A bridge crack monitoring device

Abstract

This invention discloses a bridge crack monitoring device, belonging to the field of bridge monitoring technology. The invention embeds a fracture mechanics forward model into a physical information neural network in the form of a differentiable network layer with hard constraints, achieving high-precision decoupling of the strain fields of multiple cracks. It utilizes the prior information about crack locations obtained through fiber optic strain decoupling to generate an attention heatmap, actively guiding the image segmentation network to focus on strain anomaly regions, forming a physical-data pre-fusion paradigm. Furthermore, it establishes a temperature-load coupling response separation model, decomposing the crack width time-series data to drive a physical mechanism cumulative damage model based on the Paris formula, achieving quantitative separation of environmental effects and damage effects, and prediction of crack remaining life. This invention ensures physical consistency throughout the entire chain from raw monitoring data to risk assessment results, effectively improving the decoupling accuracy of multiple cracks, the robustness of crack identification in complex environments, and the scientific nature of risk assessment. It can be widely applied to the full life-cycle monitoring of cracks in various bridge structures.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring technology, and more specifically to a bridge crack monitoring device. Background Technology

[0002] During long-term operation, bridge structures inevitably develop various cracks due to factors such as vehicle loads, temperature changes, material aging, and natural disasters. The initiation and propagation of cracks are important indicators of structural damage evolution, and accurate, real-time, and comprehensive monitoring of them is crucial for ensuring bridge operational safety and extending structural lifespan. With the continuous expansion of transportation infrastructure, a large number of in-service bridges are entering a period of high incidence of defects, making the need for efficient and reliable crack monitoring technologies increasingly urgent.

[0003] Currently, bridge crack monitoring technologies can be mainly classified into the following categories:

[0004] (I) Machine Vision-Based Crack Detection Methods. This method uses industrial cameras to acquire images of the bridge surface and identifies cracks through image processing algorithms (such as edge detection and threshold segmentation) or deep learning algorithms (such as convolutional neural networks and semantic segmentation networks). Such methods are highly intuitive and visually appealing. However, they are greatly affected by environmental factors such as lighting, shadows, water stains, and dirt, resulting in low detection accuracy for micro-cracks less than 0.3 mm wide and difficulty in obtaining crack depth information. More importantly, existing deep learning methods only extract features from the image data itself, without incorporating any prior physical information related to the structural stress state. When there are crack-like interferences such as dirt, construction joints, and formwork joints on the bridge surface, the network cannot distinguish the essential difference in stress between these and real cracks, leading to a significantly higher false detection rate in complex environments.

[0005] (II) Strain Monitoring Method Based on Distributed Fiber Optic Sensing. Distributed fiber optic sensing technology uses fiber optic cables deployed along the bridge structure to obtain continuous strain distribution along the fiber axis using Brillouin scattering or Rayleigh scattering principles. The location and width of cracks are inferred by identifying strain anomaly regions. This method has advantages such as full distribution, long distance, and resistance to electromagnetic interference, with spatial resolution reaching millimeter or even sub-millimeter levels, capable of capturing minute local strain anomalies. However, when multiple adjacent cracks appear in a bridge, the strain peaks generated by each crack overlap spatially, and traditional peak detection methods cannot effectively distinguish the independent contribution of a single crack. In existing technologies, some schemes use independent component analysis for blind source separation to decouple crack signals from background deformation signals. However, this method is purely statistical and does not introduce any mechanical or physical mechanisms. Its separation results lack physical consistency constraints, and the inverted crack parameters may seriously deviate from the actual mechanical state. In recent years, researchers have attempted to introduce physical constraints into neural networks, interpreting multi-crack strain distributions by adding a physical constraint penalty term to the loss function, or using physical information neural networks for stress-strain field reconstruction. However, the above methods all introduce physical constraints in the form of penalty terms in the loss function, which belongs to the "soft constraint" paradigm. Physical constraints only affect network parameters in a penalized manner during the training phase. The forward propagation of the network itself does not guarantee that the output satisfies the physical equation. The guarantee of physical consistency is limited. When the strain peaks of multiple cracks overlap severely, inversion results that violate the basic principles of mechanics may still occur.

[0006] (III) Risk Assessment Methods Based on Static Crack Parameters. Existing crack risk assessment methods mainly rely on static crack width, length, and other geometric parameters obtained through manual detection or sensors as judgment indicators, and classify warnings according to thresholds specified in the standards. However, crack width is significantly affected by temperature changes. When the temperature rises, the concrete expands and the crack width decreases; when the temperature falls, the concrete shrinks and the crack width increases. This variation can be several times greater. Existing research shows that the periodic changes in crack width caused by temperature may far exceed the irreversible propagation caused by structural damage. Existing methods do not effectively separate the reversible thermal expansion and contraction effect caused by temperature from the irreversible cumulative effect caused by structural damage, and directly assess based on the original width data, resulting in seriously distorted risk assessment results and frequent false alarms and false negatives. Some improved schemes attempt to integrate response characteristics such as the dynamic propagation rate after vehicles cross the bridge for risk assessment, but they focus on the transient dynamic response after vehicles cross the bridge and do not establish a long-term temperature-load coupling monitoring mechanism, thus still failing to solve the long-term separation problem between temperature effects and damage effects. Another approach uses a purely data-driven probabilistic model to capture crack propagation rates in real time. However, due to the lack of a driving force from the physical mechanisms of fracture mechanics, its prediction of crack development trends lacks mechanical basis, and the extrapolation accuracy is difficult to guarantee.

[0007] In summary, the existing technologies have at least the following shortcomings: (1) In the distributed optical fiber strain interpretation method, the purely statistical method lacks physical mechanism constraints, the soft constraint method has limited physical consistency guarantees, and the decoupling accuracy is low in multi-crack scenarios; (2) The crack image segmentation method does not introduce the physical prior information of the structural stress state, resulting in a high false detection rate in complex environments; (3) The crack risk assessment method does not effectively separate the temperature reversible effect from the damage irreversible effect, resulting in inaccurate assessment results. The above three problems involve the three key links of crack perception, identification, and assessment. The deficiency of any link will lead to a significant reduction in the overall monitoring effect, making it difficult to meet the actual needs of accurate monitoring of cracks throughout the entire life cycle of bridge structures. Summary of the Invention

[0008] In view of this, the present invention provides a bridge crack monitoring device, aiming to solve the following technical problems: In existing distributed optical fiber strain interpretation methods, purely statistical methods (such as independent component analysis) lack physical mechanism constraints, and soft constraint methods (such as PC-FiberNet, stress or strain field reconstruction methods based on physical perception neural networks) have limited physical consistency guarantees and low decoupling accuracy in multi-crack scenarios; existing crack image segmentation methods do not introduce physical prior information about the structural stress state, resulting in a high false detection rate in complex environments; existing crack risk assessment methods do not effectively separate temperature reversible effects from damage irreversible effects, leading to inaccurate assessment results.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention discloses a bridge crack monitoring device, comprising:

[0011] The multi-source data acquisition module includes distributed fiber optic sensors deployed along the bridge structure axis, temperature sensors deployed at key sections of the bridge, and a mobile inspection platform equipped with a high-resolution industrial camera.

[0012] The physical constraint strain decoupling module is connected to the distributed optical fiber sensor signal and is used to receive the measured strain distribution data collected by the distributed optical fiber sensor, run the multi-crack strain field hard constraint decoupling method based on physical information neural network, and output the location parameters and width parameters of the multi-crack.

[0013] The physical prior image segmentation module is connected to the physical constraint strain decoupling module and the high-resolution industrial camera signal respectively. It is used to receive the crack location parameters output by the physical constraint strain decoupling module and the bridge surface image acquired by the high-resolution industrial camera. It generates a strain prior attention heatmap based on the crack location parameters, and uses the strain prior attention heatmap as a spatial attention weight to guide the image segmentation network to segment the bridge surface image and output the crack edge contour.

[0014] The physical mechanism risk assessment module is connected to the physical constraint strain decoupling module and the temperature sensor signal, respectively, and is used to receive vehicle load response data. It is used to establish a temperature-load coupling response separation model based on the crack width time series data output by the physical constraint strain decoupling module, the temperature time series data collected by the temperature sensor, and the vehicle load response data, to separate the cumulative amount of irreversible damage, calculate the cumulative damage risk index, and output the risk assessment result.

[0015] The digital twin visualization module is connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module, respectively, to construct a three-dimensional digital twin model of the bridge. The output data of each module is mapped to the digital twin model in real time for three-dimensional visualization.

[0016] Preferably, the physical constraint strain decoupling module includes:

[0017] The encoder unit employs a multi-layer fully connected network, using the measured strain distribution data as input to extract high-dimensional features.

[0018] A parameter regression unit, connected to the encoder unit, is used to map the high-dimensional features into predicted values ​​of crack parameters.

[0019] The physical forward modeling unit is a differentiable network layer, in which a multi-crack superimposed physical forward modeling model based on linear elastic fracture mechanics theory is embedded. The physical forward modeling unit is connected to the parameter regression unit and is used to calculate the theoretical strain distribution based on the predicted values ​​of the crack parameters.

[0020] The training unit, connected to the physical forward modeling unit, is used to perform end-to-end training by using the data fitting term between the theoretical strain distribution and the measured strain distribution as the loss function and backpropagating the error to the parameter regression unit and the encoder unit through the chain rule.

[0021] Preferably, the multi-crack superposition physical forward model linearly superimposes the theoretical strain distributions generated by each individual crack and adds a noise term to form the total theoretical strain distribution; wherein the theoretical strain distribution of each individual crack is determined based on the crack center position, crack width, strain amplitude coefficient related to the elastic modulus of the material, and spatial attenuation constant related to the shear hysteresis effect of the fiber bonding layer.

[0022] Specifically, the process of constructing a physical forward model for multiple crack superposition:

[0023] (1) Constructing a crack-strain physical forward model. Based on the theory of linear elastic fracture mechanics, the theoretical strain distribution function generated by a single crack along the optical fiber axis is established:

[0024]

[0025] in, For the first The center of the crack Location, The width of the crack. This is the strain amplitude coefficient related to the material's elastic modulus. This is the spatial attenuation constant related to the shear hysteresis effect of the fiber bonding layer.

[0026] (2) Construct a physical forward model of the superimposed strain field. The strain contributions from multiple cracks are linearly superimposed, and a noise term is added to form the overall theoretical strain distribution:

[0027]

[0028] in The set of crack parameters to be inverted. This represents the number of cracks.

[0029] (3) Constructing a hard-constraint architecture for a physical information neural network. Unlike the "soft constraint" paradigm in the prior art, which introduces physical constraints in the form of a loss function penalty term, this invention directly embeds the above-mentioned physical forward model into the neural network architecture in the form of differentiable network layers.

[0030] Specifically, the physical information neural network constructed in this invention includes:

[0031] Encoder: based on measured strain distribution As input, a multi-layer fully connected network is used to extract high-dimensional features;

[0032] Parametric regression layer: Connected to the encoder, it maps high-dimensional features to crack parameters. The predicted value;

[0033] Physical forward modeling layer: This is a differentiable network layer in which the above physical forward modeling model is embedded. It is connected to the parameter regression layer and used to calculate the theoretical strain distribution based on the predicted values ​​of crack parameters;

[0034] Hard constraint mechanism: During the forward propagation of the network, there is a fixed physical mapping relationship between the output of the physical forward modeling layer and the output of the encoder-parameter regression layer. This mapping relationship is defined by the fracture mechanics equation and cannot be changed through learning.

[0035] The network's training loss function contains only the data fitting term:

[0036]

[0037] in This represents the number of strain sampling points. Since the physical forward modeling layer is a differentiable layer, the error can be backpropagated to the parameter regression layer and the encoder using the chain rule, achieving end-to-end training.

[0038] (4) Automatic determination of the number of cracks. The Bayesian Information Criterion (BIC) is used to determine the number of cracks. Automatic selection:

[0039]

[0040] in The number of cracks is The maximum likelihood value at time , The number of model parameters, This represents the number of strain sampling points. Select the smallest BIC value. The optimal number of cracks.

[0041] Preferably, the physical prior image segmentation module includes:

[0042] The coordinate mapping unit is used to map the crack location parameters output by the physical constraint strain decoupling module from the one-dimensional coordinates of the optical fiber to the two-dimensional coordinate system of the image.

[0043] A heat map generation unit, connected to the coordinate mapping unit, is used to generate a strain prior attention heat map based on the mapped crack location, wherein the value of each pixel in the heat map reflects the probability that a crack exists at that location.

[0044] The segmentation network unit includes a backbone encoder and a prior attention injection module. The backbone encoder is used to extract multi-scale features of the bridge surface image, and the prior attention injection module is used to inject the strain prior attention heatmap as spatial attention weights into the intermediate feature layer of the backbone encoder.

[0045] The inference control unit, connected to the segmentation network unit, is used to control the segmentation network unit to use high-resolution fine inference for high-attention regions in the heatmap and low-resolution fast inference for low-attention regions, and output the crack segmentation result.

[0046] Preferably, the heatmap generation unit generates attention weights by spreading them outwards to the surrounding area according to a Gaussian distribution, with each crack location as the center. The attention weights of each crack are then superimposed to form the strain prior attention heatmap.

[0047] Specifically, the strain prior attention heatmap is generated using the crack location information obtained from decoupling, which actively guides the image segmentation network to focus on the strain anomaly region.

[0048] (1) Generation of strain prior attention heatmap. The decoupled crack location is then... Mapped to the image coordinate system, a strain prior attention heatmap is generated.

[0049]

[0050] in For the first The projected coordinates of the crack in the image. This parameter controls the radius of the attention region. In the heat map, a larger value indicates a higher probability that the fiber optic strain data indicates the presence of a crack in that area.

[0051] (2) Strain-prior-guided segmentation network.

[0052] A crack image segmentation network is constructed, comprising a backbone encoder and a prior attention injection module. The backbone encoder employs a lightweight convolutional neural network to extract multi-scale features from the image; the prior attention injection module injects strain prior attention heatmaps. Spatial attention weights are injected into the intermediate feature layer of the backbone encoder; for high attention regions in the heatmap, the network uses high-resolution fine-grained inference; for low attention regions, the network uses low-resolution fast inference, thereby achieving optimized allocation of computing resources.

[0053] Preferably, the physical mechanism risk assessment module includes:

[0054] The time-series data synchronization unit is used to synchronously collect crack width time-series data, bridge temperature time-series data, and vehicle load response time-series data.

[0055] The response separation unit, connected to the time-series data synchronization unit, is used to establish a temperature-load coupled response separation model, which decomposes the crack width time-series data into the initial crack width, the reversible crack width change component caused by temperature, the reversible elastic deformation contribution component caused by vehicle load, and the irreversible crack propagation component caused by structural damage accumulation.

[0056] The cumulative damage calculation unit, connected to the response separation unit, is used to establish a cumulative damage model based on the fatigue crack propagation law, identify material parameters online through continuous monitoring data, and calculate the cumulative amount of irreversible damage.

[0057] The risk index calculation unit is connected to the cumulative damage calculation unit and is used to calculate the cumulative damage risk index based on the amount of irreversible damage and its expansion rate. When the risk index reaches a preset threshold, an early warning signal is output.

[0058] Preferably, the response separation unit calculates the reversible crack width change component caused by temperature using a thermoelastic constitutive model, and calculates the reversible elastic deformation contribution component caused by vehicle load using a bridge influence line model. After deducting the above two reversible components from the crack width time series data, the remaining part is taken as the irreversible crack propagation amount caused by the accumulation of structural damage.

[0059] Preferably, the risk index calculation unit performs dimensionless processing on the cumulative amount of irreversible damage and its propagation rate, respectively, and obtains the cumulative damage risk index by weighted summation of the two dimensionless indicators based on the critical values ​​of crack width and propagation rate.

[0060] Specifically, the physical mechanism risk assessment process is as follows:

[0061] (1) Synchronous acquisition of multi-source time series data. Three types of time series data were acquired synchronously: crack width time series data. (Output from the first layer), bridge temperature time series data (Acquired through deployed temperature sensors) Vehicle load response time-series data (Obtained through dynamic weighing systems or bridge strain response identification).

[0062] (2) Temperature-load coupled response separation model. The crack width... The model is a superposition of four contributions:

[0063]

[0064] in: The initial width of the crack (constant term); The temperature-induced crack width variation (reversible thermal expansion and contraction effect) is described using a thermoelastic constitutive model. The contribution of elastic deformation (reversible elastic response) caused by vehicle load is described using a bridge influence line model; The amount of irreversible crack propagation caused by the accumulation of structural damage (the core quantity that needs to be assessed).

[0065] (3) Fatigue cumulative damage model based on Paris formula. For the separated irreversible damage components... A cumulative damage model was established based on Paris's fatigue crack propagation law:

[0066]

[0067] in The number of load cycles, The stress intensity factor amplitude (derived from vehicle load) (Calculated) and Material parameters. Material parameters are identified online through continuous monitoring data. and This enables real-time prediction of the remaining life of cracks.

[0068] (4) Cumulative Damage Risk Index. A comprehensive index considering irreversible damage is constructed. and its expansion rate Risk index:

[0069]

[0070] in This is the critical value for crack width. This is the critical value for the expansion rate. , For weighting coefficients. When An alert is triggered at any time.

[0071] Preferably, the digital twin visualization module includes:

[0072] The 3D model building unit is used to build a 3D digital twin model of the bridge structure based on the bridge design parameters.

[0073] The data mapping unit is connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module, respectively, and is used to map the output data of each module to the corresponding spatial position of the three-dimensional digital twin model.

[0074] The early warning display unit, connected to the data mapping unit, is used to mark cracks of different risk levels on the three-dimensional digital twin model with different colors and / or icons, and to display crack development trend prediction information.

[0075] Preferably, it also includes a data communication module, which is signal-connected to the multi-source data acquisition module, the physical constraint strain decoupling module, the physical prior image segmentation module, the physical mechanism risk assessment module, and the digital twin visualization module, respectively, to realize data interaction between the modules and data transmission with the remote monitoring center.

[0076] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a bridge crack monitoring device. The present invention adopts a unified concept of physical information-driven approach, and by embedding a fracture mechanics forward model into a neural network in the form of a differentiable network layer with hard constraints, the decoupling results of the multi-crack strain field naturally satisfy the basic principles of mechanics in each forward propagation. This fundamentally solves the problems of the lack of physical constraints in existing pure statistical methods and the insufficient physical consistency guarantee of soft constraint methods, significantly improving the decoupling accuracy and physical reliability in multi-crack overlapping scenarios. Based on this, it uses the crack location prior obtained from distributed fiber strain decoupling to generate an attention heatmap, actively guiding the image segmentation network to focus on strain anomaly regions. It replaces the post-fusion method of image feature extraction and subsequent fusion with the pre-fusion of physical priors and image data, enabling the segmentation process to have physical prediction capabilities, significantly reducing the false detection rate caused by interference factors such as illumination, shadows, and surface stains in complex environments. Furthermore, by establishing a temperature- The load-coupled response separation model decomposes crack width time-series data into reversible temperature components, elastic load components, and irreversible damage components, achieving quantitative separation of environmental and damage effects. This overcomes the shortcomings of existing crack risk assessments caused by temperature interference, which leads to assessment distortion. Furthermore, the separated irreversible damage component drives a physical mechanism cumulative damage model based on the Paris formula to predict remaining life, giving the assessment and prediction results a clear mechanical basis and significantly improving the scientific rigor and reliability of risk assessment. A complete physical information driving chain is formed through a progressive logic of "strain perception providing physical constraints, physical constraints guiding visual recognition, and multi-dimensional data driving mechanism assessment." Distributed fiber optic sensing capabilities and high-resolution visual imaging capabilities work in deep synergy to ensure physical consistency throughout the entire chain from raw monitoring data to the final risk assessment results. This effectively solves the technical problems of isolated links and inconsistent physical meanings in existing technologies, providing a complete technical support solution for accurate monitoring of bridge cracks throughout their entire life cycle. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0078] Figure 1 The structural block diagram provided for this invention;

[0079] Figure 2 The structural block diagram of the physical constraint strain decoupling module provided by the present invention;

[0080] Figure 3The block diagram of the physical prior image segmentation module provided by the present invention;

[0081] Figure 4 The structural block diagram of the physical mechanism risk assessment module provided by this invention;

[0082] Figure 5 This is the interface corresponding to the digital twin visualization module provided by the present invention. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] Example 1: System Overall Architecture

[0085] like Figure 1 As shown, the present invention includes a multi-source data acquisition module, a physical constraint strain decoupling module, a physical prior image segmentation module, a physical mechanism risk assessment module, and a digital twin visualization module.

[0086] The multi-source data acquisition module includes distributed fiber optic sensors deployed along the bridge structure's axial direction, temperature sensors positioned at key bridge sections, and a mobile inspection platform equipped with a high-resolution industrial camera. The distributed fiber optic sensors utilize optical frequency domain reflectance (OFDR) technology, with a spatial resolution no greater than 0.65 mm, and are continuously laid along the axial direction of the bridge's main beams and piers. Temperature sensors are deployed at key bridge sections to collect time-series temperature data of the bridge structure. The mobile inspection platform is a drone or a wall-climbing robot. The high-resolution industrial camera is mounted on the mobile inspection platform, which is equipped with a positioning module to register the acquired image coordinates with the bridge's spatial coordinates.

[0087] The physical constraint strain decoupling module is connected to the distributed optical fiber sensor signal, receives the measured strain distribution data collected by the distributed optical fiber sensor, runs the hard constraint decoupling method for multi-crack strain field based on physical information neural network, and outputs the location parameters and width parameters of the multi-crack.

[0088] The physical prior image segmentation module is connected to the physical constraint strain decoupling module and the high-resolution industrial camera signal respectively. It receives the crack location parameters output by the physical constraint strain decoupling module and the bridge surface image acquired by the high-resolution industrial camera. Based on the crack location parameters, it generates a strain prior attention heatmap. The strain prior attention heatmap is used as a spatial attention weight to guide the image segmentation network to segment the bridge surface image and output the crack edge contour.

[0089] The physical mechanism risk assessment module is connected to the physical constraint strain decoupling module and the temperature sensor signal, respectively, and is used to receive vehicle load response data. Based on the crack width time-series data output by the physical constraint strain decoupling module, the temperature time-series data collected by the temperature sensor, and the vehicle load response data, a temperature-load coupling response separation model is established to separate the cumulative irreversible damage, calculate the cumulative damage risk index, and output the risk assessment results. The vehicle load response data is obtained through a dynamic weighing system deployed on the bridge pavement or by identifying the bridge strain response through strain sensors deployed on the bridge structure.

[0090] The digital twin visualization module is connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module to construct a three-dimensional digital twin model of the bridge. The output data of each module is mapped to the digital twin model in real time for three-dimensional visualization.

[0091] It should be noted that the "hard constraint" described in this invention is in contrast to the "soft constraint" paradigm in existing technologies, which introduces physical equations as penalty terms in the loss function. In existing soft constraint methods, the output of the neural network... It does not inherently satisfy physical equations; physical constraints only indirectly affect network parameters through gradient signals during backpropagation, and the network forward propagation process lacks physical interpretability.

[0092] In this invention, the physical forward modeling unit is embedded in the network in the form of a fixed architecture layer. Its input-output mapping relationship is predefined by the fracture mechanics equations and cannot be changed through learning. Even if the measured data contains noise, the crack parameters output by the network remain consistent. In each forward propagation round, the theoretical strain distribution is calculated according to the physical equations, ensuring that the prediction results naturally satisfy the principle of mechanical superposition at the structural level. For model errors and measurement noise, this invention attributes them to the physical forward model. The processing of these items does not affect the deterministic nature of the physical mapping relationship.

[0093] Example 2: Specific Implementation of the Physical Constraint Strain Decoupling Module

[0094] like Figure 2 As shown, the physical constraint strain decoupling module includes an encoder unit, a parameter regression unit, a physical forward modeling unit, and a training unit.

[0095] The encoder unit employs a multi-layer fully connected network, using measured strain distribution data as input to extract high-dimensional features.

[0096] The parametric regression unit is connected to the encoder unit to map high-dimensional features into predicted values ​​of crack parameters.

[0097] The physical forward modeling unit is a differentiable network layer, which embeds a multi-crack superimposed physical forward model based on the theory of linear elastic fracture mechanics. The physical forward modeling unit is connected to the parameter regression unit, and the theoretical strain distribution is calculated based on the predicted values ​​of the crack parameters.

[0098] The training unit is connected to the physical forward modeling unit. The data fitting term between the theoretical strain distribution and the measured strain distribution is used as the loss function. The error is backpropagated to the parameter regression unit and the encoder unit through the chain rule for end-to-end training.

[0099] The multi-crack superposition physical forward model linearly superimposes the theoretical strain distributions generated by each individual crack and adds a noise term to form the total theoretical strain distribution. The theoretical strain distribution of each individual crack is determined based on the crack center location, crack width, strain amplitude coefficient related to the material's elastic modulus, and spatial attenuation constant related to the shear hysteresis effect of the fiber bonding layer.

[0100] Since the physical forward model of multiple crack superposition is based on the assumptions of uniform linear elastic medium and ideal shear hysteresis, while actual concrete / steel bridges may have material nonlinearity, interface slip, local defects, etc., systematic deviations may occur between the model and the actual structure after long-term operation.

[0101] The physical forward modeling unit also includes a model error compensation layer, which takes the predicted values ​​of the crack parameters as input and uses a low-order polynomial network or Gaussian process regression to output compensated strain. .

[0102] The overall theoretical strain distribution is corrected to:

[0103]

[0104] The complexity of the model error compensation layer is subject to regularization constraints (such as...). Regularization or early stopping strategies are used to ensure that the compensation layer is used only to absorb systematic biases in the physical model, rather than drowning out the dominant mechanical response. During the initial training phase, the output of the compensation layer is initialized to zero.

[0105] In one specific embodiment, the measured strain distribution data is acquired using an OFDR distributed fiber optic sensor, with the number of spatial sampling points being [number missing]. The measured strain distribution is input into the encoder unit, and features are extracted by a multi-layer fully connected network. The parametric regression unit then outputs predicted values ​​of the crack parameters. The physical forward modeling unit calculates the theoretical strain distribution based on the predicted crack parameters. The training unit calculates the mean square error between the theoretical and measured strain distributions as a loss function and updates the network parameters through backpropagation. After sufficient training, the network can accurately decouple the location and width of each crack from the superimposed strain field.

[0106] For cases where the number of cracks is unknown, the Bayesian information criterion is used to automatically determine the number of cracks. Let the number of cracks be 1, 2, 3, ... Train the corresponding network and calculate the BIC value, then select the N with the smallest BIC value as the optimal number of cracks.

[0107] Example 3: Specific Implementation of the Physical Prior Image Segmentation Module

[0108] like Figure 3 As shown, the physical prior image segmentation module includes a coordinate mapping unit, a heatmap generation unit, a segmentation network unit, and an inference control unit.

[0109] The coordinate mapping unit maps the crack location parameters output by the physical constraint strain decoupling module from the one-dimensional coordinates of the optical fiber to the two-dimensional coordinate system of the image.

[0110] The heat map generation unit is connected to the coordinate mapping unit. Based on the mapped crack location, a strain prior attention heat map is generated. The value of each pixel in the heat map reflects the probability that a crack exists at that location.

[0111] The segmentation network unit includes a backbone encoder and a prior attention injection module. The backbone encoder extracts multi-scale features from the bridge surface image, and the prior attention injection module injects the strain prior attention heatmap as spatial attention weights into the intermediate feature layer of the backbone encoder.

[0112] The inference control unit is connected to the segmentation network unit, which controls the segmentation network unit to use high-resolution fine inference for high-attention regions in the heatmap and low-resolution fast inference for low-attention regions, and outputs the crack segmentation results.

[0113] The heatmap generation unit generates attention weights by spreading outwards to the surrounding area from the location of each crack in a Gaussian distribution. The attention weights of each crack are superimposed to form a strain prior attention heatmap.

[0114] Furthermore, the coordinate mapping unit embeds a positioning uncertainty evaluation subunit, which is used to calculate in real time the spatial uncertainty of each crack position mapped to the image coordinate system based on the positioning signal quality of the mobile inspection platform (such as the number of GNSS satellites, RTK fixed solution status, and visual odometry cumulative drift). The heatmap generation unit dynamically adjusts the attention diffusion radius corresponding to each crack based on the spatial uncertainty, specifically as follows:

[0115]

[0116] That is, when the positioning uncertainty increases, the diffusion range of the attention heatmap around the crack expands accordingly, compensating for the mapping error in a "blurring" manner to ensure that the real crack remains within the high attention region. Furthermore, when the positioning uncertainty exceeds a preset threshold, the inference control unit automatically increases the inference resolution in low attention regions to prevent missed detections due to mapping offset.

[0117] In one specific embodiment, the physical constraint strain decoupling module outputs the position parameters of the three cracks, and the coordinate mapping unit maps the positions of the three cracks from the one-dimensional coordinates of the optical fiber to the two-dimensional coordinate system of the image.

[0118] The heatmap generation unit generates an attention heatmap centered on three mapped coordinates and following a Gaussian distribution. The parameter σ, which controls the radius of the attention region, is set to 50 pixels.

[0119] The backbone encoder of the segmentation network unit uses a lightweight convolutional neural network (such as MobileNetV2), and the prior attention injection module injects attention heatmaps into the 3rd and 5th layers of the encoder, respectively. The inference control unit divides the image region into high attention regions (heatmap values ​​greater than a threshold) and low attention regions based on the heatmap values. For high attention regions, fine inference is performed at the original resolution, while for low attention regions, fast inference is performed after downsampling, followed by upsampling for recovery.

[0120] The final output is a pixel-level segmentation result of the crack.

[0121] Example 4: Specific Implementation of the Physical Mechanism Risk Assessment Module

[0122] like Figure 4 As shown, the physical mechanism risk assessment module includes a time-series data synchronization unit, a response separation unit, a cumulative damage calculation unit, and a risk index calculation unit.

[0123] The time-series data synchronization unit is used to synchronously collect time-series data on crack width, bridge temperature, and vehicle load response.

[0124] The response separation unit is connected to the time-series data synchronization unit to establish a temperature-load coupled response separation model. The crack width time-series data is decomposed into the initial crack width, the reversible crack width change component caused by temperature, the reversible elastic deformation contribution component caused by vehicle load, and the irreversible crack propagation component caused by structural damage accumulation.

[0125] The cumulative damage calculation unit is connected to the response separation unit. A cumulative damage model is established based on the fatigue crack propagation law. Material parameters are identified online through continuous monitoring data, and the cumulative amount of irreversible damage is calculated.

[0126] The risk index calculation unit is connected to the cumulative damage calculation unit. It calculates the cumulative damage risk index based on the amount of irreversible damage and its expansion rate. When the risk index reaches a preset threshold, it outputs an early warning signal.

[0127] The response separation unit calculates the reversible crack width variation component caused by temperature using a thermoelastic constitutive model and the reversible elastic deformation contribution component caused by vehicle load using a bridge influence line model. After deducting the above two reversible components from the crack width time series data, the remaining part is taken as the irreversible crack propagation amount caused by the accumulation of structural damage.

[0128] The risk index calculation unit performs dimensionless processing on the cumulative amount of irreversible damage and its propagation rate, respectively. Based on the critical values ​​of crack width and propagation rate, the two dimensionless indicators are weighted and summed to obtain the cumulative damage risk index.

[0129] In one specific embodiment, the time-series data synchronization unit synchronously acquires crack width, temperature, and vehicle load response data at a sampling rate of 1Hz for 30 consecutive days. The response separation unit first establishes a crack width-temperature regression model to separate the reversible change component caused by temperature. Then, a crack width-load response regression model was established to separate the contribution of elastic deformation caused by vehicle load. The remaining portion represents the amount of irreversible damage propagation. The cumulative damage calculation unit is based on the Paris formula and uses data from the previous 15 days to identify material parameters. and Verification and prediction will be conducted over the next 15 days. The risk index calculation unit is set with a critical crack width value. The critical value for the spread rate is 0.3 mm. The value is 0.01 mm / day, and the weighting coefficients α and β are both 0.5. A yellow alert is triggered when the value is ≥1. A red alert is triggered when the value is ≥1.5.

[0130] Example 5: Specific Implementation of the Digital Twin Visualization Module

[0131] like Figure 5 As shown, the digital twin visualization module includes a 3D model building unit, a data mapping unit, and an early warning display unit.

[0132] The 3D model building unit constructs a 3D digital twin model of the bridge structure based on bridge design parameters (including bridge geometry, material properties, structural form, etc.). The data mapping unit is signal-connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module, mapping the output data of each module to the corresponding spatial locations in the 3D digital twin model. The early warning display unit is connected to the data mapping unit, marking cracks of different risk levels on the 3D digital twin model with different colors and / or icons, and displaying predicted crack development trends.

[0133] In one specific embodiment, the 3D model building unit constructs a 3D digital twin model based on the bridge's BIM (Building Information Modeling) design files. The data mapping unit displays the crack locations output by the physical constraint strain decoupling module as marker points on the 3D model, overlays the crack edge contours output by the physical prior image segmentation module as texture maps onto the model surface, and displays the risk index output by the physical mechanism risk assessment module using color mapping (green for low risk, yellow for medium risk, and red for high risk). The early warning display unit automatically pops up an early warning window when the risk index exceeds a threshold, highlights the location and extent of dangerous cracks on the 3D model, and generates a monitoring report containing crack location, size, risk level, development trend prediction, and maintenance recommendations.

[0134] Example 6: Data Communication and Timing Synchronization

[0135] The system of this invention also includes a data communication module, which is connected to the multi-source data acquisition module, the physical constraint strain decoupling module, the physical prior image segmentation module, the physical mechanism risk assessment module, and the digital twin visualization module, respectively, to realize data interaction between the modules and data transmission with the remote monitoring center.

[0136] Data communication between modules utilizes industrial Ethernet or 5G wireless communication. Distributed fiber optic strain data, temperature data, and image data acquired by the multi-source data acquisition module are transmitted in real-time to the physical constraint strain decoupling module and the physical prior image segmentation module via the data communication module. The processing results from each module are aggregated to the digital twin visualization module for unified display via the data communication module, and can also be uploaded to a remote monitoring center for remote viewing and analysis by management personnel.

[0137] In terms of timing synchronization, all data acquisition channels use a unified clock source for synchronization, ensuring that fiber optic strain data, temperature data, image data and vehicle load response data have a consistent time reference, providing accurate timing alignment for subsequent temperature-load coupling analysis and cumulative damage assessment.

[0138] It also includes a hierarchical working mode control module:

[0139] Normal monitoring mode: Only the distributed fiber optic sensor and temperature sensor are activated to collect strain and temperature data at a low sampling rate (e.g., 0.1Hz). The physical constraint strain decoupling module runs in lightweight inference mode. The high-precision rendering functions of the physical prior image segmentation module, physical mechanism risk assessment module, and digital twin visualization module are in a dormant or reduced-frequency state.

[0140] Fine detection mode: When the physical constraint strain decoupling module detects strain anomalies (such as local strain peak exceeding a preset threshold or identifying the initiation of new cracks), the hierarchical working mode control module wakes up the mobile inspection platform to perform bridge surface image acquisition tasks, and fully activates the physical prior image segmentation module, physical mechanism risk assessment module and digital twin visualization module to perform high-precision crack identification and risk assessment.

[0141] Regular full inspection mode: All modules are fully activated according to a preset cycle (such as once a week) to perform systematic data calibration and comprehensive evaluation.

[0142] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0143] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bridge crack monitoring device, characterized in that, include: The multi-source data acquisition module includes distributed fiber optic sensors deployed along the bridge structure axis, temperature sensors deployed at key sections of the bridge, and a mobile inspection platform equipped with a high-resolution industrial camera. The physical constraint strain decoupling module is connected to the distributed optical fiber sensor signal and is used to receive the measured strain distribution data collected by the distributed optical fiber sensor, run the multi-crack strain field hard constraint decoupling method based on physical information neural network, and output the location parameters and width parameters of the multi-crack. The physical prior image segmentation module is connected to the physical constraint strain decoupling module and the high-resolution industrial camera signal respectively. It is used to receive the crack location parameters output by the physical constraint strain decoupling module and the bridge surface image acquired by the high-resolution industrial camera. It generates a strain prior attention heatmap based on the crack location parameters, and uses the strain prior attention heatmap as a spatial attention weight to guide the image segmentation network to segment the bridge surface image and output the crack edge contour. The physical mechanism risk assessment module is connected to the physical constraint strain decoupling module and the temperature sensor signal, respectively, and is used to receive vehicle load response data. It is used to establish a temperature-load coupling response separation model based on the crack width time series data output by the physical constraint strain decoupling module, the temperature time series data collected by the temperature sensor, and the vehicle load response data, to separate the cumulative amount of irreversible damage, calculate the cumulative damage risk index, and output the risk assessment result. The digital twin visualization module is connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module, respectively, to construct a three-dimensional digital twin model of the bridge. The output data of each module is mapped to the digital twin model in real time for three-dimensional visualization.

2. The bridge crack monitoring device according to claim 1, characterized in that, The physical constraint strain decoupling module includes: The encoder unit employs a multi-layer fully connected network, using the measured strain distribution data as input to extract high-dimensional features. A parameter regression unit, connected to the encoder unit, is used to map the high-dimensional features into predicted values ​​of crack parameters. The physical forward modeling unit is a differentiable network layer, in which a multi-crack superimposed physical forward modeling model based on linear elastic fracture mechanics theory is embedded. The physical forward modeling unit is connected to the parameter regression unit and is used to calculate the theoretical strain distribution based on the predicted values ​​of the crack parameters. The training unit, connected to the physical forward modeling unit, is used to perform end-to-end training by using the data fitting term between the theoretical strain distribution and the measured strain distribution as the loss function and backpropagating the error to the parameter regression unit and the encoder unit through the chain rule.

3. The bridge crack monitoring device according to claim 2, characterized in that, The multi-crack superposition physical forward model linearly superimposes the theoretical strain distributions generated by each individual crack and adds a noise term to form the total theoretical strain distribution; wherein the theoretical strain distribution of an individual crack is determined based on the crack center position, crack width, strain amplitude coefficient related to the elastic modulus of the material, and spatial attenuation constant related to the shear hysteresis effect of the fiber bonding layer.

4. The bridge crack monitoring device according to claim 1, characterized in that, The physical prior image segmentation module includes: The coordinate mapping unit is used to map the crack location parameters output by the physical constraint strain decoupling module from the one-dimensional coordinates of the optical fiber to the two-dimensional coordinate system of the image. A heat map generation unit, connected to the coordinate mapping unit, is used to generate a strain prior attention heat map based on the mapped crack location, wherein the value of each pixel in the heat map reflects the probability that a crack exists at that location. The segmentation network unit includes a backbone encoder and a prior attention injection module. The backbone encoder is used to extract multi-scale features of the bridge surface image, and the prior attention injection module is used to inject the strain prior attention heatmap as spatial attention weights into the intermediate feature layer of the backbone encoder. The inference control unit, connected to the segmentation network unit, is used to control the segmentation network unit to use high-resolution fine inference for high-attention regions in the heatmap and low-resolution fast inference for low-attention regions, and output the crack segmentation result.

5. A bridge crack monitoring device according to claim 4, characterized in that, The heatmap generation unit generates attention weights by spreading them outwards from the location of each crack in a Gaussian distribution. The attention weights of each crack are then superimposed to form the strain prior attention heatmap.

6. The bridge crack monitoring device according to claim 1, characterized in that, The physical mechanism risk assessment module includes: The time-series data synchronization unit is used to synchronously collect crack width time-series data, bridge temperature time-series data, and vehicle load response time-series data. The response separation unit, connected to the time-series data synchronization unit, is used to establish a temperature-load coupled response separation model, which decomposes the crack width time-series data into the initial crack width, the reversible crack width change component caused by temperature, the reversible elastic deformation contribution component caused by vehicle load, and the irreversible crack propagation component caused by structural damage accumulation. The cumulative damage calculation unit, connected to the response separation unit, is used to establish a cumulative damage model based on the fatigue crack propagation law, identify material parameters online through continuous monitoring data, and calculate the cumulative amount of irreversible damage. The risk index calculation unit is connected to the cumulative damage calculation unit and is used to calculate the cumulative damage risk index based on the amount of irreversible damage and its expansion rate. When the risk index reaches a preset threshold, an early warning signal is output.

7. A bridge crack monitoring device according to claim 6, characterized in that, The response separation unit calculates the reversible crack width change component caused by temperature using a thermoelastic constitutive model and the reversible elastic deformation contribution component caused by vehicle load using a bridge influence line model. After deducting the above two reversible components from the crack width time series data, the remaining part is taken as the irreversible crack propagation amount caused by the accumulation of structural damage.

8. A bridge crack monitoring device according to claim 6, characterized in that, The risk index calculation unit performs dimensionless processing on the cumulative amount of irreversible damage and its propagation rate, respectively. Based on the critical values ​​of crack width and propagation rate, the two dimensionless indicators are weighted and summed to obtain the cumulative damage risk index.

9. A bridge crack monitoring device according to claim 1, characterized in that, The digital twin visualization module includes: The 3D model building unit is used to build a 3D digital twin model of the bridge structure based on the bridge design parameters. The data mapping unit is connected to the physical constraint strain decoupling module, the physical prior image segmentation module, and the physical mechanism risk assessment module, respectively, and is used to map the output data of each module to the corresponding spatial position of the three-dimensional digital twin model. The early warning display unit, connected to the data mapping unit, is used to mark cracks of different risk levels on the three-dimensional digital twin model with different colors and / or icons, and to display crack development trend prediction information.

10. A bridge crack monitoring device according to claim 1, characterized in that, It also includes a data communication module, which is connected to the multi-source data acquisition module, the physical constraint strain decoupling module, the physical prior image segmentation module, the physical mechanism risk assessment module and the digital twin visualization module, respectively, to realize data interaction between the modules and data transmission with the remote monitoring center.

Images