Verification of Response Evaluation Using Prior-Knowledge-Infused Neural Network in Structural Health Monitoring With Limited Observation Data

This study proposes a Seismic Response Estimation method using a Prior-Knowledge-infused Neural Network (PKNN) to overcome data scarcity in Structural Health Monitoring. Unlike Physics-informed Neural Networks (PINNs) that require strict differential equations, PKNN integrates structural dynamics as "soft constraints," enabling flexible learning for nonlinear structures. The framework employs two-stage learning: pre-training with low-fidelity physics models and incremental learning with limited high-fidelity FEM data. Results demonstrate that expanding the training data's response range effectively reduces estimation bias. Furthermore, an optimal balance exists between response diversity and sample size, which is crucial for maintaining generalization in few-shot learning.