Hyperparameter Optimization-Based Structural Field Reconstruction for Structural Health Monitoring Under Limited Data Regime


Özkan R. B., Altaş B., Uyar M., Rutçi A.

International Kayseri Scientific Research and Innovation Congress, Kayseri, Türkiye, 30 - 31 Mayıs 2026, ss.253-263, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Kayseri
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.253-263
  • Bursa Uludağ Üniversitesi Adresli: Evet

Özet

In structural health monitoring (SHM), reconstructing the full-field strain distribution across a surface from a sparse set of physical sensors is a fundamental challenge for aerospace applications, inherently

constrained by limited sensor deployment and high computational costs. This study proposes a data￾driven framework for reconstructing the full-field strain distribution of a delta-wing aircraft using only 

five optimal sensors. The underlying dataset comprises a finite element analysis (FEA) model with approximately 107,000 nodes, generated under 13 distinct load scenarios using ANSYS Mechanical.

Proper orthogonal decomposition and the discrete empirical interpolation method were utilized to determine the optimal sensor locations. Sensor measurements were enriched through spatial coordinates,

polynomial interaction terms, and physics-informed features. These descriptors were subsequently used to train a weighted ensemble combining Elastic-Net and Gaussian process (GP) regressors. Under a 

leave-one-out cross-validation protocol, the proposed ensemble achieved a weighted R² of 0.891, outperforming seven alternative architectures. Furthermore, five hyperparameter optimization methods 

(TPE, CMA-ES, NSGA-II, PSO, and simulated annealing) were systematically benchmarked across multiple performance metrics. While PSO yielded the highest wR² (0.913), TPE provided the most 

balanced node-based median R² (0.862). The observed discrepancies among evaluation metrics highlight the importance of multi-metric assessment when optimizing structural reconstruction models 

under limited-data conditions.