A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing

arXiv:2610.08822 · cs.LG · Submitted 2026-09-25 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing".

Jane: The gist A neural operator digital twin of a real truss bridge was trained on simulated vehicle crossings with member-level damage in the drive-by configuration,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're looking at this paper titled "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing". Jane, can you give us the quick rundown on what this is all about?

Jane: Okay, it sounds like they’re tackling how to monitor aging bridges without those expensive sensors or subjective visual inspections. The main idea here is using a vehicle driving by the bridge to gather structural information from how the vehicle interacts with it.

Tom: Right, so it's this drive-by inspection method that recovers structural data from vehicle-bridge interaction and vehicle-road interaction responses. The thesis seems to be that this approach offers a scalable alternative because you can use a sensorised vehicle instead of fixed sensors on every bridge.

Lu: What makes it different is that they’re building an AI digital twin of the bridge, which acts like a computational replica updated with every single crossing you make over it. This twin has to embed the physics of the truss into its structure so it can handle the structural state accurately.

Meng: So, this digital twin does two things: first, it predicts what response a certain structural condition would produce; and second, it infers what that structural condition actually is from the measurements you take. That’s a pretty ambitious setup for a single system.

Jane: Exactly. They use this digital twin framework to continuously monitor both the bridge and the road conditions using machine learning combined with physics-based modelling. It aims to unify those two approaches for continuous monitoring of bridge and road conditions through vehicle integration, which is what the paper calls "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing".

Tom: The paper claims they construct surrogate models for the vehicle-bridge interaction and vehicle-road interaction using a Fourier Neural Operator. This operator learns function-to-function mappings from operating conditions to vehicle responses, which means it can give millisecond scale inference instead of needing full order analyses which take too long.

Lu: That FNO is trained on simulated and field data, which is good because it’s being validated across different settings. They are using this operator to replace computationally intensive full-order analyses with something much faster for monitoring purposes.

Jane: And then they optimize the design of a custom electric inspection vehicle—its layout and sensor placement—using Bayesian optimisation to maximize the information yield while trying to suppress road and vehicle noise, which is a big challenge in real-world data collection.

Meng: I worry about that optimization part, though. If you’re tuning the sensor layout based on an electric vehicle's response, what happens when you move to a different type of bridge or a very rough road? It sounds like the system might be highly specific to that initial setup.

Paper summary: Tom: That’s a valid concern for deployment at scale. They address that by validating the complete workflow through coordinated multi-site field trials in Australia and Japan, covering different bridge types, traffic conditions, and environmental settings.

Jane: So they tested this whole digital twin framework on real bridges across different locations to see if it generalizes beyond just one specific structure or road type. That’s a crucial step for any practical application of this technology.

Lu: The paper specifically uses the Old Ada Bridge in Nara Prefecture, Japan, which is a simply supported steel truss with a main span of fifty-nine point two meters and a width of three point six meters as their testbed for this work on "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing".

Tom: The digital twin itself is built based on an FE model of the truss, where its twenty-nine members are grouped into four categories: bottom chord, top chord, verticals, and diagonals <ref:2610.08822#pg2,bottom chord, top chord, verticals, and diagonals>. Damage in this model is represented as a reduction in the Young’s modulus of one entire member.

Jane: So they have this finite vector of twenty-nine severities that represents the structural state of the bridge at any given time <ref:2610.08822#pg2>. The operator they use is a Fourier Neural Operator, which acts spectrally on the frequency content of those responses.

Lu: In the forward configuration, they input these twenty-nine severities as constant channels into the operator, and it returns the response measured at the vehicle's two axles over the entire crossing <ref:2610.08822#pg2,returns the response measured at the vehicle's two axles over the>. Because damage alters that response far less than what a normal crossing excites it with, they trained this operator on the residual—the difference between a damaged and healthy crossing.

Tom: That’s clever because you're training it on what *changes* due to the damage, not just what changes due to driving itself, which should make the learning process more focused.

Jane: Now for the inverse configuration, they take the axle responses as input and have the operator return both the damaged member and its severity through two separate output heads. They trained this on one thousand one hundred sixty simulated crossings, with forty crossings for each of those twenty-nine members at severities between zero point one zero and zero point six zero.

Meng: Ten hundred simulated crossings sounds like a decent training set size for an operator like this, Tom. But what is the computational cost to run this trained model in real-time on a mobile device or even a standard server? That’s where I see the first big hurdle for deployment at scale.

Paper summary: Tom: The paper says that in the forward direction, the predicted vehicle response closely tracks the FE solution, and they point out that this FNO requires only a small fraction of the computational cost of doing a full FE solution. That's a key advantage for real-time monitoring.

Jane: But then there’s this trade-off they mention in Figure four: prediction accuracy depends on how strongly the damage manifests in the measured response, and if you misidentify a member, it often selects one immediately adjacent to the true one within the same chord <ref:2610.08822#pg1>.

Lu: That suggests that distinguishability is harder than just observability; adjacent members can alter the vehicle response in nearly identical ways, and the estimated damage then splits between them. It’s a real physical limitation they are working around.

Tom: Even with that ambiguity, they report that in their inverse configuration, the FNO digital twin correctly identifies the damaged member in one hundred two out of one hundred sixteen withheld crossings and accurately recovers the severity, with a mean absolute error of zero point zero three <ref:2610.08822#pg3,the FNO digital twin correctly identifies the damaged member in 102>.

Jane: That accuracy number is quite solid for an unsupervised system. But what this means for someone listening to this show is that if you only listen to the results, you need to understand that while it finds the damage okay, it can’t always tell which specific piece of steel is broken in a complex spot.

Meng: If I were deploying this, I'd be concerned about those failure cases—the misidentification where it picks an adjacent member. That sounds like a liability if you're making maintenance decisions based on that output without further verification.

Tom: They admit they are still working on overcoming that distinguishability issue, but they point out that the vehicle’s advantage is that it passes over the entire span and captures damage effects in members near the supports which a fixed sensor at a single location can barely see.

Jane: So, for someone who just listens to this show, the implication is that this system moves beyond simply checking if *something* is damaged; it’s moving toward understanding where the damage is located along the bridge structure.

Lu: This paper sets up a framework where you combine physics-based modelling and machine learning for continuous monitoring, which they call "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing". It shows how this can work with drive-by inspection.

Tom: The conclusion is that while it’s still a numerical case study right now, the authors plan to validate this twin against field data from the Old Ada Bridge in Japan in their future work.

Jane: So, for us listeners, it means we should keep an eye on how these vehicle-integrated digital twin frameworks move from simulation and testing into actual deployment on real infrastructure. That’s what they are pointing toward next.

Conclusion: Jane: So we've been looking at how this vehicle-integrated digital twin works for bridge monitoring, and now we need to talk about what this paper is actually called and who wrote it, right?

Tom: Yeah, so the title is "A Vehicle-Integrated Approach to Digital Twin Deployment for Bridges Through Drive-By Sensing". It sounds like a mouthful.

Jane: It really does. I mean, if you strip away all the technical terms about Fourier Neural Operators and Young's modulus, what is the actual simple idea behind this paper?

Lu: The core idea is creating a digital replica of a bridge that gets updated with every time you drive over it, using vehicle movement as the way to measure structural health.

Meng: So it’s not just one static sensor checking a bridge; it’s like having a computer that learns how the bridge *should* move and compares that to how your truck actually moves.

Tom: Exactly. The authors are using this framework to combine physics with machine learning for continuous monitoring of both the bridge and the road conditions.

Jane: It's about making inspections scalable so you don't need a fixed sensor on every single bridge, which is a big problem right now when you have thousands of structures.

Lu: They use this digital twin to perform two things: first, predicting how a certain damage condition would cause the bridge to respond, and second, figuring out what the actual structural condition is from the measured response.

Tom: It’s using surrogate models built with a Fourier Neural Operator that learns how operating conditions map to vehicle responses in milliseconds.

Jane: So what does this mean for you as someone who just listens to this show? It means we're moving toward a system where structural information is gathered passively while you are driving by, rather than needing someone to physically inspect it every time.

Lu: The paper uses a specific bridge called the Old Ada Bridge in Nara Prefecture, Japan, as their testbed for this numerical case study.

Tom: And they're calling this a numerical case study right now because they still need to validate the twin against real field data from that bridge in their future work.

Jane: That’s where we’ll go next: we’ll see how robust this model actually is when it tries to handle messy, real-world data from those actual bridges.

Zihao Liu, Daigo Kawabe, Jiaji Wang, Chul-Woo Kim, §Mehrisadat Makki Alamdari

School of Civil and Environmental Engineering, University of New South Wales · KyoCenseo Inc. · Department of Civil Engineering, The University of Hong Kong · Department of Civil & Earth Resources Engineering, Kyoto University

cs.LG

Submitted: 2026-09-25

Updated: 2026-09-25

Comments: Extended abstract for 2nd International Conference on Engineering Structures (ICES2026)

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 72/100

The gist: The gist A neural operator digital twin of a real truss bridge was trained on simulated vehicle crossings with member-level damage in the drive-by configuration, and it correctly identifies damaged

Key concepts

Fourier Neural Operator (FNO): This is a type of machine learning model designed to learn complex mappings between different functions. In this context, it learns the relationship between the physical conditions of the bridge and how a vehicle interacts with it, allowing for fast prediction without needing slow, detailed structural simulations.
Digital Twin: A digital twin is a virtual replica of a physical asset that receives continuous updates from real-world data. This paper uses one to create a computational model of the bridge that can predict its behavior under various conditions and infer damage based on sensor readings.
Vehicle-Bridge Interaction (VBI): VBI refers to the structural response of a bridge when a vehicle crosses it. The study uses this interaction data, captured by sensors on an inspection vehicle, as the primary input for the digital twin to assess bridge health and detect potential damage.
Inverse Configuration: This is the part of the model where researchers take measured axle responses from a crossing and try to figure out which specific structural member is damaged and how severe that damage is. The model uses this input to output the location and severity of the defect.

Terminology

Digital Twin — 1× in this paper · explained in 6 episodes

Summary

The gist

A neural operator digital twin of a real truss bridge was trained on simulated vehicle crossings with member-level damage in the drive-by configuration, and it correctly identifies damaged members in 102 out of 116 withheld crossings while accurately recovering the severity.

The Problem and Approach

Ageing and deteriorating bridge infrastructure is a growing challenge globally, where conventional assessment methods like visual inspections are subjective and fixed sensor networks are costly to deploy across large inventories <ref:2610.08822#pg3> Drive-by, or indirect, inspection promises a scalable middle path <ref:2610.08822#pg2>. The approach taken in this paper is to create a digital twin of the bridge: a computational replica, updated with every crossing, that embeds the bridge-specific physics the inversion needs <ref:2610.08822#pg4> This twin performs two complementary tasks: predicting the response that a hypothesised structural condition would produce, and inferring the condition from the measured response.

The Digital Twin Model

The testbed is the Old Ada Bridge in Nara Prefecture, Japan, which is a simply supported steel truss with a main span of 59.2 m and a width of 3.6 m The digital twin itself is the neural operator surrogate described below Its training data are generated by an FE model of the truss, whose 29 members fall into 4 groups (bottom chord, top chord, verticals and diagonals) Damage is represented as a reduction in Young’s modulus of one entire member The structural state is therefore a vector of 29 severities. The operator is a Fourier Neural Operator (FNO) [2], in which the learned mapping is parameterised spectrally and acts directly on the response's frequency content.

Training and Inference

In the forward configuration, the 29 severities are entered as constant input channels, and the operator returns the response measured at the vehicle's two axles over the entire crossing Because damage alters the response far less than the crossing itself excites it, the operator is trained on the residual, the response difference between the damaged and healthy crossings. In the inverse configuration, the axle responses are the input, and the operator returns the damaged member and its severity through two output heads The operator is trained on 1,160 simulated crossings, 40 per member, at severities between 0.10 and 0.60.

Performance Results

In the forward direction, the predicted vehicle response closely tracks the FE solution, as shown in Figure 4 The FNO requires only a small fraction of the computational cost of the FE solution. The inverse configuration correctly identifies the damaged member in 102 of the 116 withheld crossings and accurately recovers the severity, with a mean absolute error in the estimated severity of 0.03. A key finding is that prediction accuracy depends on how strongly the damage manifests in the measured response, and a misidentification often selects a member immediately adjacent to the true one within the same chord.

Conclusion

The moving vehicle is advantageous in this respect because it passes over the entire span, capturing damage effects in members near the supports which a fixed sensor at a single location can barely see. The present work is a numerical case study, and future work will validate the twin against field data from the Old Ada Bridge [6].

ACKNOWLEDGEMENTS

This work is supported by the Australia-Japan Innovation Fund (AJIF).

REFERENCES

[1] Calderón Hurtado A., Makki Alamdari M., Atroshchenko E., Chang K.-C. & Kim C.-W. (2024) A datadriven methodology for bridge indirect health monitoring using unsupervised computer vision, Mechanical Systems and Signal Processing, 210, 111109

[2] Kovachki N., Lanthaler S. & Mishra S. (2021) On universal approximation and error bounds for Fourier neural operators, Journal of Machine Learning Research, 22(290), 1-76

[3] Kaewnuratchadasorn C., Wang J. & Kim C.-W. (2024) Neural operator for structural simulation and bridge health monitoring, Computer-Aided Civil and Infrastructure Engineering, 39(6), 872-890

[4] Chen R., Kim C.-W. & Wang J. (2025) Real-time surrogate model of vehicle-bridge interaction using a multiple-input neural operator, Automation in Construction, 180, 106576

[5] Chen R., Kim C.-W. & Wang J. (2026) Physics-informed neural operator for forecasting vehicle-induced bridge vibration trajectories, Engineering Applications of Artificial Intelligence, 176, 114708

[6] Kim C.-W., Zhang F.-L., Chang K.-C., McGetrick P.J. & Goi Y.

Improvements for AI systems

The improved system can perform real-time damage localization by using the vehicle is a moving sensor that passes over the entire span and responds to damage anywhere along the bridge, which enables the FNO digital twin to identify and localise damage. This capability allows for continuous monitoring without relying on costly fixed sensors.
The system can provide rapid structural state estimation by utilizing the inverse configuration, as the operator returns the damaged member and its severity through two output heads, accurately recovering the severity in 102 of the withheld cases with a mean absolute error of 0.03.
The system's predictive capability can be made practical for continuous evaluation because it can closely reproduce the vehicle response of the full FE simulation, several orders of magnitude faster, making continuous forward evaluation practical.

Abstract

Ageing bridge infrastructure is a growing global concern, yet conventional Structural Health Monitoring (SHM) systems are costly and difficult to scale, and routine visual inspections remain subjective. Drive-by, or indirect, bridge inspection, in which a sensorised vehicle recovers structural information from vehicle-bridge interaction (VBI) and vehicle-road interaction (VRI) responses, offers a scalable alternative. However, key challenges remain unresolved, including separating bridge responses from road roughness, detecting damage under normal traffic, and generalising across diverse bridge types. This paper presents a vehicle-integrated digital twin framework that unifies physics-based modelling and machine learning for continuous monitoring of bridge and road conditions. The framework comprises three pillars. First, surrogate models of VBI and VRI are constructed using a Fourier Neural Operator that learns function-to-function mappings from operating conditions to vehicle responses. Trained on both simulated and field data, these surrogates deliver millisecond-scale inference, replacing computationally intensive full-order analyses. Second, the design of a custom electric inspection vehicle, its sensor layout, and signal processing chain are optimised through Bayesian optimisation to maximise bridge information yield while suppressing road and vehicle noise. Unsupervised damage-assessment pipelines based on adversarial autoencoders, matrix profiles, and transformer architectures have been developed and validated to process the resulting vehicle data. Third, the complete workflow is validated through coordinated multi-site field trials in Australia and Japan, covering a range of bridge types, traffic conditions, and environmental settings.