A bridge full-life-cycle damage prediction and evaluation method and system driven by digital twin

Abstract

The application discloses a kind of digital twin driven bridge full life cycle damage prediction and evaluation method and system, belong to bridge structure health monitoring field, the method includes: obtaining the monitoring dataset of bridge structure;Build and train initial digital twin model with embedded differentiable physical layer, the differentiable physical layer is based on multi-source monitoring data, to build the damage evolution model with monotonicity physical constraint, based on the physical parameter vector, the damage evolution trajectory of the bridge in future time period is predicted by damage evolution model;Combined with uncertainty quantification method, generate the time-varying reliability index of the bridge in future time period;Based on time-varying reliability index, build and solve one maintenance decision optimization model to generate the maintenance decision of bridge.The application ensures the physical authenticity and reliability of long-term prediction result.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring, and in particular to a digital twin-driven method and system for predicting and assessing bridge damage throughout its entire life cycle. Background Technology

[0002] Bridges are critical infrastructure in transportation networks. With the rapid development of transportation infrastructure construction, the safe operation and maintenance of large bridges has become a pain point in the industry. To ensure the safe operation of bridges, bridge health monitoring systems have been widely deployed to collect bridge structural response data. Utilizing the collected monitoring data and combining it with finite element models to construct digital twins of bridges has become a current technological hotspot.

[0003] However, most existing finite element models are based on design drawings, and their geometric parameters, material properties, and boundary conditions differ significantly from those of the actual bridges after construction. Although model correction techniques exist, traditional manual trial-and-error methods are inefficient and highly subjective. While purely data-driven black-box models can fit the data, their output parameters often lack clear physical meaning, resulting in poor interpretability and unconvincing results. Furthermore, during the maintenance decision-making phase, existing optimization models typically only consider numerical indicators, making it difficult to effectively utilize the large amount of unstructured textual information contained in manual inspection reports. This leads to a disconnect between the final decision and actual engineering experience.

[0004] Therefore, the challenge lies in how to construct a bridge-like digital twin technology that can accurately reflect the current state of physical entities, scientifically predict future evolution, intelligently optimize maintenance decisions, and integrate multimodal information to assist in decision-making. Summary of the Invention

[0005] One of the objectives of this invention is to provide a digital twin-driven method and system for predicting and assessing bridge damage throughout its entire life cycle, in order to solve the problems of poor physical consistency of digital twin models, life prediction that violates physical common sense, and low utilization of unstructured information in the prior art.

[0006] This invention is achieved through the following technical solution: a digital twin-driven method and system for predicting and assessing bridge damage throughout its entire life cycle, comprising the following steps: acquiring a monitoring dataset of the bridge structure, wherein the monitoring dataset includes at least fast-timescale response data reflecting the dynamic characteristics of the structure and slow-timescale data reflecting changes in the environment and loads; constructing and training an initial digital twin model embedded with a differentiable physical layer, wherein the differentiable physical layer, based on multi-source monitoring data, obtains a physical parameter vector characterizing the current physical state of the bridge through inversion, wherein the differentiable physical layer is used to calculate the effect of the model prediction error on the physical parameter vector. The gradient of the quantity is used for end-to-end training; a damage evolution model with monotonic physical constraints is constructed, and based on the physical parameter vector, the damage evolution trajectory of the bridge in the future time period is predicted by the damage evolution model, wherein the monotonic physical constraints are used to ensure that the structural stiffness predicted by the damage evolution model decreases monotonically with time; based on the damage evolution trajectory, combined with an uncertainty quantification method, a time-varying reliability index of the bridge in the future time period is generated; and based on the time-varying reliability index, a maintenance decision optimization model is constructed and solved to generate the maintenance decision of the bridge.

[0007] Furthermore, the steps for obtaining the monitoring dataset of the bridge structure include: collecting vibration time-series data obtained by acceleration sensors deployed at key locations on the bridge as the fast time-scale response data; collecting synchronously acquired ambient temperature data, as well as fatigue stress amplitude and cycle number obtained based on strain data processing, as the slow time-scale data; collecting high-resolution images of the bridge appearance obtained by UAV inspection, and using image processing algorithms to quantify and extract the geometric feature information of cracks to form structured defect data.

[0008] Furthermore, constructing and training an initial digital twin model with an embedded differentiable physical layer also includes: fusing the structured disease data into the geometric information of the initial digital twin model to correct the geometric deviations of the theoretical model.

[0009] Furthermore, the differentiable physical layer, by transforming the matrix assembly and eigenvalue solving process of the finite element model into a differentiable computational layer, includes: configuring an encoder neural network to map the input monitoring data into the physical parameter vector, wherein the physical parameter vector contains at least the elastic modulus correction factor and mass density correction factor of each structural unit; based on the physical parameter vector, multiplying the stiffness correction factor of each structural unit with a preset unit stiffness degradation matrix and superimposing it onto the reference stiffness matrix, and multiplying the mass density correction factor with a preset unit mass matrix and superimposing it onto the reference mass matrix, thereby generating a global stiffness matrix and a global mass matrix; performing dynamic solving on the assembled global stiffness matrix and global mass matrix to output predicted modal parameters; wherein the explicit matrix assembly operation and the differentiable eigenvalue solving operation are both configured to allow gradient backpropagation, enabling the encoder neural network to automatically adjust the weights according to the error between the predicted modal parameters and the measured modal parameters.

[0010] Furthermore, the physical parameter mapping and matrix assembly can be calculated using the following formula:

[0011]

[0012] in, For neural networks The output physical parameter vector includes the elastic modulus correction factor for each element. and mass density correction factor ; For encoder neural networks, Its weight; The set of measured modal parameters (including frequency and mode shape) is the input. The stiffness matrix of the baseline finite element model; The mass matrix of the baseline finite element model; The first Stiffness correction factor for each element; For the first Quality correction factor for each unit; For the first The stiffness degradation identity matrix of each element; For the first Unit mass unit matrix.

[0013] Furthermore, the solution for differentiable eigenvalues ​​and the calculation of gradients can be represented by the following equation:

[0014]

[0015]

[0016] in, For the first Eigenvalues ​​of order (corresponding to the square of the frequency). For the first First-order eigenvectors (mode shapes); This represents the sensitivity of the eigenvalues ​​to the physical parameters. In the above formula, the gradient formula is the core of the differentiable physics layer. It allows data errors to directly drive the update of the physical parameters, enabling the neural network to understand: if you want to adjust the frequency... How should the stiffness be fine-tuned to match the observed values? and quality The parameters in.

[0017] Furthermore, a composite loss function is used to train the differentiable physical layer. The composite loss function includes: a frequency error term, used to constrain the consistency between predicted and measured feature values; a mode shape correlation error term, used to maximize the modal confidence between predicted and measured mode shapes; and a sparse regularization term, used to apply L1 norm constraints to the physical parameter vector, forcing the model to find the parameter distribution with the fewest non-zero elements in order to locate the local damage location.

[0018] Furthermore, the composite loss function can be expressed by the following equation:

[0019]

[0020] in, These are the weighting coefficients; These are the predicted and measured feature values, respectively. This is a modal confidence criterion used to measure the correlation between predicted and measured mode shapes (the closer to 1, the better). This is the L1 norm sparse regularization term. It should be noted that the sparse term... The introduction of this not only prevents overfitting, but more importantly, it improves the accuracy of damage localization, forcing the model to find the simplest damage interpretation.

[0021] Furthermore, the damage evolution model with monotonic physical constraints specifically includes: employing a state additive decomposition strategy to decompose the physical parameter vector into a reversible environmental fluctuation component and an irreversible cumulative damage component; establishing a neural network ordinary differential equation for the cumulative damage component to describe the derivative relationship of damage over time; configuring the monotonic physical constraints at the output of the neural network ordinary differential equation, the monotonic physical constraints including: applying a constant positive activation function to the output of the neural network and assigning a negative weight to force the derivative of the cumulative damage component over time to always be non-positive, thereby preventing the model from predicting structural self-healing phenomena.

[0022] Furthermore, the state additive decomposition strategy can be expressed by the following equation:

[0023] Specifically, it can be expanded as follows: ,in, This refers to the irreversible parameter degradation caused by cumulative damage. This refers to the environmental fluctuation component. This is the environmental impact regression coefficient matrix; The input is an environmental vector such as temperature. This decomposition ensures that the model will not misjudge damage repair due to environmental improvements.

[0024] Furthermore, the constant differential equation can be viewed as follows:

[0025] To satisfy physical constraints, the specific functional form can be shown in the following equation: ,in, The weights of the constant differential equations; For cumulative load input, the equivalent fatigue stress amplitude is typically obtained using the rainflow counting method. and cycle number Calculated; It is an activation function whose output is always positive. The preceding negative sign forces the derivative to be positive. It is always negative, meaning the stiffness decreases monotonically. This serves as a mask for critical structural components. It's worth noting that this design mathematically eliminates the possibility of AI models predicting structural self-healing, thus significantly enhancing the overall reliability of the system.

[0026] Furthermore, the damage evolution trajectory can be calculated using the following integral formula:

[0027] ,

[0028] Furthermore, after predicting the damage evolution trajectory of the bridge over a future time period, the method further includes: defining an intelligent decision-making objective function based on minimizing maintenance costs; introducing a large language model as an auxiliary decision-making agent, inputting the unstructured text description of the bridge inspection report and a preset set of maintenance actions into the large language model; obtaining the recommendation confidence of each maintenance action output by the large language model based on engineering knowledge reasoning; and introducing a decision preference loss term into the intelligent decision-making objective function, wherein the decision preference loss term is configured to increase the cost of the objective function when the maintenance action selected by the optimization model does not match the recommendation confidence of the large language model.

[0029] Furthermore, the intelligent decision-making objective function based on minimizing maintenance costs can be expressed by the following formula:

[0030]

[0031] in, For decision variables, representing the point in time when maintenance is performed; For decision variables, representing the type of maintenance measures (e.g., traffic restriction, carbon fiber reinforcement, bearing replacement); For the designed service life; The discount rate; In order to be in Take action at all times The cost of the measures; The total economic and social losses caused by structural failure; That is Failure probability at time step ; The decision suggestion loss term output by the large language model is used to integrate expert experience and text report information; This is the adjustment coefficient.

[0032] Furthermore, by combining uncertainty quantification methods, a time-varying reliability index for the bridge within the future time period is generated. Specifically, this includes: treating the network weights of the damage evolution model as random variables; performing multiple Monte Carlo samplings on the network weights using Bayesian inference methods to generate multiple physical parameter evolution trajectories; calculating the difference between the structural resistance and the estimated maximum load effect at each time step for each evolution trajectory; statistically analyzing the proportion of differences less than zero in all sampled trajectories as the time-varying failure probability of the bridge, and calculating the time-varying reliability index accordingly.

[0033] Furthermore, the time-varying reliability index can be calculated using the following formula:

[0034] in, For a moment Reliability indicators; It is the inverse function of the standard normal distribution function; As an indicator function, when the condition is met (i.e. The value is 1 if the structure fails, otherwise it is 0.

[0035] Furthermore, the evaluation method also includes the step of dynamically updating the digital twin model based on Bayesian data assimilation theory.

[0036] Furthermore, the steps for dynamically updating the digital twin model include: setting update trigger conditions, and initiating a parameter synchronization process when the trigger conditions are met; a prediction step: starting from the state of the system's last update, using the damage evolution model to extrapolate to the current time to obtain a priori estimates of the physical parameters at the current time; an observation step: retrieving monitoring data at the current time, and independently retrieving the observed physical parameters at the current time using the differentiable physical layer; a fusion step: calculating the Kalman gain that measures the relative reliability of the prior estimate and the observation estimate, and using the Kalman gain to perform a weighted fusion of the two to obtain the posterior updated physical parameters; and writing the posterior updated physical parameters into the initial digital twin model to replace the original baseline parameters, thereby realizing the synchronous dynamic evolution of the digital twin model as the physical entity ages.

[0037] Another aspect of the present invention provides a digital twin-driven bridge life-cycle damage prediction and assessment system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the digital twin-driven bridge life-cycle damage prediction and assessment method as described above.

[0038] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the digital twin-driven method for predicting and assessing bridge life-cycle damage as described above.

[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0040] 1. This invention introduces a differentiable physical layer when constructing a digital twin model, deeply integrating the powerful fitting ability of neural networks with the rigorous physical connotation of the finite element method. This makes the model no longer a black box operation, but directly guided by the mechanical laws of the real physical world. This ensures that the derived physical parameters can accurately match the monitoring data and have real physical meaning. At the same time, by introducing a closed-loop update mechanism based on Bayesian data assimilation, the digital twin model can periodically absorb the latest measured data, perform self-correction and dynamic evolution, and ensure that the digital twin remains highly synchronized with the physical entity throughout the entire life cycle of the bridge.

[0041] 2. This invention constructs a damage evolution model with monotonic physical constraints and adopts a state additive decomposition strategy to effectively eliminate environmental interference, fundamentally preventing the occurrence of absurd predictions such as structural self-healing, ensuring the physical authenticity and reliability of long-term prediction results, and realizing reliable and credible long-term state prediction.

[0042] 3. The optimization model of this invention, which aims to minimize the total life cycle cost, directly links the reliability indicators at the technical level with the cost-effectiveness at the economic level. By introducing a Large Language Model (LLM) to process unstructured inspection reports, it realizes the integration of quantitative data and expert experience knowledge for decision-making, transforming maintenance decisions from reactive response and passive execution to proactive prediction and optimization. Attached Figure Description

[0043] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0044] Figure 1 The above is a flowchart of the overall method provided in Embodiment 1 of the present invention.

[0045] Figure 2 This is a schematic diagram of sensor network deployment provided in Embodiment 1 of the present invention.

[0046] Figure 3 The inversion process convergence curve is shown in Embodiment 1 of the present invention.

[0047] Figure 4 This is a damage rate response surface diagram provided in Embodiment 1 of the present invention.

[0048] Figure 5 This is a schematic diagram of the state additive decomposition strategy provided in Embodiment 1 of the present invention.

[0049] Figure 6 This is a diagram showing the maintenance decision optimization results provided in Embodiment 1 of the present invention.

[0050] Figure 7 This is a schematic diagram of the decision boundary offset provided in Embodiment 1 of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] Example 1

[0053] Figure 1 The overall method flowchart of this embodiment is shown. As can be seen from the figure, this embodiment includes the following steps:

[0054] Step 1: First, acquire a multi-source heterogeneous monitoring dataset for the bridge. This can be achieved by deploying a sensor network at key locations on the bridge (such as main beams, towers, and cables). This network includes accelerometers, strain gauges, displacement meters, and GPS receivers to continuously collect real-time temporal response data of the structure. Simultaneously, environmental monitoring stations are deployed to collect environmental data synchronized with the bridge's location, such as temperature, humidity, wind speed, and wind direction. Furthermore, this dataset can integrate data on bridge surface defects obtained through regular manual or drone inspections, such as the width, length, and location of cracks on the concrete surface, and the area of ​​corrosion in the steel structure. These multi-source, heterogeneous monitoring datasets together constitute a comprehensive description of the bridge's condition.

[0055] Specifically, real-time time-series response data acquisition can include: sensor network deployment can be achieved by densely deploying sensor networks at key sections of the bridge, such as the mid-span locations of the side spans and middle spans, the tops of the piers, and near the main girder supports. Triaxial accelerometers can be installed on the inner and outer sides of the web at L / 4, L / 2, and 3L / 4 (L being the span length) of each span along the length of the main girder to capture the structure's vibration characteristics; high-precision fiber Bragg grating (FBG) strain gauges can be attached to the upper and lower edges of the mid-span section of the main girder and at the base of the main piers to measure the structure's strain response; in addition, GPS receivers or static levels can be installed on the main girder and cables to monitor the structure's long-term displacement and creep. All dynamic sensors (acceleration, strain) are configured to continuously and synchronously acquire data at a sampling frequency of 100Hz, and the acquired data is collected by a data acquisition instrument. The acquired real-time time-series response data includes: acceleration time series, strain time series, displacement time series, etc., at each measuring point.

[0056] Environmental monitoring data can be collected synchronously with a sensor network by installing a small weather station near the bridge site. This data includes ambient temperature, bridge surface temperature, humidity, wind speed, and wind direction. Temperature data can be collected at a lower frequency, such as once every 5 minutes, while wind load data needs to be collected at a frequency of 100Hz, synchronized with the structural response. Figure 2 A schematic diagram of the sensor network deployment in this implementation is shown.

[0057] By conducting a comprehensive routine inspection every six months, drones (UAVs) equipped with high-definition zoom cameras are used to take all-round, close-up photos of the bridge, capturing high-definition images of the bridge deck, main beams, piers, and other parts. Then, using image recognition algorithms in computer vision, these images are post-processed to automatically identify and quantify cracks on the concrete surface, extracting their precise location, length, average width, and direction information in the bridge's three-dimensional coordinate system, thereby forming a structured database of defects.

[0058] The accelerometer collects vibration data at a frequency of 100Hz, which is used for subsequent extraction of modal parameters such as frequency and mode shape. ), that is, fast timescale data ( The environmental temperature, humidity, and traffic flow data collected simultaneously, along with the daily average equivalent fatigue stress amplitude and cycle number obtained by processing strain data using the rainflow counting method, constitute the environment-load vector. ), that is, slow timescale data ( ).

[0059] Step 2: Construct an initial baseline digital twin model with an embedded differentiable physical layer. Based on the bridge's design drawings and construction data, a three-dimensional finite element model (FEM) of the bridge is established using finite element analysis software (such as ANSYS, ABAQUS, etc.). This model includes detailed geometric information, material properties, and element meshing. Specifically, an UAV equipped with an oblique photogrammetry camera can be used to scan the actual bridge to obtain its high-precision three-dimensional point cloud data. Then, using a point cloud registration algorithm, the measured three-dimensional point cloud model is aligned and compared with the theoretical model in the design drawings to identify and correct any geometric deviations that may exist between the theoretical model and the actual construction (such as construction errors in beam height and slab thickness).

[0060] To eliminate the uncertainty of physical parameters (such as actual elastic modulus and boundary constraint stiffness) of digital twin models, this embodiment calibrates the model by constructing a differentiable computing architecture that deeply integrates physics and data.

[0061] By defining a core physical parameter vector containing all parameters to be corrected (including the elastic modulus correction factor and mass density correction factor of each superelement), a differentiable finite element assembly operator is introduced as a bridge connecting the neural network and the physical model. Specifically, this operator receives the physical parameter vector predicted by the neural network and explicitly assembles the current global stiffness matrix and mass matrix using the pre-calculated derivatives of the element stiffness matrix and mass matrix. Subsequently, by connecting a differentiable eigenvalue solving layer, the assembled global matrix is ​​dynamically solved, thereby outputting the predicted modal frequencies and mode shapes. Automatic differentiation techniques or adjoint methods are then used to calculate the predicted modes and the measured modes in step 1. The error gradient between the eigenvalues ​​and the physical parameters is calculated. In this way, the error gradient can propagate back through the complex eigenvalue solving process, driving the neural network to automatically adjust the physical parameter vector until the predicted dynamic characteristics highly match the measured data. Finally, the converged physical parameters are assigned to the finite element model, thus completing the construction of the initial baseline digital twin model.

[0062] Specifically, in this embodiment, the differentiable computing architecture with deep fusion of physical and data can be constructed through the following steps:

[0063] 1) First, define the core physical parameter vector (i.e. manifold coordinates) of the entire life cycle of the bridge, and model the entire life cycle of the bridge as a dynamic system on a time-varying parameter differential manifold. Through the state recognition algorithm with embedded differentiable physical layer, the current physical state is inverted from the monitoring data.

[0064] Specifically, the health status of a bridge is not directly measurable but is determined by its material properties. Therefore, to accurately describe the bridge's behavior at different time scales, a core state variable needs to be defined. In this embodiment, the bridge's physical parameters are considered as a variable that varies with a 'slow time scale'. '(used to describe damage evolution, month / grade) change vector, while the structural response is the parameter at a 'fast timescale'. '(Used to describe vibration, in milliseconds) performance.'

[0065] For example, in this embodiment, the core physical parameter vector is defined as follows:

[0066]

[0067] in, for The physical parameter vector at time t, its dimension is , This represents the number of finite element super elements. This is the normalized stiffness (elastic modulus correction factor) of each element. This is the mass density correction factor for each unit; This is the transpose symbol for a matrix.

[0068] 2) After establishing the core physical parameter vector, in order to inversely determine the true distribution of internal physical parameters of the structure from massive monitoring data containing only surface features such as frequencies and mode shapes, a mapping from observed data to physical parameters needs to be constructed. To ensure that the inverted parameters conform to mechanical principles and are not meaningless values ​​generated by a black-box neural network, this embodiment introduces a differentiable finite element assembly operator. This involves constructing an encoder network embedded with the differentiable finite element assembly operator, forcibly mapping the neural network output to stiffness and mass matrices with explicit physical meaning, and explicitly assembling the correction factors output by the neural network into global stiffness and mass matrices. This ensures that the network generates not meaningless numbers, but physical parameters that conform to the principles of structural mechanics.

[0069] For example, in this embodiment, the physical parameter mapping and matrix assembly can be calculated using the following formula:

[0070]

[0071] in, For neural networks The output physical parameter vector includes the elastic modulus correction factor for each element. and mass density correction factor ; For encoder neural networks, Its weight; The set of measured modal parameters (including frequency and mode shape) is the input. The stiffness matrix of the baseline finite element model; The mass matrix of the baseline finite element model; The first Stiffness correction factor for each element; For the first Quality correction factor for each unit; For the first The stiffness degradation identity matrix of each element; For the first Unit mass unit matrix.

[0072] The above equation assembles the global stiffness matrix explicitly. and global quality matrix The prediction results of deep learning are forcibly constrained within a solution space that conforms to the principles of mechanics.

[0073] 3) After assembling the two global matrices, in order for the neural network to self-correct based on the errors in the dynamic equations, the usually non-differentiable numerical process of solving for eigenvalues ​​needs to be transformed into a differentiable layer in the network. By solving the generalized eigenvalue problem and using Nelson's method or matrix calculus to calculate the gradient of the eigenvalues ​​with respect to the physical parameters, the error signal can be backpropagated through the physical equations back to the neural network, thereby using measured data to correct the physical parameters.

[0074] For example, in this embodiment, the solution for differentiable eigenvalues ​​and the calculation of gradients can be represented by the following equation:

[0075]

[0076]

[0077] in, For the first Eigenvalues ​​of order (corresponding to the square of the frequency). For the first First-order eigenvectors (mode shapes); This represents the sensitivity of the eigenvalues ​​to the physical parameters. In the above formula, the gradient formula is the core of the differentiable physics layer. It allows data errors to directly drive the update of the physical parameters, enabling the neural network to understand: if you want to adjust the frequency... How should the stiffness be fine-tuned to match the observed values? and quality The parameters in.

[0078] 4) Finally, to better train the above model, a composite loss function can be constructed that not only requires the predicted frequencies and mode shapes to be consistent with the measured data, but also introduces a sparse regularization term. This is because, in most cases, the damage is localized, meaning that most elements remain healthy (parameters unchanged), and only a few elements experience abrupt parameter changes.

[0079] For example, in this embodiment, the physical consistency loss function can be expressed as follows:

[0080]

[0081] in, These are the weighting coefficients; These are the predicted and measured feature values, respectively. This is a modal confidence criterion used to measure the correlation between predicted and measured mode shapes (the closer to 1, the better). This is the L1 norm sparse regularization term. It should be noted that the sparse term... The introduction of this not only prevents overfitting, but more importantly, it improves the accuracy of damage localization, forcing the model to find the simplest damage interpretation.

[0082] Figure 3 This is a convergence curve of the inversion process of the differentiable computational architecture parameters in the embodiment. The horizontal axis represents the number of training rounds, and the vertical axis represents the stiffness correction coefficients of each unit derived by the neural network. In the figure, the curves for Unit 1 and Unit 2 represent healthy units, whose parameters oscillate in the early stages of training before rapidly converging to high values; while the curve for Unit 3 represents damaged units, whose parameters converge to lower values. The figure shows that introducing a sparse regularization term into the loss function... By forcing most units to remain healthy and only allowing the parameters of truly damaged units to decrease, accurate damage localization was achieved, proving that the model was not a "black box" fit but rather physically interpretable.

[0083] Step 3: After obtaining the current accurate physical state, establish a dynamic model of damage evolution over time to make a preliminary prediction of the future damage evolution of the bridge.

[0084] Specifically, changes in bridge physical parameters are typically composed of two parts: fluctuations caused by environmental factors and degradation caused by cumulative damage. To eliminate the interference caused by seasonal temperature differences, a state additive decomposition strategy can be adopted to decompose the physical parameters into reversible environmental fluctuation terms and irreversible damage terms. A damage evolution manifold model can then be constructed on a slow timescale. The trajectory of the extrapolated parameters is used to predict the entire life cycle.

[0085] Specifically, in this implementation, the damage evolution manifold model can be constructed through the following steps:

[0086] 1) First, seasonal temperature changes can cause reversible fluctuations in material modulus, but this does not necessarily mean that the health of the structure has changed. Therefore, a state additive decomposition strategy can be adopted to decompose physical parameters into irreversible damage terms and reversible environmental fluctuation terms, thereby separating the influence of environmental factors from the changes in physical parameters.

[0087] For example, in this embodiment, the additive decomposition of states can be expressed by the following formula:

[0088]

[0089] Specifically, this can be expanded as follows:

[0090]

[0091] in, This refers to the irreversible parameter degradation caused by cumulative damage. This refers to the environmental fluctuation component. This is the environmental impact regression coefficient matrix; The input is an environmental vector such as temperature. This decomposition ensures that the model will not misjudge damage repair due to environmental improvements.

[0092] 2) Then, for the core damage evolution part, a differential equation describing its change over time can be established. Since the damage is irreversible, the structural stiffness will only decrease monotonically with time. To this end, a monotonic projection operator can be introduced to force the derivative of the neural network output to always be negative (or positive, depending on the parameter definition), thereby preventing the model from predicting phenomena that violate physical common sense, such as structural self-healing.

[0093] For example, the ordinary differential equations (ODEs) in this embodiment are shown below:

[0094]

[0095] To satisfy physical constraints, the specific functional form can be shown in the following equation:

[0096]

[0097] in, The weights of the constant differential equations; For cumulative load input, the equivalent fatigue stress amplitude is typically obtained using the rainflow counting method. and cycle number Calculated; It is an activation function whose output is always positive. The preceding negative sign forces the derivative to be positive. It is always negative, meaning the stiffness decreases monotonically. This serves as a mask for critical structural components. It's worth noting that this design mathematically eliminates the possibility of AI models predicting structural self-healing, thus significantly enhancing the overall reliability of the system.

[0098] Figure 4 This is the damage rate response surface plot of the damage evolution manifold model in this embodiment. The plot is a three-dimensional surface showing the nonlinear relationship between the damage rate (Z-axis) and the current stiffness (X-axis) and the external load (Y-axis). The Z-value is less than 0 throughout the entire domain. As can be seen from the plot... The dynamic characteristics of the surface are learned by a neural network, reflecting the nonlinear physical law that the greater the load, the faster the failure, and the weaker the structure, the faster the failure. The negative value across the entire domain directly verifies the effectiveness of the monotonicity constraint.

[0099] 3) Finally, based on the constructed differential equation, the future state is obtained by integrating over time.

[0100] For example, in this embodiment, the integral of the evolution trajectory can be calculated using the following formula:

[0101]

[0102] The integration process employs the Runge-Kutta 45 (RK45) solver. This solver automatically adjusts the step size based on the error, accurately calculating the time step from the current moment. To the future The cumulative amount of damage.

[0103] Figure 5 This diagram illustrates the state additive decomposition strategy of the damage evolution manifold model in the embodiment. The dashed lines in the diagram represent the raw physical parameter observations directly acquired by the sensors. The curve exhibits obvious periodic fluctuations, corresponding to environmental noise; while the solid line shows the damage evolution trend obtained after decomposition by this algorithm. The curve is smooth and monotonically decreasing; the gray shaded area represents the isolated environmental fluctuation component. As can be seen from the figure, by dividing the dynamic system into a fast timescale (vibration response) and a slow timescale (damage evolution), and introducing a linear regression layer to handle the environmental term, the problem of spurious stiffness fluctuations caused by seasonal temperature differences was successfully solved, ensuring that the subsequent life prediction model is trained only on the actual cumulative structural damage.

[0104] Step 4: The future state obtained through the above steps is a single deterministic prediction. However, because both the load and model parameters are random, a single state is often insufficient to support decision-making in practical engineering. A decision-making layer needs to be constructed to transform this single deterministic prediction into probabilistic indicators required for engineering decisions.

[0105] In this embodiment, the weights of the ODE network are... Treating them as random variables, we combine Bayesian inference methods to perform multiple samplings to generate a set of possible future evolution trajectories, thereby covering all possible future states.

[0106] For example, in this embodiment, Bayesian-based fluid network prediction can be expressed by the following formula:

[0107]

[0108] in, A set for predicting the distribution of physical parameters; This represents the total number of Monte Carlo samplings. For the first The evolution curves of physical parameters obtained from Monte Carlo sampling.

[0109] After obtaining the distribution of physical parameters, it is necessary to transform them into a judgment on whether the structure will fail. By defining the limit state equation, the physical parameters are mapped to the resistance of the structure and compared with the maximum load effect in the future.

[0110] For example, in this embodiment, the limit state equation can be expressed as follows:

[0111]

[0112] in, For the first The trajectory in The function value at time; This is a structural resistance function (e.g., section bending capacity) calculated based on current physical parameters. The maximum estimated load effect in the future. When When this occurs, it indicates structural failure.

[0113] Finally, based on the limit state equation, the time-varying reliability index is calculated by statistically analyzing the proportion of structural failures in multiple samplings. This transforms the complex physical state into an intuitive risk probability value, thereby converting technical predictions into risk probabilities that managers can understand.

[0114] For example, in this embodiment, the time-varying reliability index can be calculated using the following formula:

[0115]

[0116] in, For a moment Reliability indicators; It is the inverse function of the standard normal distribution function; As an indicator function, when the condition is met (i.e. The value is 1 if structural failure occurs, and 0 otherwise. It should be noted that this formula outputs a reliability curve that decreases over time. This allows managers to clearly see that in a future year, the structural reliability index will fall below the warning threshold (e.g., This allows for the precise formulation of maintenance plans.

[0117] It should be noted that the final output of this formula is no longer a simple lifetime value, but a reliability curve with a confidence interval; when When the probability of failure drops to a preset threshold, it means that the probability of structural failure has exceeded the acceptable range, thus triggering a maintenance decision. This method effectively quantifies the decision risk, enabling engineering managers to clearly understand the degree of certainty in predicting the outcome.

[0118] Step 5: Using the above reliability indicators as the core input for optimization, construct an intelligent decision-making model based on minimizing maintenance costs to determine when and how to maintain the bridge.

[0119] After obtaining the reliability index that changes over time Subsequently, in order to balance maintenance costs and failure risks, a global optimization model is needed to minimize the total expected cost over the entire lifecycle.

[0120] Specifically, in this embodiment, the total cost can be composed of two parts: first, the direct cost of preventive maintenance (such as reinforcement and repair); and second, the expected cost of the risk brought about by failure (the probability of failure multiplied by the consequences of failure). Furthermore, in order to handle complex unstructured information (such as crack descriptions in visual inspection reports), this embodiment introduces a Large Language Model (LLM) as an auxiliary decision-making term, thereby enabling the correction of the optimization objective through penalty or reward.

[0121] For example, in this embodiment, the intelligent maintenance decision optimization model can be represented by the following formula:

[0122]

[0123] in, For decision variables, representing the point in time when maintenance is performed; For decision variables, representing the type of maintenance measures (e.g., traffic restriction, carbon fiber reinforcement, bearing replacement); For the designed service life; The discount rate; In order to be in Take action at all times The cost of the measures; The total economic and social losses caused by structural failure; That is Failure probability at time step ; The decision suggestion loss term output by the large language model is used to integrate expert experience and text report information; This is the adjustment coefficient.

[0124] It should be noted that this model uses the failure probability calculated by the physical model. This directly translates into economic indicators, allowing engineering decisions to move away from intuition and instead be based on quantitative cost-risk analysis. By solving this optimization problem, the optimal decision variables can be directly output. (Best maintenance year) and (Best maintenance method). Figure 6 This is a diagram showing the maintenance decision optimization results obtained from the intelligent maintenance decision optimization model in this embodiment. The blue area on the left Y-axis represents the predicted trajectory cloud map of the future physical state generated through Bayesian inference; the purple curve on the right Y-axis represents the total expected cost curve changing over time. The lowest point of the purple curve is the calculated optimal maintenance time.

[0125] In addition, in this embodiment, the Large Language Model (LLM) assisted decision-making process includes the following sub-steps:

[0126] 1) First, collect the text description fields from the most recent bridge periodic inspection report, such as: "Several transverse cracks appeared in the bottom slab of the mid-span, with a maximum width of 0.2mm, accompanied by white crystal precipitation, suspected to be the loss of hydration products."

[0127] At the same time, define a standard set that includes all feasible maintenance actions:

[0128] .

[0129] Next, structured cue words specific to the civil engineering field are constructed and input into a large language model that has been fine-tuned for the field.

[0130] 2) After receiving the input, the large language model uses its pre-trained massive engineering knowledge base to perform inference chain analysis. For example, for "transverse crack + crystal precipitation", the model will infer that this often means through cracks and possible steel corrosion risk. Simple surface coating cannot solve the stress problem, so it will be given a low score; while pasting steel plates or external prestressing can actively improve the cross-sectional bearing capacity, so it should be given a high score.

[0131] Finally, the model outputs a suggestion confidence vector. .

[0132] For example, in this embodiment, the final suggested confidence vector output by the fine-tuned Qwen-30B is: This vector intuitively quantifies the degree of matching between various measures and the current actual disease characteristics.

[0133] 3) In order to incorporate the LLM recommendations into the intelligent maintenance decision optimization model, this embodiment defines a decision preference loss term. for:

[0134] ,

[0135] In other words, if the optimization algorithm chooses a measure with a very low LLM score (not recommended), It will approach 1 (high penalty); otherwise it will approach 0.

[0136] Figure 7 This diagram illustrates the decision boundary offset driven by LLM in this embodiment. The solid line in the diagram represents the cost curve after incorporating the LLM analysis report. Driven by the risk penalty term generated by the severely corroded text report, the optimal maintenance point is advanced to the 3rd year. This embodiment utilizes a large language model to process unstructured data and transforms qualitative expert experience into quantitative cost penalties, thereby correcting the decision bias of the pure mathematical model and improving the safety of the decision.

[0137] Step 6: Finally, based on Bayesian data assimilation theory, the theoretically predicted state output in Step 3 can be fused with the measured physical state obtained in Step 2 at the current moment, thereby updating the physical parameters of the digital twin model and eliminating accumulated errors. This is a key step in achieving a closed information loop between the physical entity and the digital twin, enabling the digital twin to evolve dynamically over time.

[0138] Specifically, this includes: First, defining a dynamic update trigger mechanism. The system is set to initiate the parameter synchronization process based on a fixed time interval (e.g., every Monday morning) or event-driven conditions (e.g., detecting a peak acceleration at mid-span of the main beam exceeding 0.1g). Next, performing model parameter updates based on Bayesian filtering (such as Extended Kalman Filter (EKF) or Particle Filter (PF)) involves the following three sub-steps:

[0139] 1) Generate prior state estimates (i.e., prediction steps), using the neural ordinary differential equations trained in step 3, based on the system's last updated state. Starting from this point, and combining the accumulated load during this period, the integral is extrapolated to the current moment. To obtain the prior estimates of the physical parameters at the current moment. and its predicted covariance matrix .

[0140] 2) Obtain the observation state estimate (i.e., the observation step) and retrieve the current time. The latest vibration acceleration data is used to run the physical-data fusion network (i.e., the encoder and differentiable assembly layer) in step 2 again. Without relying on historical information, it independently retrieves the measured physical parameter vector at the current moment. .

[0141] 3) Calculate the posterior update (i.e., the fusion step) and calculate the Kalman gain. This gain measures whether the model predictions or the experimental inversions should be trusted more. The optimal posterior physical parameters are calculated using the following formula:

[0142] .

[0143] 4) Finally, the calculated results (Including the updated elastic modulus of all elements of the entire bridge) and mass density The stiffness matrix is ​​automatically written into the material property table of the finite element software through a scripting language (such as APDL or Python API), replacing the original reference parameters and regenerating the stiffness matrix.

[0144] Through this mechanism, the initial baseline digital twin model constructed in step 2 can be updated in real time, enabling it to evolve from a static digital model into a time-varying digital model that can age synchronously with the physical entity, absorb new data in real time, and self-correct. This provides a decision-making benchmark that is highly consistent with the actual physical state for the full life cycle health management of the bridge.

[0145] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A digital twin-driven method for predicting and assessing bridge damage throughout its entire lifecycle, characterized in that, The evaluation method includes: Obtain a monitoring dataset of the bridge structure, wherein the monitoring dataset includes at least fast timescale response data to reflect the dynamic characteristics of the structure and slow timescale data to reflect changes in the environment and loads. Construct and train an initial digital twin model with an embedded differentiable physical layer. The differentiable physical layer, based on multi-source monitoring data, obtains a physical parameter vector characterizing the current physical state of the bridge through inversion, wherein... The differentiable physical layer is used to calculate the gradient of the model prediction error with respect to the physical parameter vector, and to perform end-to-end training using the gradient. A damage evolution model with monotonic physical constraints is constructed, and based on the physical parameter vector, the damage evolution trajectory of the bridge in future time periods is predicted using the damage evolution model. The monotonic physical constraint is used to ensure that the structural stiffness predicted by the damage evolution model decreases monotonically over time. Based on the damage evolution trajectory and combined with uncertainty quantification methods, a time-varying reliability index for the bridge within the future time period is generated; and Based on the time-varying reliability index, a maintenance decision optimization model is constructed and solved to generate maintenance decisions for the bridge. The differentiable physical layer transforms the matrix assembly and eigenvalue solving process of the finite element model into a differentiable computational layer, including: Configure an encoder neural network to map the input monitoring data into the physical parameter vector, wherein the physical parameter vector includes at least the elastic modulus correction factor and the mass density correction factor of each structural unit; Based on the physical parameter vector, the elastic modulus correction factor of each structural unit is multiplied by the preset unit stiffness degradation matrix and superimposed on the reference stiffness matrix, and the mass density correction factor is multiplied by the preset unit mass matrix and superimposed on the reference mass matrix, thereby generating the global stiffness matrix and the global mass matrix. The global stiffness matrix and global mass matrix after assembly are solved dynamically, and the predicted modal parameters are output. In this context, both the explicit matrix assembly operation and the differentiable eigenvalue solving operation are configured to allow gradient backpropagation, enabling the encoder neural network to automatically adjust the weights based on the error between the predicted mode parameters and the measured mode parameters. The damage evolution model with applied monotonic physical constraints specifically includes: The physical parameter vector is decomposed into a reversible environmental fluctuation component and an irreversible cumulative damage component by adopting a state additive decomposition strategy. Establish a neural ordinary differential equation for the cumulative damage component to describe the derivative relationship of the damage over time; The monotonic physical constraints are configured at the output of the divine constant differential equation. The monotonic physical constraints include: A constant positive activation function is applied to the output of the neural network and a negative weight is assigned to force the derivative of the accumulated damage component with time to always be non-positive, thereby preventing the model from predicting the self-healing phenomenon of the structure. After predicting the damage evolution trajectory of the bridge over a future time period, the method also includes: Define an intelligent decision-making objective function based on minimizing maintenance costs; A large language model is introduced as an auxiliary decision-making agent, and the unstructured text description of the bridge inspection report and the preset set of maintenance actions are input into the large language model; Obtain the suggested confidence scores of each maintenance action output by the large language model based on engineering knowledge reasoning; A decision preference loss term is introduced into the intelligent decision objective function. The decision preference loss term is configured to increase the cost of the objective function when the maintenance action selected by the optimization model does not match the suggestion confidence of the large language model. The evaluation method also includes: The steps for dynamically updating the digital twin model based on Bayesian data assimilation theory.

2. The digital twin-driven bridge life-cycle damage prediction and assessment method according to claim 1, characterized in that, The steps for obtaining the monitoring dataset of the bridge structure include: Vibration time-series data acquired by acceleration sensors deployed at key locations on the bridge are used as the fast time-scale response data; The ambient temperature data acquired synchronously, as well as the fatigue stress amplitude and cycle number obtained based on strain data processing, are used as the slow time-scale data. High-resolution images of the bridge surface were collected by drone inspections, and image processing algorithms were used to quantify and extract the geometric features of cracks to form structured defect data.

3. The digital twin-driven bridge life-cycle damage prediction and assessment method according to claim 1, characterized in that, The method of combining uncertainty quantification to generate a time-varying reliability index for the bridge in the future time period specifically includes: The network weights of the damage evolution model are treated as random variables; The network weights are sampled multiple times using the Bayesian inference method to generate multiple physical parameter evolution trajectories; For each evolution trajectory, calculate the difference between the structural resistance and the estimated maximum load effect at each moment; The proportion of differences less than zero in all sampled trajectories is used as the time-varying failure probability of the bridge, and the time-varying reliability index is calculated accordingly.

4. The digital twin-driven bridge life-cycle damage prediction and assessment method according to claim 1, characterized in that, The steps for dynamically updating the digital twin model include: Set update trigger conditions, and start the parameter synchronization process when the trigger conditions are met; Prediction step: Starting from the state of the system in the last update, the damage evolution model is used to extrapolate to the current time to obtain a priori estimate of the physical parameters at the current time; Observation step: Retrieve the monitoring data at the current moment and use the differentiable physical layer to independently invert the observed physical parameters at the current moment; Fusion step: Calculate the Kalman gain that measures the relative confidence of the prior estimate and the observation estimate, and use the Kalman gain to perform a weighted fusion of the two to obtain the posterior updated physical parameters; The a posteriori updated physical parameters are written into the initial digital twin model to replace the original baseline parameters, thereby realizing the synchronous dynamic evolution of the digital twin model as the physical entity ages.

5. A digital twin-driven bridge life-cycle damage prediction and assessment system, characterized in that, The evaluation system includes: processor; The memory stores a computer program that, when executed by a processor, implements the digital twin-driven method for predicting and assessing bridge life-cycle damage as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the digital twin-driven method for predicting and assessing bridge damage throughout its entire life cycle as described in any one of claims 1 to 4.

Images