 1,000 validation cases). To ensure robustness against measurement uncertainty, a parallel model was trained on data augmented with 5% Gaussian noise. Compared with a single-head baseline of comparable capacity, the multi-head model improves reconstruction fidelity over 1,000 validation examples, reducing mean squared error from 0.205 to 0.0927 (≈55%) and increasing PSNR from 7.34 to 10.74 dB, SSIM from 0.769 to 0.850, and Pearson correlation from 0.9545 to 0.9807. The approach is further validated on in situ measurements acquired with a 32-electrode linear array on a historic church, where reconstructed moisture gradients agree with independent dielectric point measurements and infrared thermography. The proposed output decomposi- tion acts as an architectural regularizer that stabilizes low-to-high dimensional regression, suppresses artifacts, and enhances the interpretability of EIT-based moisture imaging for heritage diagnostics. |
