A deep‐learning framework for classifying the type, location, and severity of bridge damage using drive‐by measurements

This paper proposes a new deep‐learning framework for drive‐by bridge condition monitoring. The proposed approach represents a bridge monitoring regime that enables the presence, type, location, and severity of bridge damage to be identified purely from measurements taken on a passing vehicle, without needing any pre‐measured training data. The computational framework adopts a numerical vehicle–bridge interaction (VBI) model, which is automatically calibrated using only the vehicle vibration measurements. The calibrated model is used to generate labeled training data, eliminating the practical difficulties associated with collecting training data for damaged bridge scenarios. The numerically simulated data are used to train a convolutional neural network that can then classify the actual damage characteristics of a bridge. The method is tested using a laboratory‐scale VBI model, and results show that the algorithm can accurately identify the presence of seized bearings and cracking in the bridge beams. The algorithm shows good accuracy when identifying the type and extent of damage. The location of bearing damage can generally be identified; however, the location of cracking is less accurate at lower damage levels. The proposed framework represents a significant improvement on existing techniques for indirect bridge monitoring.


INTRODUCTION
Transportation networks rely on well-maintained infrastructure to allow them to operate efficiently without disruption or delays.Bridges are particularly important as the failure of a bridge can have catastrophic consequences and can result in significant delays, which can last for months while repair works are completed.Bridges that fail to meet serviceability requirements (e.g., excessive deflection or cracking) may need to be closed or have weight restrictions implemented while they undergo strengthening works.In more extreme cases, sudden struc-by a trained inspector on a periodic basis.The frequency of inspections can vary between countries; however, the time between inspections for any given bridge is often as long as 5-6 years (Hearn, 2007).Thus, visual inspection techniques cannot provide a continuous understanding of the condition of a bridge.Where there are particular concerns about a bridge, or when specific monitoring is required, structural health monitoring (SHM) is often carried out using sensors mounted directly on the structure (Favai et al., 2014).Lin et al. (2022) developed an approach for structural damage detection that aimed to identify features of the measured structural response, which are damage sensitive and domain-invariant (i.e., those that are not affected by environmental effects, modeling error, etc.).Their domain adaption neural network was designed and trained on the data from both a numerical model and the real structure at the same time and was shown to outperform a traditional convolutional neural network (CNN) based on laboratory testing of a steel beam.The method requires acceleration measurements recorded directly from the beam, and damage was simulated during testing by using a lumped mass attached at different locations.Yulong Zhang et al. (2022) also carried out laboratory testing of a beam, where a lumped mass was used to simulate the effects of damage.Their approach used acceleration measurements recorded from the beam to identify damage from the nonlinear structural response.
The proposed method used a sliding window to calculate a probability density function from the time history of the vibration signal, and changes in the density function were used as an indicator of the presence of damage.Youqi Zhang et al. (2019) carried out vibration-based damage classification using a CNN and tested their approach using three case studies: (i) a T-shaped steel beam in the laboratory, (ii) a short steel girder bridge that had been moved to a test site, and (iii) an in-service steel girder bridge.Again, damage was simulated by artificially changing the structural properties by adding lumped masses or clamping plates to the structure to change the structural stiffness.
Their method was shown to be capable of classifying damage using the measured acceleration responses from the vibrating structure.However, it should be noted that measurements are required to be taken from the healthy structure and also the damaged structure in order to train the CNN, something that is not practical in real applications.Pavlou (2022) proposed an alternative approach for estimating fatigue damage estimation for structures that experience cyclical loading.A new cycle-counting algorithm was proposed, and the loading sequence and its effect on fatigue damage was considered.Similar vibration-based approaches have also been proposed for identifying the dynamic properties of buildings for SHM purposes.Various signal processing algorithms have been proposed to facilitate modal identification from the measured vibration signals.Perez-Ramirez et al. (2016) proposed a new approach for modal parameter identification of civil structures, using the synchrosqueezed wavelet transform.They demonstrated the application of the approach on a four-story steel frame structure and also on a bridge, and the results showed improved frequency and damping identification accuracy, compared to other signal processing approaches.Zhijun Li et al. (2017) applied a similar approach for modal identification and SHM of super high-rise buildings.They tested their method using measured signals from the 555-m tall Lotte World Tower in Seoul, Korea, using measurements taken during the construction stage.Their approach was capable of extracting both natural frequencies and damping ratios from the signals recorded on the structure.Amezquita-Sanchez et al. ( 2017) also tested a novel approach using data recorded from the Lotte World Tower.They proposed a multiple signal classification algorithm and used the empirical wavelet transform and Hilbert transform to successfully identify the natural frequencies and damping ratios of the tower.
Besides vibration-based approaches, various studies have used image analytics to automatically identify and classify structural damage based on photographs or imagery from structural inspections.Zheng et al. (2022) demonstrated the ability to identify cracks and measure crack widths based on image analysis.They proposed a CNN-based algorithm that was tested in the laboratory initially and then verified using imagery from a real bridge.Zou et al. (2022) proposed a post-earthquake damage identification approach in which images were analyzed to assess the component failure mode and the relevant damage level.The method combined object detection and recognition with finite element (FE) modeling to infer the failure mechanism and estimate the severity of damage.Pan and Zhang (2022) proposed a dual attention deeplearning network for identifying and classifying surface defects on steel members.They tested their algorithm using a database of over 6000 images that had been taken from high-frequency cameras, with defects being labeled manually.
These "direct" SHM systems allow the behavior of a bridge to be monitored over time and can provide detailed information to bridge owners, meaning that anomalies or unusual changes in structural behavior can be investigated in a timely manner.However, all of these approaches require measurements to be recorded directly from the structure or close-up photographs of the structural components.While this is often considered appropriate for bridges of particular importance, or where there is a specific concern over the condition of the bridge, the bespoke nature of most "direct" SHM approaches along with the costs and health and safety implications mean that they are generally prohibitive for large-scale monitoring campaigns (Agdas et al., 2016).As such, there has been a strong move toward using "indirect" techniques for monitoring bridges (Cerda et al., 2014), with drones (Flammini et al., 2016), imaging (Kalaitzakis et al., 2018), and various technologies now enabling remote inspections to be carried out in great detail.Using indirect techniques can eliminate many of the financial, safety-related, and logistical obstacles associated with direct SHM.Catbas and Avci (2022) provide a review of some of the latest novel developments in bridge health monitoring, including both direct and indirect methods.

Drive-by bridge monitoring
One method of indirect monitoring that has received much attention in research developments is known as "drive-by" bridge monitoring.Drive-by techniques utilize in-vehicle sensors to infer the condition of a bridge from measurements taken while the vehicle crosses the bridge.This method allows data to be collected on an ongoing basis and, most importantly, is scalable, meaning that multiple bridges could potentially be monitored without any significant increase in cost or measuring equipment.Yang et al. (2004) originally proposed the concept, proving that the bridge frequency can be extracted from measurements taken on a passing vehicle.Since the concept was originally proposed, there have been significant developments in this research field as discussed in great detail in the recent review papers by Wang et al. (2022) and Malekjafarian et al. (2022).Drive-by approaches have been used to monitor not only the natural frequency of a bridge but also bridge damping (González et al., 2012), mode shapes (Malekjafarian & OBrien, 2017), or novel concepts such as the operating deflection shape ratio between following axles (Corbally & Malekjafarian, 2022a).It has also been shown that the drive-by method can be used to infer the road profile on a bridge or estimate the rotational stiffness at the bridge supports (Feng et al., 2023).Zhenkun Li et al. (2023) proposed to combine information from two vehicles on a bridge to facilitate damage detection.Signals measured on a vehicle that was temporarily parked on the bridge were combined with data on a passing "sensing" vehicle.
The authors noted that the mass of the parked vehicle should be large enough that it influences the overall measured frequency of the vehicle-bridge system.Its location was changed intermittently to raise the sensitivity of the bridge's natural frequencies to local damage.If the mass of the vehicle is known, the expected change in frequency can be calculated, and the damage location can be accurately inferred based on a combination of the results from when the vehicle is parked at different locations.Numeri-cal simulations demonstrated that the concept could work; however, the requirement for a vehicle of known mass to be parked at different locations on the bridge during testing may limit the scalability of the approach.
With increasing quantities of SHM data being collected, there has been a shift toward the application of machine learning and data-driven techniques for damage detection in bridges.These techniques have been widely applied for direct SHM methods, but only in the very recent past have machine learning and artificial intelligence approaches been applied to drive-by bridge monitoring, as summarized below.

1.3
Artificial neural networks (ANNs) Malekjafarian et al. (2019) adopted an ANN for drive-by bridge health monitoring.The ANN was trained using simulated measurements from a series of vehicle crossings on a bridge and then used to predict the vehicle vibration and its fast Fourier transform (FFT).Results of numerical tests showed that differences between the FFT of the measured signals, compared to those predicted by the ANN, can be used to detect damage.Corbally and Malekjafarian (2022b) extended the approach to examine the influence of temperature effects on accuracy.They showed, using numerical tests, that damage-related changes can still be identified even when temperature effects are considered, if the temperature is included as part of the training process.
Although these studies demonstrate the ability to detect the presence of damage and its progression over time, they cannot quantify the extent of the damage or identify the type and location of the damage.Zhenkun Li et al. (2023b) trained a deep auto-encoder model with the frequency components of measured signals from a passing vehicle.
The method was tested experimentally in the laboratory, and the auto-encoder was trained using measurements from the vehicle crossing the healthy bridge.Damage detection capabilities were then tested using increasing levels of lumped mass attached to the bridge.The differences between the measured data and the reconstructed frequency responses were used as indicators of damage.
Results showed that the proposed method could identify the presence of the added mass with an accuracy of 86.2% across all of the different damage cases considered.

Convolutional neural networks
Locke et al. (2020) adopted a CNN-based approach for indirect bridge monitoring.They considered operational and environmental variables as inputs to the CNN.The method was successful at damage classification but does not easily identify the progression of damage over time and also requires labeled data from the damaged condition to train the CNN.This information will rarely, if ever, be available in practice.Sarwar and Cantero (2021) also adopted a CNN-based algorithm for drive-by bridge monitoring.Using detailed numerical simulations of 5-axle trucks, they demonstrated that the method is successful in detecting the progression of damage over time.In a separate paper, the same authors proved that the temperature influence can also be removed (Sarwar & Cantero, 2022).Again, however, these studies can only detect the presence of damage but cannot classify the type, magnitude, or location of the problem.
Hajializadeh presented two studies that also used CNNs for drive-by monitoring.In one of these papers (Hajializadeh, 2022a), an advanced numerical model was used to simulate the interaction between the train, track, ballast, and bridge.The continuous wavelet transform map of the simulated on-train vibrations was used to train the CNN with data from a series of different damage cases.The other paper by Hajializadeh (2022b) adopted a similar CNN-based algorithm where real measurements from a scaled model of a train and bridge were used as inputs to the CNN.Images of the time-frequency spectrograms were adopted as training inputs for the CNN.Both studies demonstrate good results; however, it appears that premeasured data for different damage scenarios must be available to train the CNN.In practical applications, it is very unlikely that this information will be available.Sarwar and Cantero (2023) proposed a probabilistic approach for data-driven bridge monitoring, which allows data from passing vehicles to be combined with information measured directly on the bridge (Sarwar & Cantero, 2023).Numerical simulations were used to test the approach, with multiple measurement scenarios being considered.The probabilistic deep neural network was trained with a sample of labeled data from the healthy bridge condition and under various damaged conditions, and then the ability of the trained network to accurately classify the signals into different damage conditions was evaluated.Results showed that the algorithm can accurately classify damage characteristics, particularly when information is available from the sensors on the bridge with classification accuracy reaching levels as high as 95%-100% for some scenarios.Unsurprisingly, classification accuracy was notably lower for the cases when only vehicle-based measurements were available, with accuracy levels ranging from 44% to 84% across the different scenarios considered.It is noted, again, that this method relies on the availability of labeled data from all bridge conditions, which will rarely be available in reality.

1.5
Other approaches Lan et al. (2023) proposed a time-domain signal processing algorithm for drive-by bridge inspection.The proposed method aims to remove the effects of measurement noise and variation in vehicle speed while reducing the dimensionality of the measurement data.The processed data are input into various machine learning algorithms to test damage detection capabilities.Laboratory experiments demonstrated that the method can effectively identify the presence of lumped masses attached to the bridge, with improvements in accuracy and computational efficiency achieved by using the proposed signal processing technique.The authors noted that their method requires a labeled dataset from the healthy bridge and from each damaged condition and acknowledge that this may not be feasible to obtain in a real-world scenario.Cheema et al. (2022) presented a new approach that utilizes a so-called Uniform Manifold Approximation and Projection (UMAP) algorithm along with a clustering technique.The method facilitates the classification of the measurements based on the bridge condition at the time of measurement.Liu et al. (2019) aimed to estimate the location and severity of damage using a multi-task learning approach that was tested using lumped masses attached to the bridge to simulate damage.Their algorithm can estimate the location and extent of damage with a high level of accuracy; however, it cannot distinguish between different damage types.In a different study by the same research group, an innovative approach was proposed that trains a supervised machine learning model using in-vehicle measurements for crossings over a particular bridge so that it can then be used to detect damage on a different bridge, using a transfer learning framework (Liu et al., 2020).The algorithm is trained to recognize features that are sensitive to damage and are similar from one bridge to another.Multi-task learning allows the algorithm to learn features that are shared between bridges.The results are very promising for the purposes of damage detection and even for damage localization and quantification; however, the algorithm does not provide an indication of the type of damage, and the laboratory testing only considers lumped masses attached to the bridge as a proxy for damage.In addition to this, the method still requires labeled training data from a bridge in different damage conditions to enable the transfer learning approach to be applied.For in-service bridges, it will almost never be feasible to collect this information.Zhenkun Li et al. (2023a) note that many drive-by bridge monitoring techniques focus on monitoring changes in the fundamental frequency of the bridge and often disregard the higher-frequency components of the measured signals, which can also contain pertinent information about the bridge condition.They included higher-frequency components and adopted mel-frequency cepstral coefficients to train a support vector machine to classify damage.Results from laboratory experiments, where added masses were used to represent damage, demonstrated that the higher-frequency components of the responses contained much more information about the bridge condition.Damage detection errors were within 5% when a heavy vehicle was used for testing.However, labeled data from each damage case are required to train the machine learning algorithm, which could be considered a practical limitation of the approach.
Despite these significant research efforts in the field of drive-by bridge monitoring, it is clear that most approaches focus on identifying the presence of damage, with only a few studies actually attempting to quantify the type, location, or magnitude of the damage.For the studies that have demonstrated the ability to classify the exact bridge damage characteristics, they typically rely on a labeled training dataset, containing measurements from the healthy bridge and also from the bridge after having experienced damage.This is a significant limitation, which is overcome by the framework presented in this paper by adopting a calibrated numerical model to generate labeled training data for different damage conditions.Although the use of numerical models to generate labeled training data is not uncommon (Seventekidis et al., 2020), calibrating a numerical model to replicate the true behavior of an in-service bridge usually requires sensors to be installed on the bridge, which conflicts with the overarching goal of drive-by monitoring.The proposed framework also overcomes this challenge by allowing the numerical model to be automatically calibrated using only measurements taken from the passing vehicle so that sensors on the structure are not required.
This paper proposes the first computational framework that can detect the presence of damage and also classify the specific damage characteristics, solely using sensors inside a passing vehicle, without any need for sensors on the bridge, or any need to measure pre-labeled training data for different damage scenarios.A numerical model of the vehicle-bridge interaction (VBI) is automatically calibrated using the vehicle vibration data recorded during a number of passages over the bridge.Then the model is used to simulate various damage scenarios and generate labeled training data to feed into an optimized deep-learning network, which can subsequently be used to classify bridge damage based on measured vehicle vibrations.
This deep-learning framework for drive-by bridge monitoring represents the first holistic approach that can be used to estimate the type, location, and severity of dam-age.The framework is presented in a generic way so that it can be extended in the future to allow more detailed models or algorithms to be incorporated depending on accuracy requirements.The approach is then tested using measured data from a laboratory-scale VBI model.

PROPOSED DEEP-LEARNING FRAMEWORK FOR DRIVE-BY BRIDGE DAMAGE CLASSIFICATION
The proposed framework for drive-by bridge monitoring aims to facilitate not only the detection of damage but also to allow the type, location, and severity of damage to be estimated.The proposed deep-learning algorithm adopts a classification-based CNN that is trained using signals for various types, locations, and severities of damage.It is noted that from a practical perspective, it is extremely unlikely that a situation will exist where training data for all potential damage types or scenarios will be available for an in-service bridge.To overcome this problem, the proposed framework, as per Figure 1, utilizes a numerical VBI model to allow in-vehicle measurements to be simulated for a range of damage scenarios.The simulated signals can then be used to train the CNN so that it can recognize how the in-vehicle measured response behaves for different types, locations, and severities of damage.Once the CNN has been trained, it can be used for ongoing monitoring of the bridge to classify the bridge condition and quantify the exact characteristics of damage when it arises.
Of course, developing an FE model, which can capture the bridge behavior, along with that of the vehicle, is not straightforward.In order to create the FE model of the bridge, some information must be known about the bridge geometry, materials, and the bearing articulation.This information is typically available to infrastructure managers, and generally historical inspection records or design drawings exist, which provide sufficient detail to allow the bridge to be modeled with a reasonable degree of accuracy.However, it is well documented that numerical models do not necessarily truly represent the actual in-service behavior of a bridge (Brownjohn et al., 2003).When measured data from the structure are available, models can be calibrated to reflect the true in-service response.However, this is not feasible in the case of drive-by monitoring, where one of the primary objectives is to avoid the requirement for taking any measurements directly from the bridge.In order to overcome this challenge, the proposed framework uses an autocalibration approach that calibrates the VBI model using only the measured vehicle response.The autocalibration approach varies the parameters of the FE model, using a particle swarm optimization (PSO) calibration algorithm, until the model outputs provide a good fit to the measured data.It is noted that the nature-inspired PSO algorithm is chosen in this study as it allows complex multivariable optimization problems to be solved in a computationally efficient manner.However, as summarized by Siddique and Adeli (2015b), many nature-inspired optimization algorithms exist.These include water drop algorithms (Siddique & Adeli, 2014), the music-inspired harmony search algorithm (Siddique & Adeli, 2015a), spider monkey optimization (Akhand et al., 2020), and gravitational search algorithms (Siddique & Adeli, 2016), among others.The PSO algorithm was shown to work well for the autocalibration of the VBI model presented in this paper; however, the deep-learning framework could easily be adapted to incorporate other optimization algorithms depending on the nature of the problem.The following sections provide a description of each stage in the deep-learning framework.

Data collection
The initial phase of the process involves the collection of vehicle vibration data as it crosses over the healthy bridge.These data will be used to calibrate the VBI model.The vibration signals measured on the vehicle each time it crosses the bridge should be recorded, along with the average vehicle speed during the passage over the bridge.Vehicle crossings should be recorded at various speeds to allow the FE model to be calibrated across a range of vehicle speeds.

Model calibration
For the model calibration, the parameters of the FE model are varied until the outputs from the VBI model give a good fit to the signals measured during the previous step.
When carrying out drive-by bridge monitoring, only data recorded on the vehicle will be available, without any measurements directly from the bridge, so the calibration of the model must be carried out using the vehicle-vibration measurements.
The proposed framework requires a model that can be used to replicate the dynamic VBI behavior during a vehicle passage over the bridge.More complex models will likely be able to provide a more accurate representation of the true VBI; however, they will also be more difficult to develop and calibrate, depending on the level of information that is available.For the purposes of demonstrating the proposed framework, a generalized VBI model is adopted in this paper as depicted in Figure 2. The vehicle is modeled by a quarter car, and the bridge is represented by an FE model that uses two-noded Euler-Bernoulli beam elements with two degrees of freedom per node (vertical displacement and rotation).The load at the point of contact between the vehicle tire and the bridge is distributed to the adjacent nodes in the FE model using Hermite shape functions.The bridge, of length , is discretized into a series of beam elements of length  with mass per unit length , second moment of area , and Young's modulus .The boundary conditions are imposed at the supports by restraining the relevant degrees of freedom (vertical translation) at the support nodes, and rotational springs, of stiffness   , are included at these nodes to allow rotational restraint at the bearing locations to be considered.
The quarter-car vehicle model considers two lumped masses,   and   , representing (i) the mass of the vehicle body and (ii) the mass of the axle + wheel, respectively.The vertical motion of these masses represents the two degrees of freedom of the model.A spring-dashpot system, to represent the stiffness (  ) and damping (  ) of the suspension, connects the two lumped masses.At the contact point between the vehicle and the bridge, the tire stiffness is modeled using a spring of stiffness   .The chosen VBI model is deemed suitable for the purposes of demonstrating the framework, as it can capture the primary global components of the VBI and can also be constructed without requiring significant detail about the bridge.Figure 3 summarizes the autocalibration approach.
The first step is to generate the preliminary VBI model, using known information about the vehicle and the bridge.It is assumed that the parameters of the vehicle used for testing are known with some degree of certainty, and these are not varied during the autocalibration process.For the bridge model, it is assumed that general geometrical information and information about the support conditions and bridge material are known, as this information is typically available to infrastructure owners.Once the preliminary VBI model parameters have been defined, the parameters that are to be varied should also be defined.Any of the parameters of the VBI model can be included in the optimization.A PSO routine is then adopted to solve the multi-variable optimization problem.Further details on PSO are presented by Poli et al. (2007); however, Figure 3 gives a general overview of how the approach is applied.
It is acknowledged that the calibration process assumes that data can be measured from the vehicle crossing over the bridge in a healthy condition; however, there may be existing damage already.In most countries, infrastructure owners carry out extensive inspection regimes, so visual inspection records (or other information) could be consulted at the outset to gain an understanding of the condition of the structure when commencing monitoring.This would allow any existing damage or issues to be considered when calibrating the FE model.The proposed approach can then be implemented to identify any further changes from the existing bridge condition.
As part of the PSO algorithm, random combinations of different parameters are generated, within specified limits, and the model is updated to include these values.For each set of bridge parameters (i.e., each particle "" within the swarm of size ""), VBI simulations are carried out to replicate the measured signals from each vehicle crossing "" in the calibration dataset.For each crossing, the difference between the measured frequency response and the simulated frequency response is quantified using the root mean square error ().The overall average  for a given set of bridge parameters is then evaluated as an indicator of the accuracy of the model, when using those bridge parameters.For each of the particles within the swarm, the average  value is calculated after each iteration of the PSO algorithm until convergence is achieved or all of the "N" iterations are completed.The bridge properties that provide the lowest  value across all of the vehicle passages are selected.

Generating labeled training data
Once the VBI model has been calibrated, it can then be used to model various damage scenarios.In this way, the model can be used to generate training data for multiple damage scenarios, for which training data would otherwise be unavailable.Various types of bridge damage can be included in the VBI model, and the response of the vehicle passing over the bridge at different speeds for each damage scenario can be simulated.For the purposes of demonstrating the deep-learning framework, two different damage types are considered in this paper: (i) seized bearings and (ii) cracking of the longitudinal bridge beams.The effects of seized bearings are modeled by increasing the rotational spring stiffness values at the supports.Cracking is modeled by reducing the stiffness of the elements of the FE model in the vicinity of the crack location.More details on the modeling approach for the specific damage cases considered in this paper are provided in Section 3.

Damage classification
The drive-by bridge monitoring framework adopts a deeplearning algorithm to identify the bridge condition and to classify the type, location, and severity of damage.A CNN Choosing a suitable CNN architecture and the appropriate values of the various hyperparameters is not a trivial task.The choice of architecture and hyperparameters is often based on trial and error or based upon previous models that have been developed for specific tasks.The proposed framework in this paper adopts an optimization procedure, using PSO, to define the most appropriate architecture and hyperparameters for the classification of bridge damage from the simulated data.The PSO optimization varies the number of convolutional layers, the number and size of the filters within each convolutional layer, the dimensions of maximum pooling after each convolution, the initial learning rate, and the maximum number of epochs for training.The optimization identifies the CNN that gives as close as possible to 100% classification accuracy across all of the training signals.More details on the PSO algorithm are provided by Poli et al. (2007), and Section 3.4 demonstrates how the PSO optimization was applied to determine the CNN architecture used in this study.

Health monitoring
Once the CNN has been trained, it can then be used for ongoing health monitoring of the bridge.Signals from subsequent passages over the bridge can be classified using the CNN, and the bridge condition can be estimated, along with the type, location, and severity of the damage, when it is present.This step, and the basis of the framework, assumes that the FE model and the damage modeling techniques used can adequately replicate the effects of real damage on the bridge.Of course, this assumption is likely to be dependent on the complexity of the vehicle and bridge model used and also the damage modeling techniques.It is therefore important that consideration is given to the chosen modeling techniques, depending on the level of detail or accuracy required.The quarter-car vehicle model and 1D FE beam model adopted in this paper to demonstrate the concept are deemed appropriate to capture high-level global changes due to damage and allow a quick screening process to alert road owners of damage.This model can be generalized to many bridges on the network with only a small number of parameters required.These parameters (length, number of spans, material, support conditions, etc.) will be known to infrastructure owners and will enable the autocalibration process to refine the parameters accordingly.For specific bridges where further detail or greater accuracy is required, more detailed numerical models can be adopted.

TESTING OF DEEP-LEARNING FRAMEWORK
The proposed deep-learning framework is demonstrated using a laboratory-scale VBI model and its ability to detect

Description of laboratory model
The bridge model, depicted in Figure 4, consists of a 5-mm deep, 600-mm wide steel deck plate, with six equal-angle steel beams (20 × 20 × 3 mm) bolted to the underside of the deck.The bridge span is 2000 mm between the centerlines of the supports.The supports were designed to allow rotation at each end to replicate simply supported end conditions.There is a 65 mm overhang, beyond the centerline of the support at each end, making the total bridge length 2130 mm.
The support system includes a steel cylindrical shaft that spans transversely between rotational bearing units at either side of the support.A custom 3D printed section, made from plastic, surrounds the cylindrical steel shaft to act as a bearing shelf for the beams, with slots and pre-drilled holes allowing the beams to be bolted into position on the bearing shelf.The bearing shelf is clamped to the steel shaft using a series of locking nuts that eliminate rotational slip between the plastic bearing shelf and the steel shaft.The laboratory model was designed to represent the behavior of a typical concrete beam and slab bridge subjected to a crossing vehicle.Modeling a bridge, at a scale below 1:1, to replicate realistic behavior, is not straightforward, and as such, different scaling laws were applied to design the bridge geometry.In this case, the bridge length was scaled by a factor of 1:12.5, and the geometry was designed to ensure (i) a realistic structural form, (ii) similar span-deflection ratio between the real bridge and scaled bridge subjected to a fully loaded typical truck, and (iii) realistic natural frequencies/dynamic interaction between the vehicle and bridge.Timber approach spans were included at either end of the bridge to allow the vehicle to speed up and slow down before and after each vehicle passage.
A remote-controlled vehicle was used for the experiments.The Axial SCX10 III Jeep (Axial-Racing, 2022), a commercially available vehicle, has three speed settings and allows some of the vehicle components to be modified or replaced.The suspension of the vehicle combines oilfilled shocks with coilover springs, and the damping level of the suspension can be adjusted.When carrying out testing, the original plastic casing on the vehicle was removed to facilitate the installation of wireless accelerometers as shown in Figure 5. Four accelerometers were mounted on the vehicle using plastic 3D-printed brackets to enable vibration measurements to be recorded on the vehicle body at the front and back and also at the center of both axles.A sampling frequency of 512 Hz was used when measuring acceleration signals for all of the experiments.

Description of numerical VBI model
The numerical model presented previously in Figure 2 was used to model the behavior of the remote-controlled vehicle crossing the laboratory-scale bridge.The FE model of the bridge, before autocalibration, was created using the known geometry of the bridge.Figure 6 shows the geometry of the 600 mm wide cross-section, which was used to calculate the second moment of area for the beam elements.The FE model was 2130 mm long and was divided evenly into 33 elements.Simply supported end conditions were applied at the second and second-last node of the bridge by restraining the vertical degrees of freedom to replicate the position of the supports at either end of the bridge.No rotational stiffness was included at the supports before the autocalibration process.The preliminary VBI model was created using a second moment of area of  = 7.07 × 10 4 mm 4 , a Young's modulus of  = 210 GPa, material density of  = 8000 kg∕m 3 , and structural damping of  = 1%, modeled using Rayleigh damping where the damping ratio is assumed to be the same for the first two modes (Clough & Penzien, 1993).The properties used for the quarter-car vehicle model were estimated based on the datasheets for the vehicle, and where parameters were not known, reasonable estimates were used.The mass of the vehicle axle and wheels,   , was assumed to be 0.475 kg, with the suspension stiffness taken as   = 455 N/m, suspension damping   = 20 Ns/m, and the tire stiffness taken as   = 2500 N/m.Exact measurements of the vehicle properties were not taken, to replicate a scenario where the vehicle properties are not known with exact certainty.
The VBI was modeled through the coupled equations of motion of the combined vehicle and bridge system, presented in Equation (1).
Equation ( 1) describes the dynamic behavior of the whole system, combining both the degrees of freedom of the bridge and those of the vehicle, where [  ], [  ], and [  ] are the global system (i) mass, (ii) stiffness, and (iii) damping matrices, respectively.The force vector, {}, includes the forces applied to each degree of freedom in the overall model at every timestep as the vehicle crosses over the bridge.These time-varying matrices are updated at each timestep in the analysis to consider the moving location of the vehicle on the bridge.MATLAB software is used to perform the VBI analysis, and the Wilson-θ approach (Tedesco et al., 1999) is adopted to solve the equations of motion of the system.

Autocalibration of numerical VBI model
In order to calibrate the FE model, the remote-controlled vehicle was driven over the bridge 30 times at its lowspeed setting, crossing the bridge in the same direction each time.It is noted that the mean vehicle speed at this setting was 0.7 m/s, which is equivalent to something in the range of 14-33 km/h on a full-scale bridge, depending on the scaling rule that is applied (as discussed by Corbally & Malekjafarian, 2023).The PSO algorithm was implemented to calibrate the parameters of the FE model so that the simulated measurements matched the real measurements from the laboratory as closely as possible.The first natural frequency of the bridge, estimated from the numerical model, was seen to be 8.9 Hz, and it was noted that the measured frequency, during a passage of the vehicle, was shown to be slightly lower due to the influence of the additional mass of the vehicle on the bridge.To calibrate the FE model, the frequency spectrum for each of the measured signals was calculated, and the PSO algorithm was applied to minimize the  between the frequency spectrum generated from the quarter-car model vibration and the frequency spectrum of the measured signals in the laboratory.The portion of the frequency spectrum between 7 and 9.2 Hz was used as this was seen to contain the fundamental bridge frequency.The PSO algorithm was implemented using a swarm size of  = 75 particles, and the algorithm was run for a maximum of  = 20 iterations.Table 1 shows the parameters that were varied in the optimization and also outlines the range over which each parameter was varied.The range was selected for most parameters by allowing them to vary by up to 30% above or below the preliminary values estimated from the bridge geometry and material.
Figure 7 shows the acceleration frequency spectrum for the signal measured on the front axle of the vehicle during a single passage across the bridge at 0.72 m/s.The simulated frequency spectrum from the numerical model is also shown for comparison, for the original uncalibrated VBI model and the calibrated version.It can be seen that the calibrated VBI model provides a much better fit to the measured frequencies, particularly in the vicinity of the peak, which corresponds to the fundamental frequency of the bridge.The model does not appear to accurately capture the frequencies at the lower range, even after calibration, likely due to the simplicity of the numerical model that cannot capture all of the 3D behavior of the real vehicle-bridge system.
Table 2 summarizes the results of the PSO optimization following the calibration of the model using the data from the 30 vehicle passages.The different bridge proper-ties, as described by the symbols in the first column, are as defined in Table 1.The autocalibration process was carried out using measurements from each of the sensors separately.Table 2 provides the calibrated parameters for each of the four sensor locations.It can be seen that all of the parameters are reasonably consistent, irrespective of the sensor location used for calibration.The PSO algorithm has reduced the stiffness, Young's modulus, and density values below the original assumed values and increased the rotational stiffness at the left support but not at the right-hand support.This is an interesting result, which suggests that there may already be some level of rotational restraint at the left support of the bridge model, which is not present at the right-hand support.This may be related to the alignment of the support during construction, which could inhibit the rotation of the support if the bearing shelf is not exactly perpendicular to the span.It is also interesting to note that the calibrated damping value is consistently estimated to be zero, which is not necessarily realistic, but yields the best fit between the measured and simulated data.This is likely related to the fact that the signals recorded during the vehicle passage are very short, and the dynamics are primarily governed by forced vibrations as the vehicle crosses the bridge, meaning that the measured signals used to calibrate the model will not contain any free vibrations, where the damping effect would be more influential.

CNN architecture optimization
To facilitate damage classification, the calibrated FE model was used to generate labeled signals for multiple vehicle passages across the bridge at various speeds and under different damage conditions, with all crossings going the same direction across the bridge, in the same way as carried out in the laboratory.The architecture and hyperparameters for the CNN were optimized using the PSO algorithm to simultaneously maximize the classification accuracy of the simulated data and minimize the computational time for training.For each of the vehicle passages over the bridge, the frequency spectrum between 7 and 9.2 Hz was extracted and labeled depending on the damage condition of the bridge.Various damage scenarios were considered as described in greater detail in the subsequent sections.The frequency spectra were re-sampled at 0.05 Hz intervals to ensure that the size of all input data to the CNN was the same (i.e., vectors of 1 × 45 frequency values), and a cubic spline function was used to interpolate between the measured values.It is noted that the portion of the signals used were those from when the vehicle was in contact with the bridge and that the resampling of the FFT graphs was carried out to ensure uniformity of the x-values between the FFT graphs for each individual vehicle passage used to train the CNN.As part of the PSO algorithm, six parameters were considered as variables, with the range over which they were allowed to vary described in Table 3.
The PSO algorithm was run with a swarm size of  = 15 particles for a maximum of  = 20 iterations if convergence was not achieved before this.For each particle in the PSO, a CNN architecture was generated using parameters sampled from the ranges outlined in Table 3, and the objective function was evaluated.The objective function for a particular particle "" in the PSO, (  ), considered a combination of classification accuracy and computational time as per Equation (2): where   is the classification accuracy, defined as the percentage of training samples for which the damage label was correctly predicted (in the range of 0%-100%), and   is a reduction factor, which depends on the training time of the CNN (i.e., the Central Processing Unit (CPU) time taken for the CNN to be trained).The values for   are provided in Table 4, and where training times fell between these values, linear interpolation was used to evaluate the reduction factor.In this way, for lower computational times, greater importance is placed on the classification accuracy, with the relative importance of classification accuracy reducing for CNNs that take longer to train.This was done to ensure that computational times were not excessive when identifying the optimized CNN parameters.For example, if two CNNs both achieved 100% accuracy and one took 10 min to train, but the other took 1 h, the value of the objective function for the second CNN would be 95% rather than 100%.The final CNN architecture found from the PSO optimization process consisted of five convolutional layers, with the number and sizes of the filters in each layer and the size of the max-pooling operation for each layer summarized in Table 5.The optimal value of the initial learning rate was found to be 0.061, and the maximum number of epochs for training was 837.The convolutional filters and max-pooling filters are all applied with a stride of 1, and F I G U R E 8 Tightening screws to lock pillow-block bearing units.
each convolutional layer includes batch normalization and Rectified Linear Unit (ReLU) nonlinear activation.It is also noted that an additional input layer is incorporated into the CNN to account for the vehicle speed associated with each input signal.The optimization of the CNN architecture took approximately 4 days to complete using an i7-9700 CPU @ 3 GHz with 8 cores and 64 GB RAM.The final classification accuracy of the optimized CNN was 93.31%.

Detection of seized bearings
In this section, the ability of the algorithm to detect changes in the rotational restraint in the bearings is investigated.The supports at either end of the laboratory bridge model consist of cylindrical steel shafts, which are housed within pillow-block bearing units to allow the steel shaft to rotate and replicate simply supported bridge bearings.
There is a plastic bearing shelf surrounding the steel shaft, in between the bearing units, onto which the bridge beams are seated.In order to replicate the situation where the bearings become seized, and there is a reduction in the level of rotation at the supports, the locking screws on the bearing units were tightened, as shown in Figure 8, to reduce the level of rotation of the support.Tightening these screws prevents the steel shaft from rotating within the bearing unit; however, there will still be some level of rotation at the support due to the flexibility of the plastic bearing shelf and the fixing connections of the beams to the shelf.Given that the exact level of rotational restraint, due to tightening the screws, could not be easily quantified, the labeled data generated from the VBI model considered three levels of rotational fixity, representing low (30%), medium (60%), and full (100%) fixity.It was estimated that locking two of the bearings at one end of the bridge would only have a small influence on the rotational restraint of the overall support, and the actual damage scenario was considered to represent a "low" level of fixity.Three different damage scenarios were considered by tightening the screws in the bearing units at each end of the bridge as described in Table 6.
The VBI model was then used to simulate different cases of seized bearings by increasing the rotational spring stiffness values at either end of the bridge model and for both ends simultaneously.Three levels of rotational restraint were considered, 30%, 60%, and 100%, at each end of the bridge, resulting in nine possible damage cases along with the healthy case.The VBI model was used to generate vehicle crossing data, for the healthy bridge and each of the potential damage locations and severities, with 20 passages simulated for each case.Following the training of the CNN with the simulated data, the results from 120 vehicle crossings over the laboratory bridge were used to estimate the actual damage condition of the bridge, with the damage label estimated based on that with the highest probability when batching 5 sequential vehicle passages.Signals from 30 crossings over the bridge in the healthy condition and 30 in each of the B1, B2, and B3 conditions were used for the damage classification.
Figure 9 shows the classification results, and it can immediately be seen from Figure 9a that the algorithm can distinctly identify the difference between the healthy condition and the damaged condition.Only in two cases of the 24 estimates does it assume the bridge is healthy, when it is actually damaged.In terms of the estimated level of damage to the bearings, as shown in Figure 9b, again, the algorithm accurately identifies the magnitude of damage.For the case of the left bearing (B1), it accurately detects the level of damage in all but one case, and for the right bearing (B2), it is correctly identified every time.For the case of both bearings being seized (B3), it over-estimates the damage in one case and underestimates another but is correct in the for the remainder of cases.Finally, the algorithm is also capable of identifying the location of the problem (Figure 9c), with a good level of accuracy, except for when the damage is at the right support.When the left bearing is damaged, it correctly identifies the location of the problem in all but one case, and when both bearings are damaged, again, only one case is incorrectly located.When the right-hand bearing is damaged, the results are poor, and the estimated location is incorrect every time.
Overall, it can be seen that the algorithm is able to identify the presence of damage and the level of damage with a high degree of accuracy, and in most cases, the damage can be located with a good level of certainty, with the exception of when it is located at the right-hand support.

Detection of cracking in the bridge beams
Cracking was introduced at midspan of the laboratory bridge model by cutting 2 mm wide slots into the bridge beams as shown in Figure 10.Three damage scenarios were considered, labeled C1-C3, as depicted in Figure 10b, to capture the effect of increasing levels of damage to the bridge.
In order to test the ability of the proposed approach to estimate the actual damage characteristics, the calibrated VBI model was used to generate labeled training data.
Cracking was simulated at various locations and various severities.The effect of cracking was simulated by reducing the stiffness of the FE beam elements in the vicinity of the longitudinal crack location.Cracking was simulated every 250 mm along the length of the bridge, and damage levels according to the number of cracked beams (from 1 to 6) were simulated at each location, resulting in 48 possible damage cases along with the healthy case.The VBI model was then used to generate labeled training data for the healthy condition and each of the different damage cases, by simulating 20 vehicle crossings over the bridge for each damage condition.
When considering the effects of cracking in the FE model, the change in the second moment of area of the overall bridge cross-section relative to the fully intact section, due to a 14 mm crack in one or more beams, was initially calculated.Then, the second moment of area of the two elements on either side of the crack location, in the FE model, was reduced by the same proportion.It is noted that the stiffness of the calibrated FE model had been adjusted as part of the calibration, and therefore the change in the second moment of area was applied as a reduction in the proportion of the calibrated stiffness rather than an absolute reduction.
Once the CNN had been trained using the simulated data, 50 passages of the remote-controlled vehicle crossing the laboratory bridge were performed, in the healthy condition, and in each of the three damage conditions C1-C3 (as per Figure 10).The trained CNN was then used to classify the bridge condition, again using batches of five vehicle passages for each estimate.Figure 11 shows the results, with Figure 11a illustrating that the algorithm can almost exactly distinguish between the healthy bridge condition and that when it has experienced cracking.Only two cases existed where the bridge was actually undamaged, but the algorithm incorrectly identified that there was cracking present.Figure 11b outlines the estimated number of cracked beams, for each case.It is evident that the algorithm tends to overestimate the number of cracks.For the case when only one beam was cracked, the algorithm indicates that there were three cracked beams, with some estimates indicating four cracked beams and one indicating that all six beams were cracked.For the case of two cracked beams (C2), only one estimate correctly identifies the damage level, with the majority of estimates indicating that five beams had been cracked.
For the damage case C3, the number of cracks was also overestimated in a number of cases.It is clear that the algorithm has a tendency to overestimate the extent of damage, and retrospective analysis of the signals from the bridge, compared to the simulated data from the model, showed that the real impact of cracking in the bridge was greater than that predicted by the modeling.This highlights the importance of the modeling approach, and its accuracy in reflecting the damage type being assessed.
Figure 12 summarizes the estimated location of cracking along the length of the bridge, for each of the three midspan cracking severities.It is observed that the algorithm fails to clearly identify the correct location of the cracking, with the exception of the most extreme damage case, where most of the estimates are at the correct location.For damage cases C1 and C2, the results indicate that the damage is most likely to be at the ¾ span location, although based on the distribution of estimated damage locations, it is certainly clear that the damage level cannot be accurately pinpointed.

Differentiating between cracking and seized bearings
A final test of the algorithm was carried out to examine its ability to differentiate between healthy, cracked, or seized bearing damage scenarios when the CNN was trained with simulated data from all of these potential scenarios.The simulated vehicle passages over the bridge in all of the damage scenarios examined in the previous two sections were used to train the CNN.Then the CNN was used to estimate the damage characteristics during the passages of the remote-controlled vehicle over the healthy bridge and for scenarios B1-B3 and C1-C3.Figure 13 outlines the results, where it is evident that the algorithm, in most cases, can identify the bridge condition (Figure 13a).The main exception is for the case of one cracked beam (C1), where the algorithm incorrectly indicates that the bearings are seized.It is also worth noting that in a small number of cases, when the bridge is healthy, the algorithm incorrectly identifies damage.
For the cases where cracking was correctly identified as being present in the bridge beams, Figure 13b shows the estimated severity of cracking.The results are similar to previous findings, where the algorithm has a tendency to overpredict the number of cracks, but for the case of C3, half of the estimates accurately identify that there are three cracks in the bridge.For the cases where seized bearings were correctly identified, Figure 13c shows the estimated level of restraint.It can be seen that almost every case correctly estimates 30% rotational fixity in the bearings.
Finally, Figure 14 shows the estimated location of the damage cases that were correctly identified.Figure 14a,b shows the estimated location of cracking for damage cases C2 and C3, respectively (C1 is not shown as it was never correctly identified).It can be seen, similar to the previous results, that C2 cannot be accurately located, but when three cracks exist in damage case C3, the algorithm correctly identifies the location of damage for most of the estimates.Figure 14c shows the results for the seized bearings damage cases, demonstrating that the accuracy is reduced, compared to the situation when only bearing damage was simulated previously.Damage to the right-hand bearing is correctly located for three of the six batches.Damage to the left bearing is always incorrectly located and assumed to be at the right-hand support.For the case where both bearings were seized, the algorithm cannot clearly identify this and tends to indicate that the right-hand bearings are seized.Overall, it is seen that the accuracy of locating the damage has reduced when all of the bearing and cracking damage cases are used to train the CNN simultaneously.

DISCUSSION
The proposed deep-learning framework for drive-by bridge monitoring has been tested using data from a laboratoryscale VBI model.The framework is implemented using a quarter-car and FE beam numerical VBI model.The vehicle speeds in the laboratory correspond to full-scale speeds of approximately 15-30 km/h, which do not represent full highway speeds but would still allow quick data acquisition during low-traffic periods (e.g., at night) where a monitoring vehicle could be driven across a bridge, without requiring any closures of traffic lanes.
The ability of the proposed framework to identify and classify changes in the rotational restraint of the bearings and cracking in the longitudinal girders is examined.It is observed that when considering seized bearings, in isola- tion, the presence of damage can be detected with a high degree of accuracy and the approach can also accurately classify the location and extent of the damage.When using the approach to identify cracking in the bridge, it is possible to identify the presence of damage in the bridge beams; however, the model overestimates the severity of the cracking.This highlights a limitation of the proposed approach, which assumes that the damage model can accurately represent the effects of real damage on the bridge.It also emphasizes the importance of choosing a suitable level of detail in the numerical model and the damage modeling approach, depending on the level of accuracy required, as it is an underlying assumption of the proposed framework that the damage model is capable of accurately replicating the effects of real damage.A more detailed 2D or 3D VBI model could be adopted; however, this type of model would be more difficult to calibrate and would likely be less scalable for adaptation for use on other bridges.The location of the cracking is not accurately identified by the algorithm until three beams are cracked, at which point most of the estimates correctly identify the location of cracking.It should also be noted that the crude modeling approach for cracking, whereby a full reduction in stiffness is applied to the elements near the crack location, may not truly capture the actual effects of cracking.Sinha et al. (2002) proposed a more detailed approach for modeling the effects of cracking in a beam, subject to transverse vibration.Using Sinha's approach, the reduction in stiffness propagates linearly away from the crack location, with the extent of influence being related to the overall depth of the beam.During preliminary testing of the proposed framework, Sinha's method was adopted as a method for simulating cracking in the FE model.However, it was found that the actual measured changes in bridge frequency, due to the cracks, were much greater than that observed with the application of Sinha's model.Applying the full stiffness reduction to the elements in the vicinity of the crack resulted in changes to the bridge frequency, which were much closer to those measured directly from the bridge and hence was considered to be a more appropriate method for modeling the effects of damage to the laboratory bridge model.This is not necessarily surprising because Sinha's method was developed to model the influence of a crack in an individual beam and was not developed to reflect the overall influence of a cracked beam on a beam and slab bridge such as that presented in this study.Nevertheless, it highlights the importance of using a suitable modeling approach for damage to maximize the effectiveness of the proposed deep-learning framework, as the accuracy of the proposed method will be strongly influenced by the ability of the chosen damage model to replicate the real effect that damage will have on the VBI.
For the final test, the calibrated VBI model is used to simulate training data for multiple potential damage scenarios for seized bearings and cracking in the deck.It is shown that when the CNN is trained to recognize a significant number of potential damage scenarios, the accuracy drops slightly.The algorithm can still identify the difference between the healthy bridge, the cracked bridge, or when the bearings have seized, for all cases except that where a single crack exists.The level of rotational restraint in the bearings can be accurately quantified, but the level of cracking is still overestimated.There is a noticeable reduction in the ability of the algorithm to accurately locate the damage when the left bearing is seized or when both bearings have seized.In the case of cracking, the crack location is not accurately identified until three cracks have been induced in the bridge.
It is acknowledged that the damage conditions considered in this study represent significant levels of damage, which in practice would be related to an extreme event (e.g., earthquake, vehicle collision, explosion, etc.).However, the ability to quickly detect such damage across multiple bridges on the network would be extremely beneficial for infrastructure owners, particularly in the aftermath of an extreme event.Despite this, it is expected that further development of this method (e.g., through more detailed vehicle/bridge models and damage mechanisms) would enable the detection of smaller levels of damage, which could be detected at an early stage and be used to enable preventative maintenance.
The proposed framework is adaptable, and future development of this approach could investigate the use of more detailed modeling of the vehicle and bridge, along with different damage models.This would potentially enhance the ability of the approach to detect different damage types and the transverse location of damage.In addition, this could facilitate the detection of smaller levels of damage allowing earlier detection and enabling pro-active maintenance interventions.Another suggested future development for the proposed approach would be to extend the autocalibration process, and the CNN classification process, to consider additional information beyond the FFT of the measured vibrations.For example, the time histories of the vibration responses could be used directly, enhancing the ability of the algorithm to differentiate between symmetric damage locations.In addition, other known damage-sensitive features could be used for one or more measured signals on the vehicle.This would potentially enable more accurate classification of damage by including additional information during the training and classification processes.Other future areas of development relate to consideration of operational and environmental factors such as vehicle speed, other vehicles on the bridge, or ambient temperature or wind-induced vibrations.The proposed method could be coupled with existing algorithms that allow such factors to be included (Corbally & Malekjafarian, 2022b;Sarwar & Cantero, 2022) or may be extended to consider the influence of these factors directly.
Overall, despite the fact that there are a small number of false positive results, which incorrectly indicate the presence of damage, the majority of the estimates are correct and when repeating the experiments, the overall accuracy of the proposed approach is very good.It is also worth noting that the laboratory-scale model may not truly represent the behavior of a real bridge.Despite being designed to maintain the VBI and dynamic response of a real structure, it may not fully capture all of the characteristics of the reallife scenario.However, as it would not be feasible to introduce damage in a real bridge for the purposes of testing different damage scenarios, the laboratory-scale experiments are deemed to provide an appropriate method to test the framework.Nevertheless, the results demonstrate that the proposed drive-by bridge monitoring framework shows great potential for damage detection and classification, beyond the capabilities of most existing approaches.

CONCLUSION
This paper presents a novel approach for identifying the type, location, and extent of damage in a bridge, using only measurements taken from a traversing vehicle.The proposed deep-learning framework can be used to estimate the damage characteristics of a bridge without needing to attach any sensors directly to the bridge and without the need for measured training data for different damage scenarios.
The framework is tested using laboratory experiments where different levels of rotational restraint and cracking are introduced into the bridge.Results show that the presence of damage can be accurately detected and that the type of damage, along with the extent and location, can be correctly classified in most cases.However, it is seen that the model tends to overestimate the level of cracking.In addition, it is shown that the damage location is not always correctly identified, particularly when increasing the number of potential damage scenarios that are used to train the CNN.
The proposed framework for drive-by bridge monitoring represents the first approach that can be used to identify damage and to quantify the type, location, and extent.The method provides greater certainty of bridge damage characteristics than other drive-by monitoring approaches and represents a useful tool for indirect monitoring of bridge safety, which also shows good potential for future expansion to increase its capabilities.

A C K N O W L E D G E M E N T S
The research conducted in this publication was funded by the Irish Research Council (IRC) under Grant Number GOIPD/2023/1588.
Open access funding provided by IReL.

F
I G U R E 1 Deep-learning framework for drive-by damage classification.CNN, convolution neural network; FE, finite element.

F
I G U R E 3 Autocalibration routine for VBI model.RMSE, root mean square error.is adopted to classify the bridge condition based on measured in-vehicle vibrations, and the calibrated VBI model is used to generate labeled data to train the CNN.

F
Laboratory bridge model in (a) final constructed condition and (b) during construction, showing support conditions and placement of beams.and classify bridge damage characteristics is examined for different damage scenarios.

F
Vehicle model and locations of accelerometers.

F
Cross-section of laboratory bridge model (dimensions in millimeter).

TA B L E 1
Overview of variables for finite element autocalibration.Modeled versus measured frequency spectrum for front axle vibrations.

TA B L E 6
Overview of damage scenarios for seized bearings.

F
Damage classification for seized bearings, classification of: (a) damage type, (b) damage severity, and (c) damage location.

F
I G U R E 1 0 Damage to individual bridge beams at midspan (a) slot cut into beam and (b) summary of damage scenarios.F I G U R E 1 1 Damage classification for cracking of bridge beams, classification of: (a) damage type and (b) damage severity.

F
Identification of seized bearings or cracking in beams: (a) classification of damage type, (b) severity of cracking, and (c) extent of bearing fixity.

F
Estimated location of damage for (a) cracking damage case C2, (b) cracking damage case C3, and (c) seized bearings.

Accelerometer position on the vehicle Property Unit Axle (front) Axle (back) Body (front) Body (back)
TA B L E 2 Overview of calibrated model parameters.