A data-driven modeling method for dynamic weighing of regional bridges

Abstract

The present application discloses a data-driven modeling method for dynamic weighing of regional bridges, comprising the following steps: constructing a bridge structure response simulation dataset based on vehicle weight, constructing a domain-aligned simulation dataset by conducting adversarial training between measured bridge structure response data and the simulation dataset; dividing the global weighing model space of all bridges into multiple independent subspaces according to the bridge type, and adopting a multi-task learning strategy in deep learning to jointly construct a pre-trained base model including the corresponding independent subspaces; determining the combination of pre-trained base models of bridges in the region, and performing cascaded progressive fine-tuning from a seed bridge-specific weighing model to a bridge cluster-specific weighing model. The beneficial effects of the present application are: accurate vehicle load estimation can be provided for data-driven modeling in the region through the clustering of bridges in the region, and the established universal pre-trained base model can be used to achieve modeling using a small amount of measured data or cascaded data fine-tuning, thereby effectively improving modeling efficiency.

Description

Technical Field

[0001] The present application relates to the technical field of bridge health monitoring, and in particular to a data-driven modeling method for dynamic weighing of regional bridges. Background Art

[0002] With the rapid development of sensor technology and artificial intelligence, deep learning-based methods for identifying moving loads driven by large-scale monitoring data have become a research hotspot in bridge health monitoring and traffic management. These methods typically install various response sensors on bridges, such as strain sensors, accelerometers, fiber optic sensors, or deflection measurement devices based on DIC (digital image correlation) principles. Combined with real-world vehicle weight data provided by physical weighing systems at the bridgehead, deep learning models are constructed to establish a nonlinear mapping relationship between vehicle loads and structural responses, thereby enabling estimation of vehicle axle weight or total weight.

[0003] However, existing data-driven bridge weighing technologies generally have the following limitations:

[0004] (1) High requirements for training data sets: The accuracy of deep learning models depends largely on the quality and quantity of training data. When the training data is insufficient or unrepresentative, the scarcity and imbalance of the data set will lead to a significant decrease in the prediction accuracy and generalization ability of the model, which places high demands on the accumulation and cleaning of training data.

[0005] (2) “One bridge, one model” leads to low modeling efficiency: Existing methods usually train models for a single bridge, and the model performs well only on the trained bridge. When faced with a cluster of regional bridges with different structural parameters and environmental conditions, each bridge needs to be modeled and trained separately, resulting in low modeling efficiency and difficulty in meeting the needs of regional vehicle load identification.

[0006] (3) Training data sources rely on the WIM system: For existing bridges in the region that do not have a physical WIM weighing system installed, obtaining high-precision vehicle load data is difficult, which becomes a bottleneck for model training. The expensive deployment and maintenance costs of the WIM system limit the widespread application of existing data-driven models in regional bridge clusters. Summary of the Invention

[0007] One of the objectives of the present application is to provide a data-driven modeling method for dynamic weighing of regional bridges that can solve at least one of the defects in the above-mentioned background technology.

[0008] To achieve at least one of the above purposes, the technical solution adopted in this application is: a data-driven modeling method for dynamic weighing of regional bridges, comprising the following steps:

[0009] S100: Construct a bridge structure response simulation dataset based on vehicle weight;

[0010] S200: Establish a lightweight monitoring system for regional bridge clusters, conduct adversarial training on measured data of bridge structural responses under actual traffic loads and simulated datasets, and then construct a domain-aligned simulation dataset;

[0011] S300: The global weighing model space of all bridges is divided into multiple independent subspaces according to bridge type. A multi-task learning strategy in deep learning is used to jointly construct pre-trained base models including the corresponding independent subspaces.

[0012] S400: Determine the pre-trained base model combination of bridges in the area and construct a measured data set; perform cascade progressive fine-tuning from a seed bridge-specific weighing model to a bridge cluster-specific weighing model based on the obtained measured data set; wherein, a bridge with a dynamic weighing system is used as a seed bridge.

[0013] Preferably, in step S100, the construction of the bridge structure response simulation data set includes the following process:

[0014] S110: Select bridges with different structural parameters, including bridge type, material parameters, cross-sectional properties, and span parameters, and establish a refined finite element model;

[0015] S120: Statistically analyzing the traffic flow information collected by the actual system to obtain a vehicle load spectrum including gross weight, axle weight, and wheelbase information, and establishing a vehicle dynamics analysis model that conforms to regional characteristics;

[0016] S130: Constructing a multi-operating condition parameter set, including vehicle speed range, lane position, and road surface roughness, to form an operating condition parameter space covering typical vehicle-bridge coupling;

[0017] S140: Generate bridge-vehicle-operating condition combinations through Latin hypercube sampling, perform vehicle-bridge coupling finite element simulations under typical operating conditions, and obtain a simulation data set of the bridge structure response.

[0018] Preferably, in step S200, the construction of the domain alignment simulation dataset includes the following process:

[0019] S210: Perform adversarial training on the simulated dataset S and the pseudo-real dataset R' generated in the forward direction, and perform adversarial training on the real measured dataset R and the pseudo-simulated dataset S' generated in the forward direction to obtain the adversarial loss function L GAN ;

[0020] S220: Reconstruct the simulation dataset S and the real measured dataset R through the cycle consistency constraint to obtain the cycle consistency loss function L cycle ;

[0021] S230: Perform weighted summation of the adversarial loss function and the cycle consistency loss function to obtain the total loss function L total And minimize; the total loss function L total =L GAN +λL cycle ; Where λ represents the weight of the cycle consistency loss;

[0022] S240: Keep the label of the original simulation dataset S, and replace the original simulation dataset S with the pseudo real dataset R' to obtain a domain-aligned simulation dataset.

[0023] Preferably, step S210 includes the following process:

[0024] S211: Input the simulation data set S and the real measured data set R into the generator G respectively S→R and G R→S , correspondingly generate pseudo real data set R' and pseudo simulation data set S';

[0025] S212: Input the real measured data set R and the pseudo real data set R' to the discriminator D at the same time R , input the simulated dataset S and pseudo-simulated dataset S' to the discriminator D at the same time S ;

[0026] S213: Setting the Discriminator D R The goal is to judge the real measured data set R as true and the pseudo-real data set R' as false; the discriminator D S The goal is to judge the simulation data set S as true and the simulation data set S' as false.

[0027] Preferably, step S220 includes the following process:

[0028] S221: Input the simulation data set S and the real measured data set R into the generator G respectively S→R and G R→S , correspondingly generate pseudo real data set R' and pseudo simulation data set S';

[0029] S222: Input the pseudo real dataset R' and pseudo simulation dataset S' to the generator G respectively R→S and G S→R , correspondingly generating the reconstructed simulation data set S" and the reconstructed measured data set R";

[0030] S223: Reconstruct the simulation data set S" and the simulation data set S through the cycle consistency constraint requirement for approximation, and reconstruct the measured data set R" and the real measured data set R for approximation.

[0031] Preferably, in step S300, data preprocessing is required before constructing the pre-trained base model, which specifically includes the following process:

[0032] S310: Input raw data, including dynamic deflection time series, bridge structure parameters and vehicle load labels;

[0033] S320: Set the target time series length T, and perform linear interpolation or compression on the original time series to make its length uniform to T;

[0034] S330: Set the maximum number of channels C max , fill the insufficient channels with zero and generate a channel mask matrix ; Among them, M[C]=1 means that the Cth channel is valid, and M[C]=0 means that the Cth channel is invalid;

[0035] S340: Each bridge type and parameter combination is considered as a task, which is further divided into a support set and a query set. The support set accounts for 80% of the total tasks, and the query set accounts for 20% of the total tasks.

[0036] Preferably, in step S300, the pre-trained base model includes a shared feature layer, a task-specific adaptation layer and a meta-learning module; the shared feature layer uses a shared neural network Extracting universal features of dynamic deflection signals ; where θ s represents the shared feature layer parameters, Indicates that invalid channels are set to zero; the task-specific adaptation layer is used for each type of bridge design task branch H k , combined with the shared common feature z and bridge parameter s k Predicted vehicle weight y k =H k (z,s k θ k ); The meta-learning module introduces a meta-learning mechanism in the task-specific adaptation layer to optimize the initial parameters θ s and θ k ; where θ k Indicates task branch parameters.

[0037] Preferably, the pre-trained base model needs to be trained after it is built. The specific training process is as follows: for each task branch H k Optimize task-specific parameters θ using support set data k ´; The support set data is passed through the support set loss function L support (θ s ,θ k ) to optimize the parameters. The specific expression is as follows:

[0038] ;

[0039] ;

[0040] Use the updated task-specific parameters θ k ´Evaluate model performance on query set data and update parameters θ s and θ k , the query set data passes the query set loss function L query (θ s ,θ k ´) to perform performance evaluation, the specific expression is as follows:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] Among them, α represents the inner loop learning rate, β represents the outer loop learning rate, and L total (θ s , {θ k}) represents the summary of the loss function of the total query set, k represents the total number of bridge types, x represents the dynamic deflection time series, y represents the measured value of vehicle weight, N support and N query Denotes the number of samples corresponding to the support set and query set, respectively, D support and D query Represents support set data and query set data respectively.

[0046] Preferably, the regional bridge cluster lightweight monitoring system includes sensors or a DIC machine vision system for collecting structural response data, and a traffic camera for monitoring traffic on the bridge deck. In step S400, the construction of the measured data set includes the following process:

[0047] S401: Vehicle weight data is collected through dynamic weighing systems, highway toll booths, and overload detection stations installed in the area;

[0048] S402: Tracking the status of vehicles of known weight on different bridges through bridge traffic cameras, thereby obtaining the vehicle's location, speed, and license plate information and uploading it to a local server database;

[0049] S403: Matching the identified vehicle crossing time and location information to improve the vehicle information in the database;

[0050] S404: The vehicle type identification is used to trigger the collection of structural response data segments, and causal constraints are introduced to ensure temporal correlation, thereby obtaining a measured data set that matches the vehicle weight with the bridge structural response.

[0051] Preferably, in step S400, the process of fine-tuning the bridge cluster-specific weighing model based on the seed bridge-specific weighing model is as follows:

[0052] S410: Directly supervise the seed bridge based on the collected measured data set, build a special weighing model for the seed bridge and continuously output vehicle weight prediction results;

[0053] S420: Using license plate recognition technology and vehicle spatiotemporal parameter information, the status of the same vehicle on different bridges is tracked to form a vehicle trajectory chain across the bridges;

[0054] S430: If the prediction error of a bridge is less than a set threshold, the prediction result is marked as a high-confidence pseudo label and used as fine-tuning data for other bridge-specific weighing models;

[0055] S440: All bridges in the area are considered as distributed dynamic weighing nodes. The vehicle weight prediction results output by the deployed bridge model are transmitted to other bridges through the model cascade mechanism.

[0056] S450: For non-seed bridges, initially fine-tune the pre-trained base model of the bridge using a small amount of high-confidence pseudo-labeled data; in the later stages, as high-confidence pseudo-labeled data accumulates, increase the adjustment range of shallow parameters;

[0057] S460: Re-evaluate the model’s performance on the query set after each fine-tuning session and adjust the model architecture or hyperparameters as needed;

[0058] S470: Once a bridge-specific weighing model has been fully fine-tuned, it is added to the model cascade network to provide more high-confidence pseudo-labeled data for other bridges.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] The regional bridge clustering can provide accurate vehicle load estimates for data-driven modeling in the region. The established general pre-trained base model can be fine-tuned using a small amount of measured data or cascade data to achieve modeling, thereby effectively improving modeling efficiency and significantly reducing dependence on expensive WIM systems or other real data sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a schematic diagram of the workflow of this application.

[0062] Figure 2A schematic diagram of the overall modeling architecture of this application. DETAILED DESCRIPTION

[0063] Below, the present application is further described in conjunction with specific implementation methods. It should be noted that, in the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like are intended to mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms should not be understood as necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0064] In the description of this application, it should be noted that for directional words, such as the terms "center", "horizontal", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and so on, indicating the orientation and position relationship are based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing this application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and cannot be understood as limiting the specific scope of protection of this application.

[0065] It should be noted that the terms "first", "second", etc. in the description and claims of this application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0066] In this application, unless otherwise expressly specified or limited, terms such as "mounted," "connected," "connect," and "fixed" should be understood in a broad sense. For example, they may refer to connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.

[0067] In this application, unless otherwise expressly specified or limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features being in contact not directly but through another feature between them. Moreover, a first feature being "above," "above," and "above" a second feature may include the first feature being directly above or obliquely above the second feature, or may simply mean that the first feature is higher in level than the second feature. A first feature being "below," "below," and "below" a second feature may include the first feature being directly below or obliquely below the second feature, or may simply mean that the first feature is lower in level than the second feature.

[0068] The terms "comprises" and "having" and any variations thereof in the specification and claims of this application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units expressly listed, but may include other steps or units not expressly listed or inherent to such process, method, product or apparatus.

[0069] One of the preferred embodiments of this application is as follows: Figure 1 As shown, a data-driven modeling method for dynamic weighing of regional bridges includes the following steps:

[0070] S100: Construct a bridge structure response simulation dataset based on vehicle weight.

[0071] S200: Establish a lightweight monitoring system for regional bridge clusters, conduct adversarial training on measured data of bridge structural responses under actual traffic loads and simulated datasets, and then construct a domain-aligned simulation dataset.

[0072] S300: The global weighing model space of all bridges is divided into multiple independent subspaces according to the bridge type, and the multi-task learning strategy in deep learning is used to jointly construct a pre-trained base model including the corresponding independent subspaces.

[0073] S400: Determine the pre-trained base model combination of bridges in the area and construct a measured data set; perform cascade progressive fine-tuning from a seed bridge-specific weighing model to a bridge cluster-specific weighing model based on the obtained measured data set; wherein, a bridge with a dynamic weighing system is used as a seed bridge.

[0074] It is understandable that the simulation data set can be simulated in simulation software. Common simulation software includes ANSYS, ABAQUS and MIDAS, etc., which can be selected according to the actual needs of technicians in this field. In a simulation environment, traffic load data under any working conditions can be simulated, that is, the obtained simulation data set can include the structural dynamic response data of all bridges in the area to all types of vehicles by default. As for the measured data set, it can only be collected on bridges equipped with weighing systems. Therefore, the data volume of the measured data set is much smaller than that of the simulation data set; then the present application compares and verifies the measured data set with the simulation data set through domain alignment, and then achieves high-precision vehicle load identification through a small amount of measured data, thereby significantly reducing the amount of calibration samples required for subsequent model training.

[0075] At the same time, a universal independent subspace can be constructed according to the types of various bridges, and then a universal pre-trained base model can be obtained for the bridges in the region by combining different independent subspaces, thereby achieving rapid adaptation of the bridge cluster in the region, avoiding the inefficiency of the traditional "one bridge, one model" problem, and greatly improving the modeling efficiency of the bridge cluster.

[0076] Finally, through dynamic pseudo-label generation and cascade fine-tuning, the model can be continuously optimized during continuous operation, significantly reducing the demand for expensive WIM systems, lowering hardware investment and maintenance costs, and providing an efficient, low-cost and scalable solution for intelligent traffic management and bridge health monitoring.

[0077] In this embodiment, the construction of the bridge structure response simulation data set in step S100 includes the following process:

[0078] S110: Select bridges with different structural parameters, including bridge type, material parameters, cross-sectional properties, and span parameters, and establish a refined finite element model.

[0079] It should be known that the types of bridges mainly include simply supported beams, continuous beams and continuous steel structures, the material parameters mainly include elastic modulus and Poisson's ratio, the cross-sectional properties mainly include cross-sectional size and cross-sectional type, and the span parameters mainly include span and number of spans.

[0080] S120: Statistically analyzing the traffic flow information collected by the actual system to obtain a vehicle load spectrum including gross weight, axle weight, and wheelbase information, and establishing a vehicle dynamics analysis model that conforms to regional characteristics.

[0081] It is understandable that the construction of the vehicle dynamic analysis model can also directly adopt the internationally accepted universal dynamic analysis model for three-axle and five-axle vehicles. According to the established vehicle dynamic analysis model, a vehicle model library based on the gradient change of vehicle gross weight can be constructed.

[0082] S130: Construct a multi-operating condition parameter set, including vehicle speed range, lane position, and road surface roughness, to form an operating condition parameter space covering typical vehicle-bridge coupling.

[0083] S140: Generate bridge-vehicle-operating condition combinations through Latin hypercube sampling, perform vehicle-bridge coupling finite element simulations under typical operating conditions, and obtain a simulation data set of the bridge structure response.

[0084] It should be noted that the structural response data of a bridge mainly include deflection, strain or acceleration.

[0085] It is understandable that the constructed simulation dataset includes input and output dimensions. The input dimension is: typical bridge parameter set {bridge type, material parameters, cross-section properties, span parameters} ∪ operating condition parameters {vehicle speed, lane position, road surface roughness} ∪ multi-point dynamic response time history {deflection 1, deflection 2, ..., deflection n} or {strain 1, strain 2, ..., strain n} or {acceleration 1, acceleration 2, ..., acceleration n}; the output dimension is: gross vehicle weight.

[0086] In this embodiment, the regional bridge cluster lightweight monitoring system constructed in step S200 includes a sensor or a DIC machine vision system for collecting structural response data, and a traffic camera for monitoring traffic on the bridge deck.

[0087] It should be noted that the specific structures and operating principles of sensors, DIC machine vision systems, and traffic cameras are well-known to those skilled in the art and therefore will not be elaborated upon here. The specific selection of sensors or DIC machine vision systems should be considered based on the bridge type, structural characteristics, and cost budget. The acquisition frequency should ideally be greater than 10 Hz. For ease of understanding, the following will illustrate several commonly used sensors.

[0088] Specifically, accelerometers are sensitive to bridge vibrations caused by vehicle loads, are relatively easy to install, and are reasonably cost-effective, making them suitable for capturing the dynamic characteristics of bridges. Strain gauges directly reflect the stress state of bridge structures and are more directly related to loads, but are relatively complex to install and sensitive to environmental factors (temperature and humidity), requiring better protection and compensation. Dynamic deflectometers directly reflect bridge deformation and are intuitively related to load magnitude. However, traditional deflectometers are complex to install and maintain, resulting in high costs. Non-contact, low-cost DIC machine vision systems can be used as an alternative.

[0089] As can be seen from the foregoing, due to the limited number of bridges equipped with WIM systems, the number of measured data sets for bridge structural responses based on vehicle weight is relatively small. Therefore, this embodiment uses a small number of measured data sets to compensate for differences in the simulation data set, effectively ensuring the accuracy of the simulation data set for all bridges within the coverage area. Specifically, in step S200, the two domains corresponding to the simulation and measured data sets are aligned using a domain alignment algorithm to compensate for differences.

[0090] Specifically, a small number of representative bridges with installed WIM systems or measured data are selected as target bridges. Measured data on the structural responses of the target bridges under actual traffic loads is collected. Domain differences between the simulated and measured data are then compared and analyzed. Sources of these differences include, but are not limited to, environmental noise and model simplification errors. A domain-adaptive alignment algorithm is then used to transform the simulation dataset generated in step S100 using both the simulated and measured data. This domain-aligned simulation dataset is then obtained by narrowing the data distribution differences between the simulated and real domains.

[0091] It should be noted that there are various types of domain adaptive alignment algorithms, including but not limited to Generative Adversarial Networks (GANs) and Cycle Generative Adversarial Networks (CycleGANs). Domain conversion involves adjusting the feature distribution of the simulated signal to approximate that of the measured signal of a real bridge, compensating for differences between the simulated model and the real bridge in terms of environmental noise, simplified boundary conditions, and other factors. For ease of understanding, the following detailed description of the domain conversion alignment process will use Cycle Generative Adversarial Networks (CycleGANs) as an example.

[0092] Specifically, such as Figure 2 As shown in Figure 2, the specific steps of using a cyclic generative adversarial network to perform domain conversion alignment to generate a domain alignment simulation dataset are as follows:

[0093] S210: Perform adversarial training on the simulated dataset S and the pseudo-real dataset R' generated in the forward direction, and perform adversarial training on the real measured dataset R and the pseudo-simulated dataset S' generated in the forward direction to obtain the adversarial loss function L GAN .

[0094] S220: Reconstruct the simulation dataset S and the real measured dataset R through the cycle consistency constraint to obtain the cycle consistency loss function L cycle .

[0095] S230: Perform weighted summation of the adversarial loss function and the cycle consistency loss function to obtain the total loss function L total And minimize; the total loss function L total =L GAN +λL cycle ; λ represents the weight of cycle consistency loss.

[0096] S240: Keep the label of the original simulation dataset S, and replace the original simulation dataset S with the pseudo real dataset R' to obtain a domain-aligned simulation dataset.

[0097] It can be understood that the domain adaptive alignment performed by CycleGAN resolves the difference between the distribution of simulation and real data. Through adversarial training and cycle consistency constraints, CycleGAN adjusts the noise level, boundary conditions and other characteristics of the simulation signal to be close to the actual measured signal of the real bridge, effectively compensating for the errors caused by the simplification of the simulation model. For example, the generator learns the statistical characteristics of the real data through adversarial loss, while the cycle consistency loss ensures the reversibility of the bidirectional mapping between simulation and real data, avoiding information loss in domain conversion. This process significantly reduces the deviation of the simulation data, allowing the model trained based on the domain-aligned simulation dataset to better adapt to real scenarios, reduce dependence on real vehicle weight data, and solve the problem of data scarcity. For ease of understanding, the specific training process of CycleGAN will be described in detail below.

[0098] (1) The structure of CycleGAN mainly includes the generator G S→R and G R→S , and the discriminator D S and D R .

[0099] (2) The forward generation and adversarial training process based on step S210 is as follows:

[0100] S211: Input the simulation data set S and the real measured data set R into the generator G respectively S→R and G R→S , correspondingly generating a pseudo real dataset R' and a pseudo simulation dataset S'.

[0101] S212: Input the real measured data set R and the pseudo real data set R' to the discriminator D at the same time R , input the simulated dataset S and pseudo-simulated dataset S' to the discriminator D at the same time S .

[0102] S213: Setting the Discriminator D R The goal is to judge the real measured data set R as true and the pseudo-real data set R' as false; the discriminator D S The goal is to judge the simulation data set S as true and the simulation data set S' as false.

[0103] It is understandable that the generator G S→R The goal is to input the simulated dataset S to generate a pseudo-real dataset R' so that the discriminator D RIt is difficult to distinguish between the pseudo-real dataset R´ and the real measured dataset R, that is, to "deceive" the discriminator D as much as possible R ; This part of the training uses the adversarial loss function L GAN (G S→R , D R , S, R) to drive. Generator G R→S The goal is to input the real measured data set R to generate a pseudo simulation data set S' so that the discriminator D S It is difficult to distinguish between the pseudo-simulation dataset S' and the simulation dataset S, that is, to "fool" the discriminator D as much as possible S ; This part of the training uses the adversarial loss function L GAN (G R→S , D S , R, S) to drive. In this way, L GAN = L GAN (G S→R , D R , S, R)+L GAN (G R→S , D S , R, S).

[0104] (3) The cycle consistency constraint training process based on step S220 is as follows:

[0105] S221: Input the simulation data set S and the real measured data set R into the generator G respectively S→R and G R→S , correspondingly generating a pseudo real dataset R' and a pseudo simulation dataset S'.

[0106] S222: Input the pseudo real dataset R' and pseudo simulation dataset S' to the generator G respectively R→S and G S→R , correspondingly generating the reconstructed simulation data set S" and the reconstructed measured data set R".

[0107] S223: Reconstruct the simulation data set S" and the simulation data set S through the cycle consistency constraint requirement for approximation, and reconstruct the measured data set R" and the real measured data set R for approximation.

[0108] It can be understood that after completing the adversarial training, the CycleGAN structure is used to align the distribution of the simulation dataset S with the real measured dataset R of the target bridge. That is, the label of the original simulation dataset S is kept, and the pseudo-real dataset R' is used to replace the original simulation dataset S to construct a domain-aligned simulation dataset, providing a more reliable data foundation for the subsequent training of the pre-trained base model.

[0109] In this embodiment, Figure 2 As shown, according to different types of bridges, the global weighing model space V of all bridges isN Divide into N independent subspaces, each subspace corresponds to a bridge type; for example, a simply supported beam subspace, a continuous beam subspace, etc. Each subspace trains its own base model based on the domain alignment simulation data set obtained in step S200, such as α1 = (1, 0, 0, ...) represents a pre-trained base model for a simply supported beam, α2 = (0, 1, 0, ...) represents a pre-trained base model for a continuous beam; and so on. V N Dedicated weighing model V for any bridge i It can be expressed by A={α1,α2,…,α N} means that V i =f(A); where the f function represents the fine-tuning process of the base model, then A is called the global weighted model space V N The pre-trained base model is denoted by N, where N is the dimensionality of the model space, i.e., the number of samples. If all bridges in the region are simply supported beams, the model space is a one-dimensional regional weighing model space V1, corresponding to A = {α1}. If the bridges in the region consist only of simply supported beams and continuous beams, the model space is a two-dimensional regional weighing model space V2, corresponding to A = {α1, α2}, and so on.

[0110] The vehicle weight inversion tasks of different bridge type subspaces can share some physical laws, but the specific characteristics are different. It can be regarded as forming a structural domain corresponding to each bridge type. Global weighing model space V N The pre-trained base model A={α1, α2, ..., α N We employ a multi-task learning strategy from deep learning to jointly construct the model. Training is performed using a multi-task network consisting of a shared feature extractor and multiple single-task output heads. The single-task output heads are structured differently, with each bridge type forming a corresponding domain. For ease of understanding, the following describes the construction of the pre-trained base model and the training process in detail.

[0111] Specifically, such as Figure 2 As shown, the data preprocessing process before building the pre-trained base model in step S300 is as follows:

[0112] S310: Input original data, including dynamic deflection time series x and bridge structure parameters s i and vehicle load label y i .

[0113] It should be noted that the dynamic deflection time series ; C i represents the number of sensors, that is, the number of channels corresponding to the bridge, T i Column length when representing data; bridge structure parameters s i Including stiffness, span and damping ratio, vehicle load label , indicating the vehicle weight.

[0114] S320: Set the target time series length T, and perform linear interpolation or compression on the original time series to unify its length to T.

[0115] S330: Set the maximum number of channels C max , fill the insufficient channels with zero and generate a channel mask matrix ; Among them, M[C]=1 means that the Cth channel is valid, and M[C]=0 means that the Cth channel is invalid.

[0116] For ease of understanding, the process of step S330 will be described in detail below using specific parameters.

[0117] Assume that the maximum number of channels of a bridge is C max =5, and the bridge is actually equipped with three sensors, each collecting response signals at time T = 4. The raw data collected by the three sensors is: Sensor 1 (Channel 1): [1.0, 2.0, 3.0, 4.0]; Sensor 2 (Channel 2): [0.1, 0.2, 0.3, 0.4]; Sensor 3 (Channel 3): [9.0, 8.0, 7.0, 6.0].

[0118] Because the maximum number of channels is 5, and only 3 channels actually have data, channels 4 and 5 need to be padded with zeros to make up the total 5 channels. Thus, channel 4 (padded): [0, 0, 0, 0]; channel 5 (padded): [0, 0, 0, 0]. This results in a 5×4 tensor X_padded, which can be used to generate the corresponding channel mask matrix M. The channel mask matrix M is a binary vector of length 5 that marks which channels are valid and which are padded. Specifically, M=[1, 1, 1, 0, 0]; where M[C]=1 indicates that the Cth channel contains real data, and M[C]=0 indicates that the Cth channel is an invalid channel padded with 0s.

[0119] S340: Each bridge type and parameter combination is considered as a task, which is further divided into a support set and a query set. The support set accounts for 80% of the total tasks, and the query set accounts for 20% of the total tasks.

[0120] Specifically, such as Figure 2 As shown, in step S300, the pre-trained base model includes a shared feature layer, a task-specific adaptation layer, and a meta-learning module. The shared feature layer uses a shared neural network Extracting universal features of dynamic deflection signals ; where θ s represents the shared feature layer parameters, Indicates that invalid channels are set to zero. The task-specific adaptation layer is used for each type of bridge design task branch Hk , combined with the shared common feature z and bridge parameter s k Predicted vehicle weight y k =H k (z,s k θ k ). The meta-learning module introduces a meta-learning mechanism in the task-specific adaptation layer to optimize the initial parameters θ s and θ k ; where θ k Indicates task branch parameters.

[0121] In this embodiment, the pre-trained base model needs to be trained after it is built. The specific training process is as follows:

[0122] Inner loop: for each task branch H k Optimize task-specific parameters θ using support set data k ´; The support set data is passed through the support set loss function L support (θ s ,θ k ) to optimize the parameters and support the set loss function L support (θ s ,θ k ) and task-specific parameters θ k The expression of ´ is as follows:

[0123] .

[0124] .

[0125] Where α represents the inner loop learning rate, k represents the total number of bridge types, x represents the dynamic deflection time series, y represents the measured value of vehicle weight, and N support represents the number of samples corresponding to the support set, D support Represents support set data.

[0126] Outer loop: Use the updated task-specific parameters θ k ´Evaluate model performance on query set data and update parameters θ s and θ k , the query set data passes the query set loss function L query (θ s ,θ k ´) for performance evaluation, query set loss function L query (θ s ,θ k ´) and the updated shared feature layer parameters θ s and task branch parameter θ k The expression is as follows:

[0127] .

[0128] .

[0129] .

[0130] .

[0131] Among them, β represents the outer loop learning rate, L total (θ s , {θ k}) represents the summary of the total query set loss function, N query Respectively represent the number of samples corresponding to the query set, D query Represents queryset data.

[0132] It should be noted that during the inner loop training, only the task branch parameters θ k Do a gradient update, during which the feature layer parameters θ are shared s Remain unchanged. During the outer loop training, the task branch parameters θ k and shared feature layer parameters θ s Perform an outer loop update.

[0133] It is understandable that this application introduces multi-task learning and meta-learning strategies to construct a universal pre-trained base model covering different bridge types. By jointly training a variety of simulation data (such as simply supported beams, continuous beams, etc.), the model shares the underlying feature extractor, and at the same time designs independent task branches for different bridge types to achieve knowledge transfer across bridge types. The meta-learning module further optimizes the model parameter initialization so that it can quickly adapt to a small amount of measured data for new bridges. For example, in the inner loop, the task branch parameters are fine-tuned for a specific bridge, and the shared feature extractor is optimized in the outer loop to enhance the model's generalization ability for changes in cross-bridge parameters. This method breaks through the limitations of the traditional "one bridge, one model" approach. It only requires fine-tuning the pre-trained base model based on a small amount of measured data, and can be quickly deployed to regional bridge clusters, greatly improving modeling efficiency.

[0134] In this embodiment, in step S400, the construction of the measured data set includes the following process:

[0135] S401: Vehicle weight data is collected through the dynamic weighing (WIM) system installed in the area, highway toll booths, and overload detection stations. If the above methods are not feasible, the vehicle weight data can be obtained by using a weight truck.

[0136] S402: Track the status of vehicles of known weight on different bridges through bridge traffic cameras, thereby obtaining the vehicle's location, speed and license plate information and uploading them to the local server database.

[0137] S403: Use the identified bridge crossing time and location information of the vehicle to perform matching and improve the vehicle information in the database, thereby providing a data basis for subsequent training of the vehicle weight prediction model.

[0138] It should be known that for matching the time and position of vehicles crossing bridges, a spatiotemporal hash table can be introduced to encode the time and coordinate position of vehicles crossing bridges into spatiotemporal feature vectors, and the vehicle trajectories of different bridges can be matched through cosine similarity, thereby solving the spatiotemporal alignment error across bridges.

[0139] S404: The vehicle type identification is used to trigger the collection of structural response data segments, and causal constraints are introduced to ensure temporal correlation, thereby obtaining a measured data set that matches the vehicle weight with the bridge structural response.

[0140] It should be noted that there are many ways to impose causal constraints. For example, based on the vehicle passing time, a sliding window of ±3 seconds is set, and only the structural response data within the window is matched, thereby avoiding mismatching caused by the passing of other vehicles.

[0141] In this embodiment, the progressive fine-tuning of the cascade model in step S400 primarily includes the following steps: determining a base model for the regional pre-trained model space, fine-tuning the seed bridge-specific model, and fine-tuning the bridge cluster-specific weighing model. For ease of understanding, each step is described in detail below.

[0142] Specifically, such as Figure 2 As shown, for the determination of the base model of the regional pre-training model space, in step S300, a pre-training base model A = {α1, α2, ..., α N These base models correspond to different bridge types (such as simply supported beams, continuous beams, etc.) and are trained based on domain-aligned simulation datasets. On this basis, the pre-trained model space dimension N and the corresponding base models for the regional bridge cluster are further determined. The specific steps are as follows: Based on the distribution of bridge types in the region, an appropriate base model combination A is selected to describe the global weighted model space V for the region. N ; For each bridge subspace, load its corresponding pre-trained base model α k , as the basis for subsequent fine-tuning; the shared feature extractor and task-specific layer parameters of the base model will be used as the initial weights.

[0143] Specifically, such as Figure 2As shown in Figure 2, the process of fine-tuning the seed bridge-specific model is as follows: using the bridge with WIM data as the seed node, accurately correlating the vehicle weight with the structural response data according to the spatiotemporal matching method in step S200 to ensure time synchronization; standardizing the input data, including unifying the time series length, zero padding, and channel masking; loading the pre-trained base model α of the corresponding bridge type. k ; Use the seed bridge dataset to fine-tune the model and optimize the support set loss function L support (θ k ) and update the task-specific parameters θ k ´;Evaluate the performance of the fine-tuned model on the query set. If the performance meets expectations, save the fine-tuned model as a dedicated weighing model for the seed bridge.

[0144] Specifically, the fine-tuning of the bridge cluster-specific weighing model based on the seed bridge-specific weighing model mainly includes the high-confidence pseudo-label data generation process and the model cascade and progressive fine-tuning process.

[0145] The high-confidence pseudo-label data generation process includes the following steps:

[0146] S410: Directly supervise the seed bridge based on the collected measured data set, build a special weighing model for the seed bridge and continuously output vehicle weight prediction results.

[0147] S420: Using license plate recognition technology and vehicle spatiotemporal parameter information, the status of the same vehicle on different bridges is tracked to form a vehicle trajectory chain across the bridges.

[0148] S430: If the prediction error for a particular bridge is less than a set threshold, the prediction result is marked as a high-confidence pseudo-label and used as fine-tuning data for other bridge-specific weighing models. The specific value of the threshold can be determined by those skilled in the art based on their actual needs. For example, the threshold can be set to 5%.

[0149] The model cascading and progressive fine-tuning process includes the following steps:

[0150] S440: All bridges in the area are considered as distributed dynamic weighing nodes, and the vehicle weight prediction results output by the deployed bridge model are transmitted to other bridges through the model cascade mechanism.

[0151] S450: For non-seed bridges, a small amount of high-confidence pseudo-labeled data is used in the initial stage to perform preliminary fine-tuning on the pre-trained base model of the bridge. In the later stages, as the high-confidence pseudo-labeled data accumulates, the adjustment range of the shallow parameters is increased to make the model more suitable for the characteristics of the target bridge.

[0152] S460: Re-evaluate the model’s performance on the query set after each fine-tuning session and adjust the model architecture or hyperparameters as needed.

[0153] S470: Once a bridge-specific weighing model has been fully fine-tuned, it is added to the model cascade network to provide more high-confidence pseudo-labeled data for other bridges.

[0154] It is understandable that in order to solve the problem of missing data on bridges without WIM systems, this application proposes a model cascade mechanism based on license plate tracking. The deployed seed bridge model outputs high-confidence pseudo-label data, and associates the vehicle trajectories across the bridge through spatiotemporal matching to form a labeled fine-tuning dataset. As the pseudo-label data accumulates, the model gradually adjusts the shallow parameters to fit the characteristics of the target bridge, and continuously optimizes the model performance of the entire cluster through a cascade feedback mechanism. For example, pseudo-labels are used to fine-tune high-level parameters in the early stage, and more data is combined to optimize the underlying features in the later stage, forming a positive feedback loop of "the more data is used, the more accurate the model is adjusted." This strategy significantly reduces the dependence on the physical WIM system, and is particularly suitable for low-cost deployment of existing bridges in the region.

[0155] The above describes the basic principles, main features, and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-described embodiments. The above-described embodiments and the specification merely illustrate the principles of the present application. Various changes and improvements may be made to the present application without departing from the spirit and scope of the present application. These changes and improvements fall within the scope of the present application for which protection is sought. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A data-driven modeling method for regional bridge dynamic weighing, characterized by: The steps include: S100: Construct a bridge structure response simulation dataset based on vehicle weight; S200: Establish a lightweight monitoring system for regional bridge clusters, conduct adversarial training on measured data of bridge structural responses under actual traffic loads and simulated datasets, and then construct a domain-aligned simulation dataset; S300: The global weighing model space of all bridges is divided into multiple independent subspaces according to bridge type. A multi-task learning strategy in deep learning is used to jointly construct pre-trained base models including the corresponding independent subspaces. S400: Determine a pre-trained base model combination for bridges in the region and construct a measured data set, and perform cascaded progressive fine-tuning from a seed bridge-specific weighing model to a bridge cluster-specific weighing model based on the obtained measured data set; Among them, the bridge with dynamic weighing system is used as the seed bridge; In step S300 , the pre-trained base model includes a shared feature layer, a task-specific adaptation layer, and a meta-learning module; Shared feature layers use shared neural networks Extracting universal features of dynamic deflection signals ; The task-specific adaptation layer is used for each type of bridge design task branch H k , combined with the shared common feature z and bridge parameter s k Predicted vehicle weight y k =H k (z,s k θ k ); The meta-learning module introduces a meta-learning mechanism in the task-specific adaptation layer to optimize the initial parameters and ; in, represents the shared feature layer parameters, Indicates setting invalid channels to zero. Indicates task branch parameters.

2. The data-driven modeling method for regional bridge dynamic weighing according to claim 1, characterized in that: In step S100, the construction of the bridge structure response simulation dataset includes the following process: S110: Select bridges with different structural parameters, including bridge type, material parameters, cross-sectional properties, and span parameters, and establish a refined finite element model; S120: Statistically analyzing the traffic flow information collected by the actual system to obtain a vehicle load spectrum including gross weight, axle weight, and wheelbase information, and establishing a vehicle dynamics analysis model that conforms to regional characteristics; S130: Constructing a multi-operating condition parameter set, including vehicle speed range, lane position, and road surface roughness, to form an operating condition parameter space covering typical vehicle-bridge coupling; S140: Generate bridge-vehicle-operating condition combinations through Latin hypercube sampling, perform vehicle-bridge coupling finite element simulations under typical operating conditions, and obtain a simulation data set of the bridge structure response.

3. The data-driven modeling method for regional bridge dynamic weighing according to claim 1, characterized in that: In step S200, the construction of the domain alignment simulation dataset includes the following process: S210: Perform adversarial training on the simulated dataset S and the pseudo-real dataset R' generated in the forward direction, and perform adversarial training on the real measured dataset R and the pseudo-simulated dataset S' generated in the forward direction to obtain the adversarial loss function L GAN ; S220: Reconstruct the simulation dataset S and the real measured dataset R through the cycle consistency constraint to obtain the cycle consistency loss function L cycle ; S230: Perform weighted summation of the adversarial loss function and the cycle consistency loss function to obtain the total loss function L total And minimize; the total loss function ;in, represents the weight of cycle consistency loss; S240: Keep the label of the original simulation dataset S, and replace the original simulation dataset S with the pseudo real dataset R' to obtain a domain-aligned simulation dataset.

4. The data-driven modeling method for regional bridge dynamic weighing according to claim 3, characterized in that: Step S210 includes the following process: S211: Input the simulation data set S and the real measured data set R into the generator respectively and , correspondingly generate pseudo real data set R' and pseudo simulation data set S'; S212: Input the real measured data set R and the pseudo real data set R' to the discriminator D at the same time R , input the simulated dataset S and pseudo-simulated dataset S' into the discriminator D at the same time S ; S213: Setting the Discriminator D R The goal is to judge the real measured data set R as true and the pseudo-real data set R' as false; the discriminator D S The goal is to judge the simulation data set S as true and the simulation data set S' as false.

5. The data-driven modeling method for regional bridge dynamic weighing according to claim 3, characterized in that: Step S220 includes the following process: S221: Input the simulation data set S and the real measured data set R into the generator respectively and , correspondingly generate pseudo real data set R' and pseudo simulation data set S'; S222: Input the pseudo real dataset R' and pseudo simulation dataset S' to the generator respectively and , correspondingly generating the reconstructed simulation data set S" and the reconstructed measured data set R"; S223: Reconstruct the simulation data set S" and the simulation data set S through the cycle consistency constraint requirement for approximation, and reconstruct the measured data set R" and the real measured data set R for approximation.

6. The data-driven modeling method for regional bridge dynamic weighing according to any one of claims 1 to 5, characterized in that: In step S300, data preprocessing is required before building the pre-trained base model, which specifically includes the following steps: S310: Input raw data, including dynamic deflection time series, bridge structure parameters and vehicle load labels; S320: Set the target time series length T, and perform linear interpolation or compression on the original time series to make its length uniform to T; S330: Set the maximum number of channels C max , fill the insufficient channels with zero and generate a channel mask matrix ; Among them, M[C]=1 means that the Cth channel is valid, and M[C]=0 means that the Cth channel is invalid; S340: Each bridge type and parameter combination is considered as a task, which is further divided into a support set and a query set. The support set accounts for 80% of the total tasks, and the query set accounts for 20% of the total tasks.

7. The data-driven modeling method for regional bridge dynamic weighing according to claim 6, characterized in that: After the pre-trained base model is built, it needs to be trained. The specific training process is as follows: For each task branch H k Optimize task-specific parameters θ using support set data k ´; The support set data is passed through the support set loss function L support (θ s ,θ k ) to optimize the parameters. The specific expression is as follows: ； ； Use the updated task-specific parameters θ k Evaluate model performance on query set data and update parameters and , the query set data passes the query set loss function L query (θ s ,θ k ´) to perform performance evaluation, the specific expression is as follows: ； ； ； ； Among them, α represents the inner loop learning rate, β represents the outer loop learning rate, represents the summary of the loss function of the total query set, k represents the total number of bridge types, x represents the dynamic deflection time series, y represents the measured value of vehicle weight, N support and N query Denotes the number of samples corresponding to the support set and query set, respectively, D support and D query Represents support set data and query set data respectively.

8. The data-driven modeling method for regional bridge dynamic weighing according to claim 1, characterized in that: The regional bridge cluster lightweight monitoring system includes sensors or a DIC machine vision system for collecting structural response data, and traffic cameras for monitoring traffic on the bridge deck. In step S400, the construction of the measured data set includes the following process: S401: Vehicle weight data is collected through dynamic weighing systems, highway toll booths, and overload detection stations installed in the area; S402: Tracking the status of vehicles of known weight on different bridges through bridge traffic cameras, thereby obtaining the vehicle's location, speed, and license plate information and uploading it to a local server database; S403: Matching the identified vehicle crossing time and location information to improve the vehicle information in the database; S404: The vehicle type identification is used to trigger the collection of structural response data segments, and causal constraints are introduced to ensure temporal correlation, thereby obtaining a measured data set that matches the vehicle weight with the bridge structural response.

9. The data-driven modeling method for regional bridge dynamic weighing according to claim 8, characterized in that: In step S400, the process of fine-tuning the bridge cluster-specific weighing model based on the seed bridge-specific weighing model is as follows: S410: Directly supervise the seed bridge based on the collected measured data set, build a special weighing model for the seed bridge and continuously output vehicle weight prediction results; S420: Using license plate recognition technology and vehicle spatiotemporal parameter information, the status of the same vehicle on different bridges is tracked to form a vehicle trajectory chain across the bridges; S430: If the prediction error of a bridge is less than a set threshold, the prediction result is marked as a high-confidence pseudo label and used as data for fine-tuning the weighing model for other bridges; S440: All bridges in the area are considered as distributed dynamic weighing nodes. The vehicle weight prediction results output by the deployed bridge model are transmitted to other bridges through the model cascade mechanism. S450: For non-seed bridges, initially fine-tune the pre-trained base model of the bridge using a small amount of high-confidence pseudo-labeled data; in the later stages, as high-confidence pseudo-labeled data accumulates, increase the adjustment range of shallow parameters; S460: Re-evaluate the model’s performance on the query set after each fine-tuning session and adjust the model architecture or hyperparameters as needed; S470: Once a bridge-specific weighing model has been fully fine-tuned, it is added to the model cascade network to provide more high-confidence pseudo-labeled data for other bridges.

Images