Transfer learning from simulation to real-world data enables accurate structural health monitoring with minimal labeled experimental data—a practical pattern for deploying deep learning in engineering applications with expensive real-world measurements.
This paper presents a transfer learning framework combining physics-based simulations and deep learning to diagnose structural damage using guided waves. By pretraining on cheap simulated data and fine-tuning on limited real experimental data, the approach achieves high accuracy for damage detection in structures with piezoelectric sensors while reducing computational costs.