The application of steel structures in infrastructure construction may exhibit early-stage damage characteristics caused by dynamic load fluctuations and connection imperfections, which, if undetected, could pose significant safety risks. The paper aims to discuss this issue.
This study presents a dynamic health monitoring framework for steel structures under construction, integrating multi-source sensing, wavelet–Kalman signal preprocessing, spatiotemporal feature extraction via a CNN–LSTM neural network and maintenance decision inference using a Bayesian probabilistic graphical model. Such components are embedded within a digital twin platform that supports end-to-end status assessment and strategy mapping. The system adopts a four-layer architecture, comprising sensing (STM32 MCU), edge computing (Raspberry Pi 4), cloud-based inference (GPU) and real-time visualization (Web Dashboard).
The framework was deployed on a 50-m cable-stayed bridge with 60 strain gauges and 30 accelerometers, continuously collecting 260 m data points over a 30-day period from the deployed sensor network, the results show that the proposed method achieves 93.6% classification accuracy and a 20% improvement in early warning sensitivity compared to conventional frequency-domain thresholds and offline probabilistic neural network models.
The system provides a robust solution for intelligent structural health monitoring and life-cycle infrastructure management.
