STAGE-EN: An ElasticNet-Gaussian Process Stacking Ensemble Approach for Delta Wing Strain Reconstruction Using Physics-Informed Feature Engineering and DEIM
11th International Congress on Engineering Sciences and Multidisciplinary Approaches, İstanbul, Türkiye, 23 - 24 Mayıs 2026, ss.46-53, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.46-53
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
This study proposes the STAGE-EN (Stacking GP Ensemble with ElasticNet) model for reconstructing the full-field strain distribution of a delta-wing structural model (~107,000 nodes under 13 loading scenarios) using only five strain gauges. Sensor locations are determined under a minimum-distance constraint using a discrete empirical interpolation method (DEIM) based on singular value decomposition (SVD), which extracts dominant deformation modes, and DEIM selects the most informative nodes. During feature engineering, the input set is enriched with spatial coordinates, random forest-filtered polynomial features, and physics-informed structural descriptors, including strain gradients, distance-weighted averages, and axial asymmetry terms. The proposed architecture models global linear behavior using ElasticNet regression and nonlinear residuals using Gaussian process regression. Final predictions are obtained through an MSE-weighted ensemble strategy. Evaluated using scenario-based leave-one-out cross-validation, the framework achieves an overall R² of 0.873 and a median node-wise R² of 0.825. The inclusion of physics-informed features improves performance by +0.0293 in R², while sensor quantity analysis indicates that five sensors provide the best trade-off between reconstruction accuracy and sensing cost. The proposed pipeline demonstrates strong potential for digital twin-based structural health monitoring applications.