희소 계측 데이터 기반 고주파 구조 진동장 예측을 위한 딥러닝 대리모델

키워드

확장요약문

Digital twins and structural health monitoring require full-field vibration reconstruction and load identification from sparse sensor measurements. Physics-informed machine learning offers a promising approach, but physics-informed neural networks (PINNs) and other learning-based solvers often struggle to converge at high excitation frequencies. Common remedies, including Fourier feature embeddings, causal weighting, and domain decomposition, still rely on learning the solution and may be insufficient for resonance-dominated forced responses. This study introduces a framework that embeds the solution structure directly into the representation. A parameter-adaptive harmonic basis remains numerically stable arbitrarily close to resonance and makes the PDE residual linear in the expansion coefficients, enabling a training-free direct solve and lightweight neural network interpolation. Using the same basis, an inverse pipeline identifies system parameters and loads from sparse measurements, and a residual-based consistency check quantifies the reliability of each estimate relative to the measurement noise. Under identical benchmark conditions, learning-based solvers equipped with frequency embeddings and causal training fail to resolve the response accurately, whereas the proposed framework delivers highly accurate solutions in real time. The inverse pipeline exhibits graceful degradation even under substantial measurement noise, with each estimate accompanied by its own reliability assessment. These results suggest that the high-frequency barrier arises from a mismatch between solution structure and model representation rather than insufficient network capacity. By addressing this mismatch, the proposed framework provides a common basis for efficient forward prediction, inverse estimation, and residual-based reliability assessment.