Vol 5 No 2 (2026)

Reliable inverse identification of mechanical system parameters from sparse and noisy measurements remains a major challenge for digital twin applications because conventional identification methods are highly sensitive to measurement uncertainty and limited sensing data. This study presents a physics-informed digital twin framework that integrates the governing equation of motion with a neural network to simultaneously reconstruct system dynamics and estimate unknown mechanical parameters while maintaining physical consistency. The framework is developed for a single-degree-of-freedom mass–spring–damper system and jointly optimizes the displacement response together with the mass, damping, and stiffness using a two-stage Adam–L-BFGS optimization strategy. Its performance is evaluated through convergence analysis, basin stability, structural identifiability, ensemble-based uncertainty quantification, comparisons with least-squares, recursive least-squares, and data-driven neural networks, validation using independent literature-based parameter sets, and sensitivity analysis under signal-to-noise ratios ranging from 5 to 40 dB. The proposed framework reconstructs the system response with RMSE values of 0.013–0.021 while identifying mass, damping, and stiffness with relative errors of approximately 1.1%, 1.1%, and 1.3%, respectively. Compared with conventional and purely data-driven approaches, it achieves more accurate parameter estimation, lower reconstruction error (RMSE = 0.016), consistent convergence across multiple random initializations, and reliable 95% confidence intervals for prediction uncertainty. Unlike previous studies that primarily demonstrate inverse parameter recovery using physics-informed neural networks, this work provides a comprehensive reliability-oriented evaluation by integrating optimization robustness, structural identifiability, uncertainty quantification, comparative benchmarking, and multi-noise validation within a single physics-informed digital twin framework, improving the reliability and interpretability of inverse parameter identification under limited sensing conditions.

This paper investigates the development and deployment of spectral-semantic analytical frameworks for local space analysis using neural networks. The proposed paradigm unifies spectral data processing—with measurements collected from spectrometers and thermal imagers—and semantic modeling, which frames spectral features as bearers of structured semantic information. This work elaborates on techniques for translating spectral modality into linguistic-semantic representations; this transformation enables dual characterization of local physical properties: quantitative description via spectral parameters, and qualitative description via semantic profiles and patterns. Special focus is given to building spectral-semantic dictionaries and matching neural architectures that extract and formalize contextual correlations between object spectral signatures and their semantic implications. Diverse spectrogram formats and spectral representations (multiband and hyperspectral data included) are adopted for local space characterization, alongside a neural network metamodel capable of generating domain-specific spectral-semantic models. Attention modules are embedded into these architectures to automatically screen high-value spectral bands and spatial regions pivotal to semantic reasoning.