This paper presents a novel approach to the drive-by Structural Health Monitoring (SHM) of bridges based on crowdsourced data collected from multiple vehicle passages. The methodology is experimentally validated on a laboratory-scale bridge model designed to replicate the dynamic behaviour of real bridge structures, exhibiting representative natural frequencies and mode shapes. Acceleration data were acquired using smartphones mounted on a specially adapted vehicle traversing the bridge under varying conditions, including different speeds, directions, and lateral positions. Multiple crossings were conducted to simulate a crowdsourcing scenario, where independent vehicle measurements contribute to a shared monitoring database. This experimental setup enabled the evaluation of how data aggregation from numerous runs can enhance the robustness and accuracy of the drive-by monitoring technique. The collected acceleration signals were processed in MATLAB to extract the bridge’s key dynamic parameters. The analysis focused on estimating natural frequencies and identifying mode shapes based on the aggregated datasets. The results were compared with reference measurements obtained from high-precision accelerometers permanently installed on the bridge model. The identified parameters showed only minor deviations—typically within a few percent—from the reference values, confirming the reliability of the proposed approach. Furthermore, combining data from multiple vehicle passages reduced measurement variability and improved the stability of the identified dynamic characteristics. The findings demonstrate the feasibility and potential of using crowdsourced data for drive-by bridge monitoring with widely available mobile sensors. By leveraging repeated measurements from multiple independent crossings, the study highlights a scalable, low-cost, and non-invasive solution for continuous bridge health assessment. This approach paves the way toward practical implementation of large-scale SHM systems capable of exploiting existing traffic as a dynamic sensing network.

Drive-by bridge health monitoring provides a low-cost means of assessing bridge integrity through vehicle-mounted sensors. Yet, its effectiveness is limited by data scarcity and the high noise of field measurements. This study investigates how different data augmentation strategies—including physics-guided synthesis, time-series generative adversarial networks (TimeGAN), and stochastic noise—can enhance the accuracy and robustness of drive-by damage detection. These augmented datasets were validated using vibration data from the MITICA (Monitoring Transport Infrastructures with Connected and Automated Vehicles) testbed bridge in Ispra, Italy. Comparative experiments show that appropriate augmentation can improve damage classification accuracy by more than 10%, under low-sample and noisy conditions. The findings demonstrate that carefully designed augmentation can mitigate data scarcity and substantially boost the reliability of AI-driven drive-by monitoring systems.

Structural health monitoring (SHM) data analysis demands domain expertise and multi-stage specialized engineering computations. Despite their general-purpose reasoning abilities, large language models (LLMs) lack deep SHM knowledge and exhibit limited numerical accuracy, which constrains their effectiveness as stand-alone engineering solutions. We present SHMClaw, an LLM-powered multi-agent framework that enables autonomous SHM data interpretation and report generation. Through multi-agent collaboration, the framework unifies user requests with LLM reasoning, domain knowledge and specialized computational tools into a fully artificial intelligence (AI)-guided, end-to-end SHM workflow. We demonstrate its effectiveness through analysis of monitoring data from a footbridge at The Hong Kong Polytechnic University, generating a professional monitoring report. Our results validate SHMClaw as a practical AI co-engineer and establish a new, extensible computational paradigm for engineering intelligence, with significant implications for the autonomous operation and maintenance of city-wide infrastructure networks.

Existing SHM methods normally consist of direct instrumentation of infrastructures (e.g., bridges) with sensors for continuous or periodic health assessment. However, traditional SHM techniques can be time-consuming and expensive due to the requirements for on-site installations (especially cable deployment). Recently, indirect vibration-based SHM strategies have been proposed in the literature, which make use of the dynamic response of instrumented vehicles to carry out “drive-by” monitoring of road pavement and bridges. Such vehicles are generally cars or trucks on which sensors are installed. The use of bicycles for indirect SHM purposes has been much less explored. In this paper, the potentialities for innovative indirect vibration-based SHM using only commercial bicycles and smartphones are proposed. Smartphones are widespread, contain several types of sensors, and have on-board computing capabilities, embedded batteries, and internet access. Hence, in this study, a smartphone, fixed on bicycles, is used to collect different types of data (acceleration, angular velocity, position, orientation, magnetic field), which are fused using a Kalman filter algorithm, thus enabling the extraction of the main dynamic parameters of footbridges. The methodology is applied to two real footbridges in Northern Italy.

This contribution presents an experimental study investigating the potential of mobile-sensing strategies for vibration-based identification (VBI) of infrastructures. The work focuses on the use of a small-scale vehicle as a mobile sensing unit, instrumented with both conventional accelerometers and a commercial smartphone, with the aim of assessing the feasibility of low-cost and easily deployable monitoring solutions versus professional sensing equipment. The proposed approach relies on vehicle–structure interaction: the instrumented vehicle is driven at controlled speeds over selected infrastructures in the urban area of Palermo (Italy), enabling the acquisition of dynamic response data associated with both the vehicle and the underlying structure. The accelerations recorded by piezoelectric sensors, considered as reference measurements, are complemented by those collected through the embedded sensors of the smartphone. This dual-sensing configuration allows for a systematic evaluation of sensitivity, noise performance, and the capability of extracting key vibration-based features. While the accelerometer-equipped setup demonstrates clear potential for identifying structural dynamic characteristics, the performance of the smartphone presents a more nuanced outcome. Issues related to noise, sampling stability, and environmental interference are considered, with the aim of understanding the limitations and possible preprocessing strategies required for reliable modal identification. The analysis highlights the conditions under which smartphone measurements may approach acceptable accuracy, as well as scenarios where their performance remains insufficient without additional filtering or calibration. While demonstrated on a scaled vehicle platform, the methodology is applicable to other structures and mobile sensing scenarios. Despite these open challenges, the study reinforces the promise of mobile-sensing techniques as scalable and cost-effective tools for structural diagnostics. The results contribute to the ongoing development of practical methodologies for vibration-based assessment that could, in the future, support widespread, rapid, and user-friendly structural monitoring campaigns in transportation networks.

To address issues concerning road pavement health, which can ultimately result in economic losses and traffic accidents, this study introduces a roughness identification method based on in-vehicle smartphone sensors, referred to as the Correlated-Wheel Augmented Kalman Filter (CW-AKF). The principal innovations of the proposed approach are as follows:
(1) The front-wheel road roughness is incorporated as an augmented state variable to be estimated in real time by the filter, while the spatial correlation between front and rear wheel excitations is explicitly modeled, thereby reducing the number of unknown road excitations to be identified within the proposed framework.
(2) The inertial pitch effect induced by longitudinal acceleration is integrated into the filtering model, enabling the method to perform effectively under both constant-speed and variable-speed driving conditions.
(3) By combining accelerometer, gyroscope, and GPS data from the smartphone, the identified spatiotemporal roughness information is geographically localized, transforming the estimation results into a road information map with practical engineering relevance.
The effectiveness of the proposed algorithm is verified through numerical simulations, and its performance is further evaluated via real-world experiments comparing roughness identification results obtained using different vehicles.