Bridge bearing capacity dynamic evaluation method based on multi-source monitoring data driving

Abstract

The application discloses a kind of bridge bearing capacity dynamic evaluation methods based on multi-source monitoring data driving, to solve the problem that bearing evaluation accuracy is difficult to guarantee under sparse sensing coverage condition, the application is through event anchor point guided cross-modal attention alignment, influence line and propagation time delay constraint mask and geometric position coding, in physical constraint neural operator model, introduce Green function kernel and time-varying boundary flexibility latent variable and use symplectic integration to promote, according to the weight of working condition Block conforming interval estimation and order preserving calibration, and coverage and physical residual coupling self-correction closed loop, calculate the space-time response and bearing utilization coefficient of whole bridge, output risk level and limit load suggestion, realize the technical effect of reliability improvement, physical consistency and long-time energy stability under complex working condition and low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering, and in particular to a dynamic assessment method for bridge bearing capacity based on multi-source monitoring data. Background Technology

[0002] With the development of bridge structural health monitoring (SHM) systems, sensors for strain, acceleration, displacement, temperature, and wind speed are widely deployed on-site. Combined with traffic monitoring of axle loads and passage time, this enables the acquisition of bridge responses and load-bearing capacity assessments under operational conditions. Existing technological approaches include: model updates and load-bearing calculations based on finite element / influence line methods; fusion and time synchronization of multi-source data; and incorporating physical constraints into machine learning (such as PINN / PINO) to improve generalization capabilities. Simultaneously, research is also being conducted using statistical and Bayesian methods for uncertainty quantification and risk classification.

[0003] However, existing technologies still have the following shortcomings:

[0004] 1. In field conditions with sparse sensor coverage, uneven signal-to-noise ratio, and missing data, commonly used cross-modal fusion and time alignment often rely on global time windows or simple correlations, failing to fully utilize the physical constraints of traffic events and influence lines / propagation delays, which can easily lead to mismatches and cause systematic deviations in the accuracy of load assessment.

[0005] 2. Physical constraint modeling often assumes constant boundary conditions, and the time-progression method does not conserve energy, making it difficult to characterize the time-varying characteristics of support flexibility and the stability of long-term dynamic response, resulting in insufficient satisfaction of dynamic equilibrium equations and boundary conditions.

[0006] 3. Regarding uncertainty assessment, the interval coverage is uneven under different traffic flow conditions and lacks monotonic consistency constraints with axle load intensity or total queue load, making it impossible to guarantee physical reasonable order preservation. At the same time, the lack of a self-correcting closed loop that couples coverage with physical residuals makes it difficult for the model to adaptively improve reliability in long-term operation.

[0007] Therefore, a dynamic assessment method for bridge bearing capacity that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a dynamic bridge load-bearing capacity assessment method driven by multi-source monitoring data. Addressing the problems of large cross-modal alignment errors, difficulty in characterizing time-varying boundary conditions, and uneven coverage in uncertainty assessments caused by sparse sensors, uneven signal-to-noise ratios, and data gaps in existing technologies, this invention proposes an event-anchor-guided cross-modal attention alignment. It constructs a physical constraint mask based on influence lines and propagation delays, combined with geometric position encoding and reliability weighting. A Green's function kernel and time-varying boundary flexibility latent variables are introduced into the physical constraint neural operator, employing symplectic integrals. Training is performed using observability weighting and dynamic equilibrium residual constraints to solve for the spatiotemporal response of the entire bridge's displacement and strain, as well as the load-bearing capacity utilization coefficient. Based on block conformal interval estimation and order-preserving calibration, a closed-loop self-correction scheme is implemented, coupling coverage and physical residuals. This invention achieves the technical effects of improving alignment and assessment accuracy under complex working conditions and low signal-to-noise ratios, ensuring long-term energy stability and physical consistency, achieving balanced coverage and order-preserving intervals, and continuously approaching the coverage target and reducing physical residuals through a closed-loop mechanism.

[0009] A method for dynamic assessment of bridge bearing capacity based on multi-source monitoring data according to an embodiment of the present invention includes:

[0010] S1. Collect time-series data of strain, acceleration, displacement, temperature, and wind speed, axle load and passage time, vehicle speed, lane, sensor spatial coordinates, bridge span, support position, and material parameters, and establish a bridge coordinate system and a unified time reference.

[0011] S2. Based on axle load and passage time, extract axle pulse event anchor points. Combine sensor coordinates, span and support position, construct a physical constraint mask according to the influence line and propagation delay and generate geometric position code. Obtain sensor reliability weights based on the signal-to-noise ratio and missing rate of response data. Perform cross-modal attention alignment on multi-source data under the mask and coding constraints and obtain dynamic balance equation parameter values.

[0012] S3. Establish a physical constraint neural operator model, set the equivalent boundary compliance as a time-varying boundary compliance latent variable and jointly estimate it with the model parameters, construct a loss containing observable weighting terms and dynamic equilibrium equation residual terms, and align features and event anchors to train or update the model.

[0013] S4. Based on the trained model and boundary flexibility latent variables, combine the alignment features and dynamic equation parameters to solve the spatiotemporal response field of the full bridge displacement and strain and calculate the physical residuals. Calculate the internal force effect of key sections based on structural geometry and material parameters and the spatiotemporal response field and form a bearing capacity utilization coefficient sequence.

[0014] S5. Using the load utilization coefficient sequence, combined with event anchor points and axle load, vehicle speed, and lane, extract traffic flow condition labels and divide time blocks according to event anchor points. Perform block shape-preserving interval estimation according to condition weighting to obtain uncalibrated confidence intervals. Apply monotonically non-decreasing constraints based on axle load strength or total queue load to perform order-preserving calibration and obtain calibrated confidence intervals.

[0015] S6. Calculate the coverage rate and compare it with the preset target. Combine the physical residuals with the coupling rules to obtain the constraint adjustment parameters. Based on this, update the physical residual loss weight and the physical constraint mask threshold of cross-modal attention in the physical constraint neural operator model for self-correction in the next evaluation cycle. Calculate the risk level and generate load limit recommendation parameters based on the load utilization coefficient sequence and calibration confidence interval.

[0016] Optionally, step S1 specifically includes:

[0017] Establish a bridge coordinate system and time reference. Determine the origin and axis of the bridge coordinate system based on the bridge span and support location. Set a unified time reference as the timestamp and sampling frequency for each time series data recording based on the clock of the acquisition equipment.

[0018] Perform sensor geometric calibration, mark the installation position and orientation of each sensor in the bridge coordinate system according to the sensor spatial coordinates, and check the consistency with the bridge span and support position;

[0019] Collect strain time-series data, acceleration time-series data, and displacement time-series data, and record the corresponding timestamps and sampling frequencies;

[0020] Collect axle load, axle passage time, vehicle speed and lane information, and record the corresponding timestamps;

[0021] Collect time-series temperature and wind speed data and record the corresponding timestamps and sampling frequencies;

[0022] Collect sensor spatial coordinates, bridge span, support location, and material parameters, and verify the consistency of the units and dimensions of the material parameters;

[0023] Output strain time series data, acceleration time series data, displacement time series data, axle load, axle passage time, vehicle speed, lane, temperature time series data, wind speed time series data, sensor spatial coordinates, bridge span, support location, and material parameters.

[0024] Terminology definition:

[0025] The bridge span is the set of horizontal projected distances between adjacent support centerlines. For single-span bridges, the nominal span is used, while for continuous or multi-span bridges, the span is given separately for each span.

[0026] The support position refers to the spatial coordinates of each support or pier top in the bridge coordinate system and its boundary type identifier, which is used to define the structural boundary conditions and influence line reference points.

[0027] The material parameters are a set of material and cross-sectional mechanical parameters required for dynamic equilibrium and internal force calculation, including but not limited to elastic modulus E, shear modulus G, density ρ, Poisson's ratio ν, cross-sectional area A, moment of inertia I, etc.

[0028] The strain time series data is a sequence of structural strain measurements recorded over time at a specified measurement point. The unit is preferably microstrain με or engineering strain, and includes a timestamp corresponding to a unified time reference.

[0029] The acceleration time series data is an acceleration response sequence along a set measurement direction at a specified measurement point, with units of m / s² or g, and includes timestamps corresponding to a unified time reference.

[0030] The displacement time series data is a displacement response sequence along a set direction at a specified measuring point, with units of mm or m, which can be directly measured by a displacement sensor or estimated from acceleration / strain.

[0031] The temperature time series data is a recorded sequence of structural or ambient temperature over time, in ℃ or K, used to characterize the effect of temperature on structural response.

[0032] The wind speed time series data is a record sequence of wind speed at the bridge site over time, in m / s, and can be a scalar average wind speed or a component along the bridge axis / lateral direction.

[0033] The axle load is the equivalent vertical load measured by the weighing or load detection system for each axle of the vehicle, in units of N or kN, used to characterize the external excitation intensity of the axle on the structure.

[0034] The axle passage time is the timestamp of each axle when it passes through a preset event reference position (e.g., mid-span or support line), which is used for event anchor point extraction and time alignment;

[0035] The vehicle speed is the speed at which the vehicle travels on the bridge or within the monitoring area, in units of m / s or km / h, and can be calculated from the continuous axle event interval or the speed measuring device.

[0036] The lanes are the bridge deck lane numbers or markings for vehicle passage, corresponding one-to-one with the lateral positions in the bridge coordinate system, and are used to determine the spatial position of vehicle loads.

[0037] Optionally, step S2 specifically includes:

[0038] Based on axle load and axle passage time, a set of axle pulse event anchor points is extracted. A physical constraint mask is constructed by combining the sensor spatial coordinates, bridge span and support position with the influence line and propagation delay. The propagation delay is calculated based on the vehicle speed and the distance between the sensor and the bridge coordinate system to limit the time window of attention association.

[0039] Encode the sensor spatial coordinates and bridge span into geometric position codes to constrain the spatial association of cross-modal attention;

[0040] Sensor reliability weights are calculated based on the signal-to-noise ratio and missing rate of strain time series data, acceleration time series data, and displacement time series data, and are used to deweight low-reliability data in cross-modal attention.

[0041] Under the constraints of physical constraint masking and geometric position encoding, cross-modal attention time alignment is performed on strain time series data, acceleration time series data, displacement time series data, temperature time series data and wind speed time series data to generate aligned multimodal time series features;

[0042] Calculate the parameter values ​​of the dynamic equilibrium equation for the beam or slab based on the bridge span, support location, and material parameters;

[0043] Output the set of axle pulse event anchor points, physical constraint mask, geometric position encoding, sensor reliability weights, aligned multimodal time series characteristics, and dynamic balance equation parameter values.

[0044] Terminology definition:

[0045] The axle pulse event anchor point set is a discrete event set determined by axle load and axle passage time, including the event timestamp of each axle and its associated lane, speed, and other identifiers, which are used as reference points for time alignment and physical association.

[0046] The influence line is the response distribution function or curve generated by a unit moving load under a given bridge structure and measuring point location, used to characterize the spatial relationship between the load location and the sensor response.

[0047] The propagation delay is the time difference between the load application point and the sensor position caused by the vehicle's movement, calculated based on the vehicle speed and the geometric distance between the two in the bridge coordinate system, and is used to limit the time correlation range.

[0048] The attention association time window is an allowable time deviation range set around the arrival time of the event anchor point, used to retain only physically reachable time associations in attention calculation;

[0049] The physical constraint mask is a constraint matrix or tensor constructed based on influence lines and propagation delay. It is used to mask or reduce the weight of attention associations that do not meet the space-time physical reachability requirement, in order to ensure the physical consistency of cross-modal alignment.

[0050] The geometric position encoding is a method of mapping sensor spatial coordinates with geometric information such as bridge span and support position into vectors, which is used to inject spatial topology and relative positional relationships into attention calculation;

[0051] The cross-modal attention is an attention mechanism that calculates the correlation strength and performs weighted aggregation between different data modes (strain, acceleration, displacement, temperature, wind speed) to fuse and align multi-source time-series data;

[0052] The cross-modal attention time alignment is a process of unifying the time index and correcting the phase of multimodal time series data based on event anchors under the constraints of physical constraint mask and geometric position encoding, and outputting the aligned feature sequence.

[0053] The sensor reliability weight is a set of weight coefficients calculated based on the signal-to-noise ratio and missing rate of each response data, which is used to reduce the impact of low-reliability sensor data in attention and subsequent solution.

[0054] The signal-to-noise ratio (SNR) is an indicator characterizing the quality of a measured signal, defined as the ratio of effective signal power or amplitude to noise power or amplitude, and is used to assess data reliability.

[0055] The missing rate is the proportion of missing data samples to the total number of samples within a given time period, used to quantify data integrity and participate in reliability weight calculation;

[0056] The aligned multimodal temporal features are feature sequences or matrices obtained under a unified time base and spatial constraints, organized by time index and containing alignment information of each mode, for subsequent physical solutions;

[0057] The parameter values ​​of the dynamic equilibrium equation are a set of parameters used for the dynamic equilibrium equation of a beam or plate, including but not limited to linear density or surface density, stiffness parameters (such as EI or plate stiffness D), damping parameters and boundary condition correlation coefficients, which are used to support the calculation of equation residuals and physical solutions.

[0058] The dynamic equilibrium equations of the beams or slabs are a collective term for the partial differential equations and their boundary conditions that describe the spatiotemporal response of the bridge structure under external loads, and are used as the basis for physical constraints and parameter calculations.

[0059] Optionally, step S3 specifically includes:

[0060] Based on the aligned multimodal temporal characteristics, the set of axle pulse event anchor points, the sensor reliability weights and the parameter values ​​of the dynamic balance equation, a physical constraint neural operator model is established by combining the bridge span, support location and material parameters. A Green's function kernel is introduced into the kernel function of the model to characterize the propagation characteristics from load to response.

[0061] The equivalent boundary flexibility is set as a time-varying latent variable of boundary flexibility and jointly estimated with the model parameters, so that the latent variable of boundary flexibility reflects the time-varying characteristics of the boundary conditions.

[0062] The symplectic integral method is used for time progression to ensure the energy stability of the dynamic response;

[0063] Based on the sensor reliability weight, an observability weighted loss is constructed, with higher weights assigned to key sections and observable areas. Equation residuals determined by the parameters of the dynamic equilibrium equation are added to the loss to constrain the model to satisfy the dynamic equilibrium equation and boundary conditions of the beam or plate.

[0064] The physical constraint neural operator model is trained or updated by utilizing the aligned multimodal temporal features and limiting the time range of the training samples by the set of axle pulse event anchor points.

[0065] Output the trained physical constraint neural operator model and the updated boundary compliance latent variables.

[0066] Terminology definition:

[0067] The physical constraint neural operator model is a model that takes continuous operator learning as its core and explicitly introduces structural dynamics equations and boundary condition constraints. It maps external loads into displacement / strain responses through kernel integration or convolution, and uses physical residuals as training constraints.

[0068] The kernel function is the parameterized form of the integral kernel or convolution kernel used to describe the mapping relationship between the input and output in the neural operator, and it is the basic component for operator learning.

[0069] The Green's function kernel is a kernel composed of Green's functions corresponding to the selected beam / slab dynamic equilibrium equations under linear or linearized conditions, used to characterize the effect of a unit impulse load propagating to the structural response under given boundary conditions.

[0070] The equivalent boundary compliance is the displacement and rotation constraint compliance equivalently characterized by parameterizable elastic / damping elements at the support, reflecting the actual flexibility of the boundary constraints under non-ideal support conditions.

[0071] The boundary compliance latent variables are a set of equivalent boundary compliance parameters that change over time in the model and cannot be directly observed. They are identified along with the model parameters during training to characterize the time-varying nature of the boundary conditions.

[0072] The symplectic integral method is a numerical time integral algorithm that preserves the symplectic structure of the Hamiltonian system and is used to maintain energy stability in long-time-domain dynamic calculations (such as Stormer–Verlet-type algorithms).

[0073] The time advancement is a process of iteratively updating the model state according to a selected time step to obtain the numerical process of the response evolving over time, combined with symplectic integrals to improve long-term stability;

[0074] The observability weighted loss is a loss function that assigns weights to each spatiotemporal sample based on sensor reliability weights and sensor coverage, so that the errors of key sections and observable areas have a higher weight in training.

[0075] The key sections are a set of structural sections used for load verification and risk assessment, preferably including typical locations such as mid-span, near supports, and design control sections;

[0076] The observable region is a structural space subdomain determined by the sensor deployment location and measurement direction. Within this subdomain, the response can be observed directly or through an identifiable mapping.

[0077] The residual term of the dynamic equilibrium equation is a measure of the remaining error obtained by substituting the displacement / strain predicted by the model and its derivatives into the dynamic equilibrium equation of the beam / plate and the corresponding boundary conditions. It is used as a penalty constraint for physical consistency.

[0078] The boundary conditions are the displacement and rotation constraint relationships of the structure at the supports, preferably represented by elastic or viscoelastic supports, and are parameterized and associated with the boundary compliance latent variables.

[0079] The joint estimation of model parameters is a process of simultaneously identifying neural operator parameters and boundary compliance latent variables during training, in order to jointly minimize the observability weighted loss and the residual terms of the dynamic equilibrium equation.

[0080] Optionally, step S4 specifically includes:

[0081] Based on the trained physical constraint neural operator model and the updated boundary flexibility latent variables, the solution is obtained by combining the aligned multimodal temporal characteristics and dynamic equilibrium equation parameter values. During the solution process, the boundary conditions are corrected with the updated boundary flexibility latent variables and updated over time to obtain the spatiotemporal response field of the full bridge displacement and strain.

[0082] The physical residuals are calculated based on the parameter values ​​of the dynamic equilibrium equations and the spatiotemporal response field to characterize the degree to which the model satisfies the dynamic equilibrium equations and boundary conditions;

[0083] The internal force effects such as bending moment, shear force and axial force of key sections are calculated based on structural geometry and material parameters and spatiotemporal response field. The bearing capacity limit of key sections is determined based on structural geometry and material parameters. The bearing capacity utilization coefficient sequence is obtained by forming a ratio between the internal force effects and the bearing capacity limit.

[0084] Output the spatiotemporal response field, physical residuals, and load-bearing utilization coefficient sequence of the full bridge displacement and strain.

[0085] Terminology definition:

[0086] The spatiotemporal response field of the full bridge displacement and strain is a set of displacement and strain functions defined in the bridge coordinate domain and time domain. The response distribution of the full bridge over time is represented by a continuous or discrete grid and used for subsequent internal force and physical residual calculations.

[0087] The physical residual is a measure of the remaining amount of the equations obtained by substituting the spatiotemporal response field into the selected dynamic equilibrium equations of the beam or plate and their boundary conditions. It includes the volume equation residuals and the boundary condition residuals and is used to measure the physical consistency of the model.

[0088] The structural geometric parameters are a set of bridge geometric features required for calculating internal force effects and bearing limits, including span length, main beam spacing, bridge deck or slab thickness, cross-sectional geometric dimensions (such as flange width and thickness, web thickness, and box girder dimensions), crossbeam arrangement, and support geometric positions.

[0089] The internal force effect is a set of mechanical effect quantities calculated at the critical section by combining the spatiotemporal response field with structural and material parameters, including bending moment, shear force and axial force, and may include torque as needed;

[0090] The bearing capacity limit is the upper limit of the allowable bearing capacity of the critical section under given material parameters, structural geometry and applicable design specifications. It includes at least the design or evaluation values ​​of bending moment bearing capacity, shear bearing capacity and axial force bearing capacity, and takes into account the corresponding safety factor and combination factor.

[0091] The load-bearing utilization coefficient sequence is a sequence of demand-capacity ratios arranged by time index, corresponding to the ratio of the internal force effect of the key section at each time point to its load-bearing limit, used to quantify the load-bearing occupancy level and subsequent risk assessment.

[0092] Optionally, step S5 specifically includes:

[0093] Based on the load utilization coefficient sequence, combined with the axle pulse event anchor point set and axle load, vehicle speed and lane to extract traffic flow condition labels, and based on the axle pulse event anchor point set to determine the time block boundary, the load utilization coefficient sequence is divided into time blocks for subsequent interval estimation;

[0094] Block conformal prediction with load condition weighting is adopted. The load utilization coefficient sequence within each time block is estimated to generate an uncalibrated confidence interval. The load condition weighting is set according to the frequency of traffic flow load condition labels and axle load intensity distribution to make the coverage of different load conditions tend to be consistent.

[0095] Based on the axle load strength or the total load of the queue, a monotonically non-decreasing constraint is applied to the uncalibrated confidence interval for order-preserving calibration. During the calibration process, the upper and lower boundaries of the interval remain monotonically non-decreasing in the working condition sequence to conform to the physical law that the load utilization factor does not decrease with the increase of load strength, and finally the calibrated confidence interval is obtained.

[0096] Output the calibrated confidence interval.

[0097] Terminology definition:

[0098] The traffic flow condition label is a set of discretized identifiers for traffic flow status, including lane, speed level, axle load intensity level, vehicle platooning / queueing status and wheelbase characteristics, etc. It is extracted from the set of axle pulse event anchor points and axle load, vehicle speed and lane, and is used as the basis for weighting and calibration sorting.

[0099] The time block boundary is determined by the start and end times or event index of the block division based on the set of axle pulse event anchor points, and is used to segment the load utilization coefficient sequence for intra-block interval estimation.

[0100] The time block is a time interval enclosed by the boundaries of adjacent time blocks, used to perform conformal estimation and calibration of the load utilization coefficient sequence within the interval;

[0101] The block conformal prediction weighted by working condition is a prediction process that groups the load utilization coefficient sequence into time blocks, sets weights according to traffic flow working condition labels, and uses a conformal method to perform interval estimation within each time block, so that the coverage of different working conditions tends to be consistent.

[0102] The load condition weighting is a weight vector or function set according to the frequency of occurrence of load condition labels and the distribution of axle load intensity, which is used to weight different load condition samples in conformal estimation and coverage balancing.

[0103] The interval estimation is the process of estimating the range of values ​​of the target quantity at a given confidence level, and outputting a prediction interval composed of upper and lower bounds.

[0104] The uncalibrated confidence intervals are the set of prediction intervals directly obtained by block conformal prediction weighted according to operating conditions, without the application of monotonic constraints.

[0105] The order-preserving calibration is a calibration process that applies a monotonically non-decreasing constraint to an uncalibrated confidence interval on a load condition sequence and readjusts the upper and lower bounds of the interval so that the interval is consistent with the physical monotonicity of the load intensity.

[0106] The monotonically non-decreasing constraint is a constraint condition that requires the upper and lower boundaries of the interval not to decrease as the axle load intensity or the total load of the queue increases, in order to ensure the physical rationality of the load-bearing utilization factor.

[0107] The axle load strength is a measure of the vertical load strength of a single axle pair structure, preferably expressed as the nominal value or statistical quantile / average value of the axle load, and is used as one of the sorting criteria for order maintenance calibration;

[0108] The total load of the queue is the cumulative value of the vertical load of all axles within a time block or vehicle queue, and is used as one of the sorting criteria for order preservation calibration.

[0109] The working condition sequence is a working condition index sequence obtained by sorting according to axle load intensity or total queue load, etc., and is used to apply monotonically non-decreasing constraints and perform order-preserving calibration.

[0110] The upper and lower bounds of the interval are the lower and upper limits of the confidence interval, which are used to define the prediction range and participate in monotonic constraints.

[0111] The calibrated confidence interval is a set of prediction intervals that satisfy the monotonically non-decreasing constraint after order-preserving calibration, and is used for risk assessment and subsequent decision-making.

[0112] The coverage rate is the proportion of true values ​​falling within the confidence interval in a selected sample set or time block. It is used to measure the reliability of the interval estimation and as an evaluation index for load weighting and equilibrium objectives.

[0113] Optionally, step S6 specifically includes:

[0114] The coverage rate is calculated based on the calibrated confidence interval and compared with the preset coverage target. The constraint adjustment parameters are obtained according to the preset coupling rule, combining the magnitude and trend of the physical residual. The constraint adjustment parameters include the physical residual loss weight adjustment amount and the physical constraint mask threshold adjustment amount.

[0115] The physical residual loss weight in the physical constraint neural operator model and the physical constraint mask threshold in cross-modal attention are updated based on the constraint adjustment parameters. This ensures that physical constraints are strengthened in the next evaluation cycle when the coverage is lower than the preset coverage target or the physical residual increases, and physical constraints are appropriately relaxed when the coverage is higher than the preset coverage target and the physical residual decreases.

[0116] The risk level is calculated based on the load utilization coefficient sequence and the calibrated confidence interval, and the recommended load limit parameters are generated based on the risk level and the upper bound of the calibrated confidence interval.

[0117] Output constraint adjustment parameters, updated physical residual loss weights, updated physical constraint mask thresholds, risk levels, and load limit recommendation parameters.

[0118] Terminology definition:

[0119] The preset coverage target is a target coverage ratio setting value used to evaluate the performance of interval estimation. It is defined on a specified set of working conditions or the overall sample and serves as the control benchmark for closed-loop parameter tuning.

[0120] The coupling rule is a rule for determining constraint adjustment parameters by jointly mapping the comparison result of coverage rate with preset coverage rate target and the magnitude and trend of physical residual. It includes threshold criteria or continuous gain function, which are used to clarify the conditions and magnitude of increasing or decreasing constraint strength.

[0121] The constraint adjustment parameters are a set of parameters used to update the model constraint strength, including at least the physical residual loss weight adjustment amount and the physical constraint mask threshold adjustment amount;

[0122] The physical residual loss weight is a weighting coefficient assigned to the dynamic equilibrium equation and boundary condition residual terms during training or updating, used to control the proportion of physical consistency constraints in the overall loss.

[0123] The physical constraint mask threshold is a threshold or parameter used to determine physically reachable associations in cross-modal attention, and is used to adjust the severity of the physical constraint mask (such as the width of the time window or the lower limit of the spatial association strength).

[0124] The physical residual loss weight adjustment amount is a numerical change or proportional coefficient used to increase or decrease the physical residual loss weight, calculated according to the coupling rule.

[0125] The physical constraint mask threshold adjustment amount is a change amount used to tighten or relax the physical constraint mask threshold, calculated based on the coupling rules.

[0126] The evaluation cycle is the time cycle unit for this method to perform one complete data acquisition, calculation, interval estimation and closed-loop parameter tuning, and is set to hour, day or event batch according to the monitoring and processing window.

[0127] The self-correction is a closed-loop parameter tuning mechanism that automatically updates the physical residual loss weight and physical constraint mask threshold based on the constraint adjustment parameters during adjacent evaluation weeks, so that the coverage rate approaches the preset coverage rate target and suppresses physical residuals.

[0128] The risk level is the structural bearing risk classification result obtained by assessing the bearing utilization coefficient sequence and the calibrated confidence interval, and is preferably represented by a multi-level discrete level for control decision-making.

[0129] The recommended load limit parameters are a set of traffic load control parameters generated based on the risk level and the upper limit of the calibrated confidence interval, including but not limited to single axle load limit, total vehicle weight limit, queue length or vehicle spacing limit, and lane restriction instructions.

[0130] The beneficial effects of this invention are:

[0131] 1. Cross-modal attention alignment guided by event anchors, combined with physical constraint masks of influence lines and propagation delays, geometric position coding and sensor reliability weighting, effectively suppresses mismatch and noise effects under conditions of sensor sparsity, uneven signal-to-noise ratio and missing data, significantly improving the accuracy of spatiotemporal response reconstruction and critical section load assessment.

[0132] 2. By introducing a Green's function kernel and time-varying boundary compliance latent variables into the physical constraint neural operator, and combining them with symplectic integral propagation, observability weighted loss and dynamic equilibrium equation residual constraints, the physical consistency and long-term energy stability of the dynamic response are guaranteed. This can accurately characterize the time-varying characteristics of the support conditions and significantly reduce physical residuals, thereby improving the accuracy of the calculation of the full-bridge response and internal force effects.

[0133] 3. We adopt block shape-preserving interval estimation and order-preserving calibration with weighted blocks according to working conditions, and use a self-correcting closed loop coupled with coverage and physical residuals to adjust parameters, so that the coverage under different traffic flow conditions tends to be consistent and the interval is monotonic and reasonable. In long-term operation, the coverage approaches the preset target and suppresses physical residuals, thereby improving the reliability of risk classification and load limit recommendations. Attached Figure Description

[0134] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0135] Figure 1 This is a flowchart of a dynamic assessment method for bridge bearing capacity based on multi-source monitoring data proposed in this invention. Detailed Implementation

[0136] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0137] refer to Figure 1 A dynamic assessment method for bridge bearing capacity based on multi-source monitoring data includes:

[0138] S1. Collect time-series data of strain, acceleration, displacement, temperature, and wind speed, axle load and passage time, vehicle speed, lane, sensor spatial coordinates, bridge span, support position, and material parameters, and establish a bridge coordinate system and a unified time reference.

[0139] S2. Based on axle load and passage time, extract axle pulse event anchor points. Combine sensor coordinates, span and support position, construct a physical constraint mask according to the influence line and propagation delay and generate geometric position code. Obtain sensor reliability weights based on the signal-to-noise ratio and missing rate of response data. Perform cross-modal attention alignment on multi-source data under the mask and coding constraints and obtain dynamic balance equation parameter values.

[0140] S3. Establish a physical constraint neural operator model, set the equivalent boundary compliance as a time-varying boundary compliance latent variable and jointly estimate it with the model parameters, construct a loss containing observable weighting terms and dynamic equilibrium equation residual terms, and align features and event anchors to train or update the model.

[0141] S4. Based on the trained model and boundary flexibility latent variables, combine the alignment features and dynamic equation parameters to solve the spatiotemporal response field of the full bridge displacement and strain and calculate the physical residuals. Calculate the internal force effect of key sections based on structural geometry and material parameters and the spatiotemporal response field and form a bearing capacity utilization coefficient sequence.

[0142] S5. Using the load utilization coefficient sequence, combined with event anchor points and axle load, vehicle speed, and lane, extract traffic flow condition labels and divide time blocks according to event anchor points. Perform block shape-preserving interval estimation according to condition weighting to obtain uncalibrated confidence intervals. Apply monotonically non-decreasing constraints based on axle load strength or total queue load to perform order-preserving calibration and obtain calibrated confidence intervals.

[0143] S6. Calculate the coverage rate and compare it with the preset target. Combine the physical residuals with the coupling rules to obtain the constraint adjustment parameters. Based on this, update the physical residual loss weight and the physical constraint mask threshold of cross-modal attention in the physical constraint neural operator model for self-correction in the next evaluation cycle. Calculate the risk level and generate load limit recommendation parameters based on the load utilization coefficient sequence and calibration confidence interval.

[0144] In this specific embodiment, S1 specifically refers to:

[0145] Establish a bridge coordinate system and a unified time reference. Determine the origin and axis of the coordinate system based on the bridge span and support location. The bridge coordinate system is denoted as... The origin is The bridge's axial, transverse, and vertical axes are respectively denoted as... and And aligned with the centerline of the span direction and support position, the sensors are geometrically calibrated based on this, and the spatial coordinates of each sensor are aligned with the centerline of the support position. The installation location and measurement direction were marked, and their consistency with the bridge span and support location was checked. Subsequently, strain time-series data, acceleration time-series data, and displacement time-series data were collected, and the corresponding timestamps and sampling frequencies were recorded. Simultaneously, axle load, axle passage time, vehicle speed, and lane location were collected, and the corresponding timestamps were recorded. Then, temperature time-series data and wind speed time-series data were collected, and the corresponding timestamps and sampling frequencies were recorded. Finally, the sensor's spatial coordinates, bridge span, support location, and material parameters were collected. The material parameters were... Let be the basic set, representing the elastic modulus, shear modulus, density, Poisson's ratio, cross-sectional area, and moment of inertia, respectively. The units and dimensions of these units are checked for consistency. To achieve time alignment with a unified time base, the following timestamp mapping formula is used to normalize the records for each mode:

[0146] ;

[0147] in, To unify the time base The first sensor The timestamp of each sample The start time for a unified time base is set by the clock of the data acquisition device. A positive integer representing the sample index. For the first The sampling frequency of each sensor, For the first The clock offset of each sensor relative to a unified time base is used to correct for time drift and trigger delay between devices. It provides a cross-modal sensor index, covering sensor types such as strain, acceleration, displacement, temperature, and wind speed;

[0148] Output strain time series data, acceleration time series data, displacement time series data, axle load, axle passage time, vehicle speed, lane, temperature time series data, wind speed time series data, sensor spatial coordinates, bridge span, support location, and material parameters.

[0149] In this specific embodiment, S2 specifically refers to:

[0150] Based on axle load and axle passage time, extract the set of axle pulse event anchor points under a unified time reference. For each event anchor point Record its event timestamp Lane markings Vehicle speed Bridge axis coordinates at the event reference location and the axle load of the vehicle and in the bridge coordinate system China-Israel Sensor Index Label the position vector of each sensor Based on the measurement modes, and then combined with the bridge span and support location, a physical constraint mask is constructed based on the influence line and propagation delay. In the time domain, the following gating function is used to limit the physical reachability associations of the attention:

[0151] ;

[0152] in This is the time-gated function used to construct the physical constraint mask. For time variables under a unified time base, The reliability weight of the sensor is determined by the sensor's signal-to-noise ratio. With missing rate Mapped to obtain, For the first Event timestamps for each axle The propagation delay is defined as the geometric distance along the bridge axis from the event reference location to the sensor. With vehicle speed The ratio, To allow for half the width of the time window, based on the sensor sampling frequency. Speed ​​uncertainty and synchronization error settings The influence line response factor represents the effect of a unit moving load on the lane. Position action at the sensor The response amplitude or weight, The threshold used to mask invalid spatial associations. This is an indicator function that takes the value 1 when the condition inside the parentheses is met, and 0 otherwise;

[0153] In the spatial domain, the sensor spatial coordinates and bridge span are encoded into geometric position codes. This encoding injects the bridge axis and support topology relationship and associates it with the lane lateral position in cross-modal attention;

[0154] Based on the signal-to-noise ratio of strain, acceleration, and displacement With missing rate Calculate sensor reliability weights This is used to deweight low-confidence data in cross-modal attention;

[0155] In physical constraint mask With geometric position encoding set Under constraints, cross-modal attention time alignment is performed on time series data of strain, acceleration, displacement, temperature, and wind speed to generate aligned multimodal time series features. Simultaneously, based on the bridge span, support location, and material parameter set... Calculate the parameter values ​​of the dynamic equilibrium equations for beams or slabs. The Including linear density Bending stiffness Damping parameters Correlation coefficient vector with boundary conditions ,in From density With cross-sectional area Sure, With elastic modulus and moment of inertia Related For damping coefficient, Determined by the location and type of support;

[0156] Output axle pulse event anchor point set Physical constraint mask Geometric position encoding set Sensor reliability weight set Aligned multimodal temporal features and the parameter values ​​of the dynamic balance equation .

[0157] In this specific embodiment, S3 specifically refers to:

[0158] Based on the aligned multimodal temporal features Axle pulse event anchor point set Sensor reliability weight set and the parameter values ​​of the dynamic balance equation A physically constrained neural operator model is established by combining the bridge span, support location, and material parameters. A Green's function kernel is introduced into the kernel function of this model, along with a kernel parameter set. Parameterization is performed to characterize the load-to-response propagation characteristics, while the equivalent boundary compliance is set as a time-varying latent variable of boundary compliance. And estimate it together with the model parameters, so that It can reflect the time-varying characteristics of support conditions. The time progression adopts the symplectic integral method to ensure long-term energy stability. The time step is set according to a unified time reference and stability conditions. The training sample is selected based on the set of axle pulse event anchor points. By limiting the time range and combining gating rules with physical constraint masks to filter observable regions and key sections, a loss function containing observability weighting terms and dynamic equilibrium equation residual terms is constructed as follows:

[0159] ;

[0160] in, This is a scalar value of the total loss function, used to measure the sum of the model fitting error and the physical residuals. This is a summation operator for the training sample index set. The training sample index set, indexed by event anchors. With sensor index Composed of and composed of the set of axle pulse event anchor points And obtained through physical constraint mask gating, The sensor reliability weights are derived from the first... The signal-to-noise ratio of each sensor is mapped to the missing rate and used for observability weighting. For the event anchor point Nearby by the The observation response values ​​obtained from each sensor can be used to calculate the data from a unified time base. Extracting scalar or component values ​​of displacement or strain from the data. For the operator mapping of the physically constrained neural operator model, the Green's function kernel parameter set is used. Enter the kernel parameters Parameter values ​​of dynamic equilibrium equations With boundary flexibility latent variables Output the predicted response, where the subscripts are... Indicates the index of the event anchor point With sensor index The corresponding predicted component values, The physical residual loss weight is used to adjust the proportion of dynamic equilibrium equation constraints in the total loss. The residual operator is used to measure the degree to which the volume equations and boundary conditions are satisfied, representing the dynamic equilibrium equations and boundary conditions. The set of parameter values ​​for the dynamic equilibrium equations includes linear density. Bending stiffness Damping parameters and the boundary condition coefficient vector related to the support location, For time The varying boundary compliance latent variable vector is used to characterize the flexibility and viscous properties of non-ideal supports. The squared L2 norm is used to measure the error energy, and the set of Green's function kernel parameters is jointly identified during the minimization of the aforementioned loss function. With boundary flexibility latent variables And update the parameters of the physically constrained neural operator model;

[0161] Output the trained physical constraint neural operator model and the updated boundary compliance latent variables.

[0162] In this specific embodiment, S4 specifically refers to:

[0163] Based on the trained physical constraint neural operator model With the updated boundary compliance latent variables Combined with aligned multimodal temporal features With the parameter values ​​of the dynamic equilibrium equation In the bridge coordinate system The solution is performed using a unified time base, and the time progression process is then carried out using... By modifying and updating the boundary conditions over time, the spatiotemporal response field of the full bridge displacement is obtained. With strain spatiotemporal response field ,in Bridge coordinate system Spatial coordinates below For time variables under a unified time base;

[0164] Subsequently based on The residuals of the above-mentioned spatiotemporal response field computational volume equations and boundary condition residuals are summarized into a physical residual scalar, which is used to characterize... The degree to which the dynamic equilibrium equations and boundary conditions are satisfied; in terms of structural geometry and material parameter sets. With the support of and For input in the key section set The effects of internal forces such as bending moment, shear force, and axial force are calculated above and denoted as follows: and ,in Index for key sections;

[0165] Based on the structural geometry and material parameters, the bearing capacity limits for each critical section are determined and denoted as bending moment bearing capacity limits. Shear bearing capacity limit With axial force bearing limit Based on this, a carrying capacity utilization coefficient sequence is formed, which is in the th... Key cross-section moments The value at is defined as:

[0166] ;

[0167] in For carrying capacity utilization factor, For the maximum value operator, This is an absolute value operator used to ensure that the internal force effect participates in the ratio calculation in terms of amplitude;

[0168] The output includes the spatiotemporal response field of the full-bridge displacement and strain. and Physical residual scalar and bearing capacity utilization coefficient sequence .

[0169] In this specific embodiment, S5 specifically includes:

[0170] Based on the carrying capacity utilization coefficient sequence And combined with the set of axle pulse event anchor points and axle load, vehicle speed and lane extraction traffic flow condition label set For each time sample, a working condition label is assigned and recorded according to information such as its lane, speed range, and axle load intensity range. Then, the time block set is determined using the event anchor point as the boundary. Each time block Enclosed by adjacent event anchor points, this segment is used for interval estimation. Before interval estimation, weighting rules are set for the working condition labels to balance the coverage of different working conditions, and a working condition weight mapping is defined. And set sample weights ,in The frequency of load conditions and the distribution of axle load intensity are jointly used to reduce the risk of under-coverage of rare and high-load load conditions. Subsequently, for each critical section... With time blocks Collect a sample set of carrying capacity utilization coefficients within this block. At confidence level The following uses block conformal interval estimation with load condition weighting to generate uncalibrated confidence intervals:

[0171] ;

[0172] in, To be at the critical section With time blocks The uncalibrated confidence interval on This is a weighted quantile operator used in weight sets. Calculation of quantile levels under action empirical quantiles, This is a sample set of the carrying capacity utilization coefficients within this block. The confidence level corresponding to the coverage target. and These are the quantile levels of the lower and upper bounds of the interval, respectively.

[0173] To satisfy the physical law that the load utilization factor monotonically does not decrease with load intensity, the load ranking values ​​for each time block are selected, including representative values ​​of axle load intensity. (Given by statistics of axle loads within this block, such as mean or quantiles) and total load of the queue. (Obtained by summing the loads of all axles within this block), in accordance with or The sorted operating condition sequence is subjected to a sequence-preserving calibration method. The upper and lower bounds are monotonically adjusted without decreasing to form Thus, a calibration confidence interval is obtained that does not decrease with load intensity or total queue load;

[0174] Output the set of calibrated confidence intervals .

[0175] In this specific embodiment, S6 specifically refers to:

[0176] Based on the calibrated set of confidence intervals Calculate empirical coverage and with preset coverage target In comparison, among them To select a set of key sections With time block set The proportion of the actual carrying capacity utilization coefficient sample falling into the calibration confidence interval The coverage target to be achieved by this method is denoted as and the physical residual scalar obtained from step S4 is denoted as . And define the physical residual scalar of the previous evaluation period as The residual baseline is Based on this, the trend of physical residual change can be obtained. In the coupling rules, the constraint adjustment parameters are jointly determined by coverage deviation, residual deviation, and residual trend. The adjustment amount is obtained by linear mapping using the gain coefficient, and is specifically calculated using the following formula:

[0177] ;

[0178] in, The adjustment amount for the weight of physical residual loss. The adjustment amount for the physical constraint mask threshold, The physical residual loss weights of the loss function in step S3, The spatial correlation threshold of the physical constraint mask in step S2, To reduce coverage deviation residual deviation With residual trend Mapped to Gain coefficient, To map the same three items to The gain coefficient and matrix multiplication are used to simultaneously generate two types of adjustment values, ensuring that constraints are strengthened when coverage is lower than the target or physical residuals increase, while constraints are appropriately relaxed when coverage is higher than the target and physical residuals decrease. Then, based on... and renew and The updated parameters are then bounded with upper and lower limits to ensure numerical stability and operability. After self-calibration, the parameter sequence is utilized. With the calibrated confidence interval Risk assessment is conducted based on the upper bound of the load capacity, and the load utilization of each key section and time block is mapped to discrete risk levels to generate load limit recommendation parameters, including single axle load limit, total vehicle weight limit, queue length or vehicle spacing limit and lane restriction instructions.

[0179] Final output constraint adjustment parameters Updated physical residual loss weights Updated physical constraint mask threshold Risk level and recommended load limits.

[0180] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0181] This invention constructs an assessment chain around traffic events and physical constraints. First, it uses axle pulse event anchor points combined with influence lines and propagation delays to form a physical constraint mask. This mask is then combined with geometric position coding and sensor reliability weights to achieve precise temporal and spatial alignment of multi-source monitoring data. Based on this, a physical constraint neural operator introduces a Green's function kernel and jointly estimates time-varying boundary flexibility latent variables. A symplectic integral is used for energy conservation over time, combined with observability weighting and dynamic balance residual constraints, to obtain the spatiotemporal response of displacement and strain, as well as the load-bearing utilization coefficient of key sections, satisfying the physical equations. Subsequently, through block conformal interval estimation and order-preserving calibration weighted according to operating conditions, the coverage rate under different operating conditions tends to be consistent, and the confidence interval remains monotonically constant with load intensity. Finally, a closed-loop self-calibration coupling coverage rate and physical residuals achieves continuous parameter tuning, approaching the coverage target and reducing physical residuals during long-term operation, outputting robust risk levels and load limit recommendations. This effectively solves the problem of ensuring the accuracy of load assessment under sparse sensor coverage conditions.

[0182] This invention makes targeted improvements to the algorithm structure and forms a verifiable closed-loop mechanism: It limits the attention window with event anchors, significantly reducing the risk of mismatch in global alignment; it suppresses invalid associations and embeds spatial topological information by using physical constraint masks and geometric position encodings composed of influence lines and propagation delays; it introduces boundary flexibility latent variables to characterize the time-varying nature of support conditions, avoiding model bias caused by constant boundary assumptions; it uses symplectic integrals to ensure long-term energy stability and improve the reliability of dynamic solutions; it incorporates observability weighting and dynamic balance residuals into the training loss, prioritizing the learning of key sections and constrained regions; in the uncertainty assessment stage, it employs block shape-preserving and order-preserving calibration weighted according to working conditions, ensuring that interval estimation simultaneously satisfies the coverage target and the physical law of load monotonicity; and it unifies statistical reliability and physical consistency through closed-loop parameter tuning of confidence and physical residuals. These structural improvements synergistically work on key aspects of alignment, solution, and interval estimation, significantly improving the accuracy and stability of load assessment under complex working conditions and low signal-to-noise ratio conditions.

Claims

1. A dynamic assessment method for bridge bearing capacity based on multi-source monitoring data, characterized in that, include: S1. Collect time-series data of strain, acceleration, displacement, temperature, and wind speed, axle load and passage time, vehicle speed, lane, sensor spatial coordinates, bridge span, support position, and material parameters, and establish a bridge coordinate system and a unified time reference. S2. Based on axle load and passage time, extract axle pulse event anchor points. Combine sensor coordinates, span and support position, construct a physical constraint mask according to the influence line and propagation delay and generate geometric position code. Obtain sensor reliability weights based on the signal-to-noise ratio and missing rate of response data. Perform cross-modal attention alignment on multi-source data under the mask and coding constraints and obtain dynamic balance equation parameter values. S3. Based on the aligned multimodal temporal features, the set of axle pulse event anchor points, sensor reliability weights, and dynamic balance equation parameter values, a physical constraint neural operator model is established in conjunction with the bridge span, support location, and material parameters. A Green's function kernel is introduced into the kernel function of the model to characterize the propagation characteristics from load to response. The equivalent boundary flexibility is set as a boundary flexibility latent variable that changes with time and is jointly estimated with the parameters of the physical constraint neural operator model. An observability weighted loss is constructed based on the sensor reliability weights. High weights are assigned to key sections and observable areas. An equation residual term determined by the dynamic balance equation parameter values ​​is added to the loss to constrain the physical constraint neural operator model to satisfy the dynamic balance equation and boundary conditions of the beam or plate. The physical constraint neural operator model is trained or updated using the aligned multimodal temporal features and the time range of the training samples is limited by the set of axle pulse event anchor points. S4. Based on the trained physical constraint neural operator model and boundary flexibility latent variables, and combined with the aligned multimodal temporal characteristics and dynamic equilibrium equation parameter values, solve the spatiotemporal response field of the full bridge displacement and strain and calculate the physical residuals. Calculate the internal force effect of key sections based on structural geometry and material parameters and the spatiotemporal response field, and form a bearing capacity utilization coefficient sequence. S5. Using the load utilization coefficient sequence, combined with event anchor points and axle load, vehicle speed, and lane, extract traffic flow condition labels and divide time blocks according to event anchor points. Perform block shape-preserving interval estimation according to condition weighting to obtain uncalibrated confidence intervals. Apply monotonically non-decreasing constraints based on axle load strength or total queue load to perform order-preserving calibration and obtain calibrated confidence intervals. S6. Calculate the coverage rate and compare it with the preset target. Combine the physical residuals with the coupling rules to obtain the constraint adjustment parameters. Based on this, update the physical residual loss weight and the physical constraint mask threshold of cross-modal attention in the physical constraint neural operator model for self-correction in the next evaluation cycle. Calculate the risk level and generate load limit recommendation parameters based on the load utilization coefficient sequence and calibration confidence interval.

2. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S1 specifically refers to: Establish a bridge coordinate system and time reference. Determine the origin and axis of the bridge coordinate system based on the bridge span and support location. Set a unified time reference as the timestamp and sampling frequency for each time series data recording based on the clock of the acquisition equipment. Perform sensor geometric calibration, mark the installation position and orientation of each sensor in the bridge coordinate system according to the sensor spatial coordinates, and check the consistency with the bridge span and support position; Collect strain time-series data, acceleration time-series data, and displacement time-series data, and record the corresponding timestamps and sampling frequencies; Collect axle load, axle passage time, vehicle speed and lane information, and record the corresponding timestamps; Collect time-series temperature and wind speed data and record the corresponding timestamps and sampling frequencies; Collect sensor spatial coordinates, bridge span, support location, and material parameters, and verify the consistency of the units and dimensions of the material parameters; Output strain time series data, acceleration time series data, displacement time series data, axle load, axle passage time, vehicle speed, lane, temperature time series data, wind speed time series data, sensor spatial coordinates, bridge span, support location, and material parameters.

3. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S2 specifically refers to: Based on axle load and axle passage time, a set of axle pulse event anchor points is extracted. A physical constraint mask is constructed by combining the sensor spatial coordinates, bridge span and support position with the influence line and propagation delay. The propagation delay is calculated based on the vehicle speed and the distance between the sensor and the bridge coordinate system to limit the time window of attention association. Encode the sensor spatial coordinates and bridge span into geometric position codes to constrain the spatial association of cross-modal attention; Sensor reliability weights are calculated based on the signal-to-noise ratio and missing rate of strain time series data, acceleration time series data, and displacement time series data, and are used to deweight low-reliability data in cross-modal attention. Under the constraints of physical constraint masking and geometric position encoding, cross-modal attention time alignment is performed on strain time series data, acceleration time series data, displacement time series data, temperature time series data and wind speed time series data to generate aligned multimodal time series features; Calculate the parameter values ​​of the dynamic equilibrium equation for the beam or slab based on the bridge span, support location, and material parameters; Output the set of axle pulse event anchor points, physical constraint mask, geometric position encoding, sensor reliability weights, aligned multimodal time series characteristics, and dynamic balance equation parameter values.

4. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S3 specifically refers to: Based on the aligned multimodal temporal characteristics, the set of axle pulse event anchor points, the sensor reliability weights and the parameter values ​​of the dynamic balance equation, a physical constraint neural operator model is established by combining the bridge span, support location and material parameters. A Green's function kernel is introduced into the kernel function of the model to characterize the propagation characteristics from load to response. The equivalent boundary flexibility is set as a time-varying latent variable of boundary flexibility and jointly estimated with the parameters of the physical constraint neural operator model, so that the latent variable of boundary flexibility reflects the time-varying characteristics of the boundary conditions. The symplectic integral method is used for time progression to ensure the energy stability of the dynamic response; Based on the sensor reliability weight, an observability weighted loss is constructed, with higher weights assigned to key sections and observable areas. Equation residuals determined by the parameter values ​​of the dynamic equilibrium equation are added to the loss to constrain the physical constraint neural operator model to satisfy the dynamic equilibrium equation and boundary conditions of the beam or plate. The physical constraint neural operator model is trained or updated by utilizing the aligned multimodal temporal features and limiting the time range of the training samples by the set of axle pulse event anchor points. Output the trained physical constraint neural operator model and the updated boundary compliance latent variables.

5. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S4 specifically refers to: Based on the trained physical constraint neural operator model and the updated boundary flexibility latent variables, the solution is obtained by combining the aligned multimodal temporal characteristics and dynamic equilibrium equation parameter values. During the solution process, the boundary conditions are corrected with the updated boundary flexibility latent variables and updated over time to obtain the spatiotemporal response field of the full bridge displacement and strain. The physical residuals are calculated based on the parameter values ​​of the dynamic equilibrium equations and the spatiotemporal response field to characterize the degree to which the physical constraint neural operator model satisfies the dynamic equilibrium equations and boundary conditions. The internal force effects of key sections are calculated based on structural geometry and material parameters and spatiotemporal response field. The internal force effects include bending moment, shear force and axial force. The bearing capacity limit of key sections is determined based on structural geometry and material parameters. The bearing capacity utilization coefficient sequence is obtained by comparing the internal force effects with the bearing capacity limit. Output the spatiotemporal response field, physical residuals, and load-bearing utilization coefficient sequence of the full bridge displacement and strain.

6. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S5 specifically refers to: Based on the load utilization coefficient sequence, combined with the axle pulse event anchor point set and axle load, vehicle speed and lane to extract traffic flow condition labels, and based on the axle pulse event anchor point set to determine the time block boundary, the load utilization coefficient sequence is divided into time blocks for subsequent interval estimation; Block conformal prediction with load condition weighting is adopted. The load utilization coefficient sequence within each time block is estimated to generate an uncalibrated confidence interval. The load condition weighting is set according to the frequency of traffic flow load condition labels and axle load intensity distribution to make the coverage of different load conditions tend to be consistent. Based on the axle load strength or the total load of the queue, a monotonically non-decreasing constraint is applied to the uncalibrated confidence interval for order-preserving calibration. During the calibration process, the upper and lower boundaries of the interval remain monotonically non-decreasing in the working condition sequence to conform to the physical law that the load utilization factor does not decrease with the increase of load strength, and finally the calibrated confidence interval is obtained. Output the calibrated confidence interval.

7. The method for dynamic evaluation of bridge bearing capacity based on multi-source monitoring data as described in claim 1, characterized in that, S6 specifically refers to: The coverage rate is calculated based on the calibrated confidence interval and compared with the preset coverage target. The constraint adjustment parameters are obtained according to the preset coupling rule, combining the magnitude and trend of the physical residual. The constraint adjustment parameters include the physical residual loss weight adjustment amount and the physical constraint mask threshold adjustment amount. The physical residual loss weight in the physical constraint neural operator model and the physical constraint mask threshold in cross-modal attention are updated based on the constraint adjustment parameters. This ensures that physical constraints are strengthened in the next evaluation cycle when the coverage is lower than the preset coverage target or the physical residual increases, and physical constraints are appropriately relaxed when the coverage is higher than the preset coverage target and the physical residual decreases. The risk level is calculated based on the load utilization coefficient sequence and the calibrated confidence interval, and the recommended load limit parameters are generated based on the risk level and the upper bound of the calibrated confidence interval. Output constraint adjustment parameters, updated physical residual loss weights, updated physical constraint mask thresholds, risk levels, and load limit recommendation parameters.

Images