Abstract

We propose a physics-constrained wavelet sparsification framework to enhance the fidelity and interpretability of digital twins for heritage preservation, addressing the challenge of processing high-frequency IoT sensor data without losing critical structural features. The method integrates domain-specific material properties and structural dynamics into a wavelet-based signal processing pipeline, enabling adaptive sparsification tailored to heritage sites. A multi-modal sensor array captures real-time data, which is then decomposed using a Morlet wavelet transform with dynamically adjusted thresholds derived from physics-informed constraints. These constraints incorporate stiffness and damping matrices to balance sparsity against structural fidelity. Furthermore, a material-aware convolutional neural network with attention mechanisms identifies and preserves low-energy features, such as micro-cracks, by conditioning attention weights on material parameters. The sparsified features are fused with a finite element model through an optimization process that minimizes residuals while accounting for sensor noise and prior knowledge of material degradation. The proposed system forms a closed-loop architecture, where updated finite element parameters feedback into the wavelet thresholding module for adaptive refinement. Unlike generic denoising techniques, our approach retains heritage-specific features by explicitly modeling material brittleness and non-stationary loads. The framework bridges the gap between high-frequency sensor data and computationally tractable digital twins, offering a scalable solution for real-time monitoring and preservation of cultural heritage sites. Experimental validation demonstrates its superiority in maintaining feature integrity while reducing data dimensionality, making it suitable for integration with conventional 3D modeling and simulation tools.

Keywords