Bridge health monitoring, structural assessment and dynamic response: overview and perspectives

2.1. Dynamic-static stiffness conversion and capacity evaluation

Tan et al. [7] propose a quick and easy evaluation method for the bearing capacity of cracked reinforced concrete bridges based on a dynamic-static stiffness conversion model. A virtual crack-based FEM links crack parameters to both dynamic and static stiffness; a surrogate model then converts measured dynamic stiffness (from vibration tests) into static stiffness and a verification factor equivalent to a static load test. Field application to a hollow slab bridge yields a bearing-capacity factor within about 3 % of a full static load test, at significantly reduced time and cost.

This work is representative of a broader trend in which measured dynamics are used as proxies for static capacity, especially where static load tests are expensive or disruptive. Similar correlations between dynamic and static stiffness have been observed in experimental and numerical studies on cracked or deteriorated members [3], [4], [8].

2.2. Data-driven capacity indicators and prestress assessment

Calò et al. [9] develop a machine-learning-based surrogate model to predict prestressing force reduction in PSC box-girder bridges with unbonded tendons, using a large FE-generated dataset calibrated against experimental tests. The framework predicts prestress loss as a function of initial prestress, concrete modulus, and strain variations at selected sections, achieving low prediction errors and interpretable patterns via SHAP analysis.

This complements dynamic–static stiffness conversion: instead of only using global dynamic stiffness, it demonstrates how monitoring data, FE modelling, and ML can infer internal, otherwise inaccessible parameters (e.g., effective prestress) at the portfolio level.

2.3. Correlation-based damage indicators

Deng et al. [10] propose a damage identification method for long-span bridges based on the correlation of monitored global dynamic responses in high-dimensional space. The method uses the ratio of mean squares between different response channels as a robust indicator that is relatively insensitive to load variability but sensitive to structural changes.

In parallel, reviews of vibration-based damage detection show a clear trend towards robust, multi-sensor features and ML-enabled indicators, including deep learning, autoencoders, and hybrid physics-ML workflows [6].

3.1. Operational traffic and surface roughness

Ho and Nishio [11] model dynamic responses of bridges under random traffic flow and surface roughness, validating the model against monitoring data from a steel box-girder bridge. They show that RMS accelerations depend strongly on truck ratios and speeds, while peak responses are sensitive to axle sequences and roughness patterns. Two wireless sensor nodes, namely sensors #1 and #2, were attached to the lower flanges at the mid-span of the two box-girders of the traffic lanes, as illustrated in Fig. 1(a) and (b). Synchronized three-axis accelerations were continuously acquired throughout the data acquisition period with a sampling frequency of 100 Hz. Thereafter, the vertical acceleration ($z$-direction) was used for validation. Developments in bridge weigh-in-motion (B-WIM) that explicitly include bridge dynamics [12] improve axle-load estimation and continuous influence-line identification, especially for longer spans. Together, these works highlight that vehicle-bridge interaction and road profile must be represented, at least approximately, when interpreting vibration data under real traffic.

Fig. 1Example for target bridge span and configuration of wireless sensor nodes

a) Target bridge spam

b) Cross-section (unit: mm)

3.2. Heavy vehicle impact and local failures

Heng et al. [8] analyze dynamic responses of highway bridges subjected to heavy truck impacts using detailed FE modelling and advanced concrete damage material laws, validated against crash tests. They outline typical failure modes of piers, caps, and girders and demonstrate how impact speed, mass, and angles govern the extent of local and global damage. The various failure modes of the impacted RC piers can be generated in actual vehicular collisions. These studies provide realistic envelopes for impact-induced demands and support robustness assessment, design of protective systems and calibration of intensity dynamic tests.

3.3. Transition zones and high-speed rail

Hu et al. [12] conduct extensive field tests on bridge-embankment transitions in high-speed railway lines, for train speeds from 5 to 360 km/h. Using wavelet-based and smoothing techniques, they characterize critical sections and dominant frequencies (mostly below 50 Hz), identify critical distances from abutments, and observe a critical speed at which dynamic amplification is maximized. In this case dynamic stress changes only slightly as train speed increases: the curve follows a gentle S-shape, indicating speed has little effect except at 12.27 m from the abutment, where stress is highest. The inflection occurs at roughly 275 km/h. Fig. 2(b) and 2(c) reveal that dynamic acceleration and displacement at the bed surface rise with train speed, with the greatest growth at 25.27 m from the WAT and increasing further as the measurement point moves away from the WAT. These relationships form steeper S-shaped curves, with inflection points again around 275 km/h. Thus, 275 km/h is regarded as a critical train speed. This illustrates how local stiffness transitions can be detected and studied via dynamic responses, which is relevant for monitoring strategies that use moving loads as “probes” of the structure.

Fig. 2Flow Chart of ML and damage detection strategies

3.4. Multi-hazard and reliability-based frameworks

Chen et al. [13] investigate deep-water bridge piers under combined seismic and wave loading using a 1:30 shaking-table model with wavemaker. They report substantial shifts in natural frequencies, strong added-mass effects, and nonlinear interactions such that the combined response is not simply the sum of individual actions. Yang et al. [14] propose a unified stochastic dynamic and system reliability framework for long-span cable-stayed bridges under near-fault ground motions, using direct probability integral methods and pulse-type ground-motion models to compute time-varying response statistics and failure probabilities. More broadly, integrated BHM reviews emphasize that multi-hazard, reliability-based approaches are becoming central, moving beyond deterministic, single-hazard analyses [1].

4.1. Dynamic modelling in BHM workflows

Several research works integrate refined numerical modelling with monitoring: dynamic B-WIM algorithms that explicitly account for bridge vibrations to estimate influence lines and axle loads [12]; shaking-table and fluid-structure interaction models for combined seismic-wave loading on deep-water piers [13] and nonlinear analysis of rocking bridge systems under 3D ground motions [15]. At a methodological level, integrated SHM frameworks increasingly adopt the “digital twin” concept: finite-element models are continuously updated using BHM data and then used for prediction, scenario analysis, and decision support [1]. Mazzeo et al. [16] propose an inerter-based dissipation system for a cable-stayed bridge, achieving improved energy dissipation and reduced seismic responses under transverse excitation compared with conventional viscous dampers. Such structural control systems are tightly linked with BHM: monitoring data can be used to tune or assess controllers, while the presence of control devices modifies the dynamic signatures that BHM algorithms must interpret.

4.2. Machine learning and intelligent damage detection

Beyond the ML-based prestressing framework of Calò et al. [9], it assisted on a rapid growth in ML and deep-learning methods for vibration-based damage detection, including convolutional neural networks, autoencoders, and hybrid physics-informed networks [6]. These methods are promising when combined with rich sensing (CV, GNSS, DFOS, IoT), but they raise challenges around generalization across structures, interpretability of learned features, and the need for large, high-quality labelled datasets.