Physics-Informed Multimodal Sensing and Machine Learning Framework with Real-Time Digital Twin for Intelligent Structural Health Monitoring of Marine Wharves

Abstract

Marine wharves, consisting of concrete or steel decks supported by pile foundations, operate in harsh marine environments involving waves, vessel impacts, and corrosive seawater. Even minor cracks or early-stage corrosion in piles can lead to sudden structural failure. Early detection is therefore essential. Traditional vibration-based structural health monitoring (SHM) methods often struggle to deliver reliable results in real operational settings because environmental and operational variations can mask damage-induced frequency changes. This study presents a multimodal SHM framework that integrates accelerometers, strain gauges, high-resolution cameras, acoustic transducers, and corrosion sensors with edge computing, advanced machine learning, and a real-time physics-informed digital twin. Edge nodes preprocess sensor data to extract modal frequencies, crack indicators, and acoustic features, reducing communication loads by over 90% while preserving diagnostic fidelity. A hybrid pipeline combined attention-based multimodal fusion, XGBoost screening, and a feedforward deep neural network trained using a composite physics-informed objective. The physical loss numerically penalised positive damage-related modal-frequency shifts after environmental and operational condition compensation and stiffness ratios outside the admissible range of 0.70–1.00. The selected physics-loss coefficients were λphys = 0.10 and η = 0.50. Sensor features were ingested continuously, whereas finite-element state updates were event-triggered by a fused damage probability of at least 0.80, agreement between at least two sensing modalities, normalised predictive entropy not exceeding 0.35, and persistence over three consecutive 10-s analysis windows. Validation using 200 finite-element simulations and 40 scaled laboratory tests achieved 95.0% accuracy for single-damage detection, compared with 78.0% for the conventional Modal Strain Energy method. For multiple-damage cases, the proposed framework achieved 92.17% precision, 88.33% recall, and an F1-score of 90.21%. Damage-severity estimation yielded R² = 0.891 and an RMSE of 2.06 percentage points, while the mean full-scale-equivalent localisation error was 4.8 cm. Edge processing reduced communication volume by 94.73% without a statistically significant reduction in diagnostic accuracy. While evaluated under controlled conditions, the framework shows strong potential for proactive, accurate, and interpretable monitoring of marine wharves in operational environments.