Physical information driven bridge multi-source data damage fusion diagnosis system and method

Abstract

This invention relates to the field of bridge structural health monitoring technology, and discloses a physical information-driven bridge multi-source data damage fusion diagnosis system and method, including: acquiring and decoupling bridge monitoring signals, extracting a low-frequency continuous field observation set and a high-frequency transient discrete event set; constructing a continuously differentiable cross-scale physical mapping functional, converting the high-frequency transient discrete event set into time-varying local stiffness boundary conditions in the structural dynamics partial differential equations; initializing a physical information neural network model, and constructing a joint loss function by combining the low-frequency continuous field observation set, the reconstructed time-varying dynamics partial differential equations, and the physical prior boundary conditions; performing a double-loop alternating optimization iteration, with the inner loop calculating and updating network weights, and the outer loop calculating and updating microscopic physical mechanism parameters; and outputting the full-field response reconstruction result, the structural local stiffness reduction matrix, and damage space coordinates when the convergence condition is met. This invention achieves deep fusion of multi-source data, improving the reliability of damage diagnosis.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring technology, specifically to a physical information-driven bridge multi-source data damage fusion diagnostic system and method. Background Technology

[0002] To diagnose the service condition of bridges, existing bridge structural health monitoring systems typically employ various types of sensors. Accelerometers and fiber Bragg grating strain gauges are primarily used to acquire low-frequency macroscopic continuous dynamic response data of the structure, while acoustic emission probes are used to capture high-frequency transient discrete strain energy signals generated during the propagation of microscopic cracks within the bridge. Because macroscopic continuous vibration and microscopic transient damage differ significantly in temporal resolution and spatial distribution scale, existing multi-source data fusion techniques often employ data feature stitching or statistical mapping matching, failing to effectively combine high-frequency transient discrete signals with the macroscopic continuous dynamic physical equations of the structure, thus hindering deep data fusion at the underlying mechanical evolution mechanism level.

[0003] In damage identification relying on monitoring data, conventional data-driven deep learning diagnostic models heavily depend on large amounts of labeled data. Due to the lack of constraints from solid mechanics principles within the model, these algorithms are prone to overfitting when faced with complex environmental background noise in real-world engineering scenarios. The damage diagnosis results output by such models often violate the physical boundary conditions of material evolution, leading to a lack of mechanical interpretability in the inverted stiffness parameters.

[0004] To incorporate engineering physical constraints, some studies have attempted to use physical information neural networks (PENs) for structural state assessment. Traditional PSNs rely on the automatic differentiation engine of deep learning frameworks when solving partial differential equations of structural dynamics. This requires that the boundary conditions and coefficient matrices of the equations possess high-order continuous differentiability within the spatiotemporal domain. When attempting to directly introduce transient and spatially concentrated local damage, represented by acoustic emission events, into continuous PSNs, the discontinuity at the spatiotemporal trigger points leads to singularities in numerical computation. These singularities can cause gradient explosion or vanishing during chain differentiation, preventing the neural network from training properly. Furthermore, in complex damaged bridge systems, the trainable weights of the neural network are deeply coupled with the parameters of unknown microscopic physical mechanisms, and conventional joint differentiation optimization can easily lead to gradient direction conflicts. The algorithm is prone to getting trapped in local minima in non-convex solution spaces, outputting parameter solutions that do not conform to the true mechanical boundaries. This results in the system being unable to accurately quantify the degree of local stiffness reduction and the three-dimensional coordinates of the damage space. Summary of the Invention

[0005] The technical problem solved by this invention is that, due to the difference in spatiotemporal scale, it is difficult to achieve deep integration of macroscopic low-frequency response data and microscopic high-frequency acoustic emission data obtained in bridge structural health monitoring at the physical mechanism level; at the same time, existing pure data-driven diagnostic models lack the constraints of physical and mechanical laws and are prone to overfitting in noisy environments, resulting in inaccurate quantitative inversion and spatial positioning of local structural damage.

[0006] To address the above problems, the present invention provides the following technical solution: The first aspect of this invention provides a physical information-driven bridge multi-source data damage fusion diagnostic system, comprising: The sensor acquisition module is deployed on the bridge structure under test to acquire structural monitoring signals under the service condition of the structure; Edge computing nodes are physically connected to the sensing and acquisition modules and are used to decouple structural monitoring signals based on the physical properties of the signals, and extract low-frequency continuous field observation sets and high-frequency transient discrete event sets. The cloud computing platform establishes data transmission links with edge computing nodes for: We construct a continuously differentiable cross-scale physical mapping functional, and use the cross-scale physical mapping functional to transform the high-frequency transient discrete event set into time-varying local stiffness boundary conditions in the partial differential equations of structural dynamics. Initialize the physical information neural network model, obtain the low-frequency continuous field observation set, the reconstructed time-varying dynamic partial differential equations, and the set physical prior boundary conditions, and combine them to construct the joint loss function of the physical information neural network model; The system performs a double-loop alternating optimization iteration. During the inner loop calculation, the network weights are updated through gradient descent, and during the outer loop calculation, the microscopic physical mechanism parameters are updated through residual backpropagation. When the convergence condition is met, the full-field response reconstruction result is output, and the converged microscopic physical mechanism parameters are used to calculate the local stiffness reduction matrix and damage space coordinates of the structure.

[0007] Furthermore, the sensing and acquisition module incorporates a heterogeneous sensor group and cross-scale clock synchronization hardware. The heterogeneous sensor group includes piezoelectric accelerometers, fiber Bragg grating strain gauges, and piezoelectric ceramic acoustic emission probes. Edge computing nodes allocate a unified global time reference to monitoring channels with different sampling frequencies through the cross-scale clock synchronization hardware. They perform time-domain interpolation alignment on the original acceleration time series and the original strain time series, carrying timestamps, and construct a low-frequency continuous field observation set by combining spatial deployment coordinates. The edge computing nodes calculate a dynamic floating threshold in real time based on ambient background noise. When the voltage signal output by the piezoelectric ceramic acoustic emission probe crosses the dynamic floating threshold, a transient time window is extracted. This is combined with the multi-channel trigger time difference to calculate the three-dimensional spatial coordinates of the micro-fracture event and the corresponding trigger time. Within the transient time window, integral calculations are performed to quantify the absolute energy, and a high-frequency transient discrete event set is constructed.

[0008] Furthermore, the cloud computing platform establishes a nonlinear mapping operator characterizing the amplitude of microscopic damage, which includes a vector of microscopic physical mechanism parameters to be inverted and calibrated. It constructs continuously differentiable spatial and temporal smoothing operators and introduces a nonlinear dissipation relaxation term characterizing the nonlinear dissipation process. The nonlinear mapping operator, spatial smoothing operator, temporal smoothing operator, and nonlinear dissipation relaxation term are then integrated to establish a cross-scale physical mapping functional for the time-varying local stiffness degradation of the output structure. The spatial smoothing operator is constructed based on an exponential decay function and incorporates spatial transition region parameters, while the temporal smoothing operator is constructed based on a log-probability smoothing curve and incorporates temporal transition region parameters. The cross-scale physical mapping functional is jointly compiled into a custom physical operation node in the computation graph.

[0009] Furthermore, the cloud computing platform performs physical domain differential transformation on the displacement field prediction values ​​output by the physical information neural network model to construct data fitting residual terms. It embeds the cross-scale physical mapping functional as a time-varying local stiffness boundary condition into the dynamic partial differential equation to construct physical law constraint residual terms. It also sets a double hard physical boundary condition that the time-varying local stiffness degradation of the structure is always greater than or equal to zero and does not exceed the set proportional limit of the initial distributed stiffness of the structure. It uses a linear rectifier function to construct a physical prior boundary penalty term. The above three terms are weighted and combined through a dynamic weight allocation strategy to establish a global joint loss function.

[0010] Furthermore, during the inner-loop computation process of alternating optimization iterations, the cloud computing platform freezes the microphysical mechanism parameter vector output by the outer-loop iteration step by blocking the gradient tracking node of the automatic differentiation engine for the microphysical mechanism parameter vector. It then uses the automatic differentiation engine to calculate the partial derivative matrix of the global joint loss function with respect to the network weights and calls the gradient descent optimizer to update the trainable weight matrix of the neural network. During the outer-loop computation process, the cloud computing platform freezes the trainable weight matrix of the neural network output by the inner-loop convergence, transforming the model into a nonlinear mechanical analytical operator. It calculates the partial derivative matrix of the outer-loop objective function with respect to the microphysical mechanism parameter vector, which is a weighted combination of the data fitting residual term and the physical prior boundary penalty term, and updates the microphysical mechanism parameters in reverse. Finally, a spatial extremum search algorithm is used to search for the minimum region where the stiffness value is below a set health threshold to lock the three-dimensional coordinates of the macroscopic damage.

[0011] A second aspect of the present invention provides a physical information-driven bridge multi-source data damage fusion diagnosis method, applied to the physical information-driven bridge multi-source data damage fusion diagnosis system provided in the first aspect, comprising: The sensor acquisition module deployed on the bridge structure under test acquires structural monitoring signals; Edge computing nodes decouple structural monitoring signals based on the physical properties of the signals, and extract low-frequency continuous field observation sets and high-frequency transient discrete event sets; The cloud computing platform receives low-frequency continuous field observation sets and high-frequency transient discrete event sets, constructs a continuously differentiable cross-scale physical mapping functional, and transforms the high-frequency transient discrete event sets into time-varying local stiffness boundary conditions in the partial differential equations of structural dynamics through the cross-scale physical mapping functional. Initialize the physical information neural network model, obtain the low-frequency continuous field observation set, the reconstructed time-varying dynamic partial differential equations, and the set physical prior boundary conditions, and combine them to construct the joint loss function of the physical information neural network model; The system performs a double-loop alternating optimization iteration. During the inner loop calculation, the network weights are updated through gradient descent, and during the outer loop calculation, the microscopic physical mechanism parameters are updated through residual backpropagation. When the convergence condition is met, the full-field response reconstruction result is output, and the converged microscopic physical mechanism parameters are used to calculate the local stiffness reduction matrix and damage space coordinates of the structure.

[0012] This invention provides a physical information-driven multi-source data damage fusion diagnostic system and method for bridges. It has the following beneficial effects: 1. This invention constructs a continuously differentiable cross-scale physical mapping functional. By extracting energy and spatial positioning data from high-frequency transient discrete event sets and combining spatial and temporal smoothing operators, the system transforms discrete microscopic fracture events into continuously differentiable time-varying local stiffness boundary conditions in the partial differential equations of structural dynamics. This feature eliminates the gradient calculation singularity problem in the solution process of physical control equations, enabling the fusion of macroscopic low-frequency continuous observation data and microscopic high-frequency acoustic emission data at the numerical calculation level within an automatic differentiation framework.

[0013] 2. This invention designs a dual-loop alternating optimization mechanism. To address the problem that synchronous optimization of network weights and microphysical mechanism parameters can easily lead to gradient coupling conflicts, the system freezes the microphysical mechanism parameters in the inner loop calculation to update the network weights, and freezes the network weights in the outer loop calculation and transforms the network into a nonlinear mechanical analytical operator to update the microphysical mechanism parameters. This feature realizes the decoupling solution of coupled parameters, ensures the numerical stability of the inversion process, and establishes a feedback link to back-calibrate the micro-damage mechanism parameters using the residual of the macroscopic dynamic response of the structure.

[0014] 3. This invention introduces a physical prior boundary penalty term. The system sets a hard condition that the stiffness degradation is non-negative and has a safe upper limit based on the physical boundary of the structural parameters during the evolution process. It also uses a linear rectified function to perform a one-sided truncation mapping on the out-of-bounds parameters and integrates it as a numerical constraint rule into the global joint loss function. This feature restricts the optimization algorithm from calculating non-physical parameter solutions that do not conform to the real mechanical laws in the non-convex solution space. It overcomes the overfitting problem that is easily caused by simple data-driven methods under noise interference and improves the engineering reliability of the stiffness reduction matrix and damage spatial location output by the diagnostic system. Attached Figure Description

[0015] Figure 1 This is an architecture diagram of a physical information-driven bridge multi-source data damage fusion diagnostic system according to an embodiment of the present invention. Figure 2 This is a flowchart of the multi-source heterogeneous data fusion diagnostic method according to an embodiment of the present invention; Figure 3 This is a comparison diagram of the predicted displacement response of the mid-span node in an embodiment of the present invention; Figure 4 This is a comparison diagram of the local stiffness inversion results of the entire bridge in an embodiment of the present invention; Figure 5 This is the cross-scale physical mapping functional and computation graph fusion logic diagram provided in the embodiments of the present invention; Figure 6 This is a dual-loop alternating optimization iterative control logic diagram provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See attached document Figure 1 The present invention provides a physical information-driven bridge multi-source data damage fusion diagnostic system, which may include: a sensor acquisition module, an edge computing node, and a cloud computing platform.

[0018] The sensor acquisition module is deployed on the bridge structure under test to acquire multi-source monitoring signals under the structural service conditions. The module internally incorporates a heterogeneous sensor array and cross-scale clock synchronization hardware. The heterogeneous sensor array includes piezoelectric accelerometers, fiber Bragg grating strain gauges, and piezoelectric ceramic acoustic emission probes. The cross-scale clock synchronization hardware assigns unified timestamps to the data channels of the heterogeneous sensor array based on a high-precision time synchronization protocol.

[0019] Edge computing nodes are physically connected to the sensing and acquisition modules via a communication bus. Each edge computing node carries a microprocessor unit, which performs threshold triggering and decoupling preprocessing operations on the raw multi-source monitoring signals output by the sensing and acquisition modules, extracting the corresponding discrete feature parameters and continuous signal matrices.

[0020] The cloud computing platform establishes a data transmission link with edge computing nodes through a network communication interface. Internally, the cloud computing platform deploys a graphics processing unit and a deep learning framework environment to build a physical information neural network model and perform optimization calculations for the joint loss function.

[0021] See attached document Figure 2 This method relies on the aforementioned system for execution, and the multi-source heterogeneous data fusion diagnostic method may include: S100, the edge computing node control sensing and acquisition module acquires structural monitoring signals, decouples the structural monitoring signals according to the physical properties of the signals, and extracts low-frequency continuous field observation sets and high-frequency transient discrete event sets.

[0022] The S200 cloud computing platform receives datasets uploaded by edge computing nodes and constructs a continuously differentiable cross-scale physical mapping functional. This cross-scale physical mapping functional is used to convert high-frequency transient discrete event sets into time-varying local stiffness boundary conditions in the partial differential equations of structural dynamics.

[0023] The S300 cloud computing platform initializes the physical information neural network model, obtains a low-frequency continuous field observation set, reconstructs the time-varying dynamic partial differential equations, and sets the physical prior boundary conditions, and combines them to construct the joint loss function of the physical information neural network model.

[0024] The S400 cloud computing platform executes a dual-loop alternating optimization iteration. During the inner loop computation, network weights are updated through gradient descent, and during the outer loop computation, microscopic physical mechanism parameters are updated through residual backpropagation.

[0025] S500 When the dual-ring alternating optimization iteration meets the convergence condition, the cloud computing platform terminates the iteration, outputs the full-field response reconstruction result based on the iteration result, and calls the converged microscopic physical mechanism parameters to calculate the local stiffness reduction matrix and damage space coordinates of the structure.

[0026] In step S100 above, in order to establish a data foundation that can reflect the true physical state of the bridge, the process of extracting the low-frequency continuous field observation set when the edge computing node control sensor acquisition module acquires structural monitoring signals and performs decoupling preprocessing operations may specifically include: S110, the edge computing node, allocates a unified global time reference to the monitoring channels with different sampling frequencies in the sensing acquisition module through cross-scale clock synchronization hardware.

[0027] Specifically, the sensing acquisition module continuously acquires the macroscopic vibration and strain responses of the bridge structure under test. During this process, the piezoelectric accelerometer outputs a continuous vibration acceleration signal, while the fiber Bragg grating strain gauge outputs a quasi-static level continuous strain signal. Because the rates of change of these two physical quantities differ significantly, direct data fusion can easily lead to timeline inconsistencies. Therefore, the cross-scale clock synchronization hardware generates a global clock pulse based on a standard high-precision time protocol and synchronously distributes this global clock pulse to the data acquisition channel of the piezoelectric accelerometer and the demodulator channel of the fiber Bragg grating strain gauge. During analog-to-digital conversion, each channel appends a unified hardware timestamp to the header of each sampled data frame.

[0028] For the specific working principle of hardware clock synchronization between network devices using standard high-precision time protocols, those skilled in the art can consult relevant communication standard documents. The clock offset measurement and delay compensation mechanisms are well-known technologies in this field and will not be elaborated here.

[0029] S120 and edge computing nodes extract continuous time-series observation data corresponding to spatial coordinates based on sampled data frames with accompanying hardware timestamps, and construct a low-frequency continuous field observation set.

[0030] In this embodiment, the edge computing node parses the data frame headers of each monitoring channel and extracts the original acceleration time series and the original strain time series carrying timestamps. Since the sampling frequency of piezoelectric accelerometers is usually higher than that of fiber optic strain gauges, the edge computing node performs time-domain interpolation alignment on the original time series based on a preset base time step, so that continuous signals with different physical properties are mapped to the same discrete time axis.

[0031] As a preferred approach, the aforementioned time-domain interpolation alignment operation employs a cubic spline interpolation algorithm to ensure the continuity of the second derivative of the reconstructed signal, thereby meeting the requirements for solving subsequent dynamic partial differential equations. The preset base time step is determined based on the Nyquist sampling theorem, and its reciprocal is typically set to be no less than two to three times the highest frequency of the main modes of interest in the bridge structure.

[0032] After aligning the time reference, the edge computing nodes, combined with the spatial deployment coordinates of each sensor on the bridge structure, expand the one-dimensional time series into a spatiotemporal observation matrix, thereby extracting a low-frequency continuous field observation set. This low-frequency continuous field observation set specifically includes an acceleration observation set and a strain observation set. The acceleration observation set consists of the set of measured acceleration values ​​of each piezoelectric accelerometer at its corresponding spatial coordinates, varying with time nodes. Correspondingly, the strain observation set consists of the set of measured strain values ​​of each fiber Bragg grating strain gauge at its corresponding spatial coordinates, varying with time nodes. To ensure the validity of the spatial data, the total number of piezoelectric accelerometers is pre-set according to the span of the bridge under test and the distribution of modal vibration nodes, while fiber Bragg grating strain gauges are typically deployed in high-stress concentration areas such as the mid-span and quarter-span sections of the main beam of the bridge.

[0033] The low-frequency continuous field observation set constructed through the above process can effectively suppress the time-dimensional misalignment interference of multi-source signals, and characterize the macroscopic dynamic response as discretized spatiotemporal field data with strict physical correspondence, thus providing reliable data support for its subsequent use as a continuous fitting benchmark for physical information neural network models.

[0034] In step S100 above, while constructing the continuous field observation set, the edge computing node simultaneously extracts the high-frequency transient discrete event set based on the high-frequency transient physical characteristics of the acoustic emission signal. This process may specifically include: S130: The edge computing node monitors the voltage signal output by the piezoelectric ceramic acoustic emission probe in real time. When the signal amplitude crosses the preset dynamic floating threshold, it triggers a transient time window interception action.

[0035] When microcracks inside the bridge under test initiate or propagate, they release strain energy. This energy propagates outward in the form of high-frequency elastic waves, which are captured by a piezoelectric ceramic acoustic emission probe and converted into discrete voltage burst pulses. To isolate background traffic noise from the bridge's service environment, the edge computing node employs a dynamic floating threshold calculated in real-time based on the ambient background noise. As a preferred approach, this dynamic floating threshold is the product of the root mean square value of the background noise signal within a preset environmental sampling period prior to the current moment and a set margin coefficient. When the real-time voltage amplitude of a certain acoustic emission channel exceeds this dynamic floating threshold for a continuously set time, the system determines that a microfracture event has occurred. Subsequently, the edge computing node uses this trigger moment as a starting point and extracts data segments of fixed duration to form a transient time window. This fixed duration is preset based on the complete evolution cycle of a typical elastic wave pulse from its rising edge to its complete decay back to the baseline.

[0036] S140, the edge computing node, based on a preset spatial positioning algorithm and combined with multi-channel trigger time difference, calculates the three-dimensional spatial coordinates of the occurrence of micro fracture events and the corresponding trigger time.

[0037] The bridge structure under test is typically deployed with multiple piezoelectric ceramic acoustic emission probes in a spatial array. When an event is triggered, the edge computing nodes extract the arrival times of the same burst pulse wave received by at least four arrayed acoustic emission probes. Using the arrival time difference between different probes and the propagation velocity of elastic waves in concrete or steel structures, a set of spatial spherical intersection equations is constructed to calculate the coordinates and absolute trigger time of the acoustic emission event.

[0038] For the specific mathematical solution process of using the time difference of arrival to locate the acoustic emission source in three-dimensional space, those skilled in the art can consult relevant nondestructive testing literature. The iterative solution algorithm of its nonlinear equation system is a well-known technology in this field and will not be elaborated here.

[0039] S150: The edge computing node performs integral calculations on the intercepted voltage signal within the transient time window to quantify the absolute energy released by the event, and constructs a high-frequency transient discrete event set by combining spatial coordinates and trigger time.

[0040] To correlate the acoustic emission signal with the subsequent stiffness degradation physical model, it is necessary to obtain the energy released with a relatively macroscopic physical meaning. In this embodiment, the edge computing node determines the integration interval based on the start and end times of the extracted transient time window. Within this integration interval, the edge computing node divides the square of the real-time voltage amplitude by the matching impedance constant of the piezoelectric ceramic acoustic emission probe data acquisition circuit, and then performs discrete digital integration with respect to time to obtain the absolute energy value representing the intensity of the microfracture. The matching impedance constant is pre-calibrated based on the hardware circuit design parameters of the front-end amplifier.

[0041] After quantizing the energy of a single event, the edge computing node associates and binds the independently calculated three-dimensional spatial coordinates, absolute trigger time, and corresponding absolute energy into a triplet. Throughout the entire structural monitoring cycle, the system continuously monitors and records each pulse behavior that exceeds the threshold, and summarizes all valid acoustic emission events with this triplet information captured within the monitoring cycle to construct a high-frequency transient discrete event set.

[0042] The high-frequency transient discrete event set extracted through the above process completes the dimensionality reduction and quantization of physical quantities, and can characterize the occurrence, spatiotemporal and intensity of local micro-damage in bridges with a lower data dimension.

[0043] In step S200 above, refer to the appendix. Figure 5 After the cloud computing platform receives the dataset uploaded by the edge computing nodes, in order to establish a numerical channel between the discrete acoustic emission data and the dynamic equations of continuous structures, it is necessary to perform a cross-scale physical mapping functional construction operation. This process can specifically include: S210 and cloud computing platforms establish nonlinear mapping operators that characterize the amplitude of microscopic damage based on the received high-frequency transient discrete event set.

[0044] In the actual service life of bridge structures, there is a direct physical mapping relationship between the absolute energy of acoustic emission released from local micro-fractures and the degree of stiffness degradation of the local structural materials. Because this mapping relationship is affected by the specific materials of the bridge and environmental factors, it is difficult to accurately describe it using fixed empirical formulas. As a preferred approach, a cloud computing platform constructs a nonlinear function containing unknown parameters to be inverted as a mapping operator. In this embodiment, the mapping operator can be a multinomial regression function or a feedforward neural network structure with a single hidden layer. The mathematical expression of this operator is denoted as... . In the formula, For the input high-frequency transient discrete event set, the first The absolute energy of this event; This represents the parameter vector of the microscopic physical mechanism to be inverted and calibrated. When this operator uses a polynomial regression function... The undetermined weight coefficients correspond to polynomials of various orders; when a feedforward neural network is used... The connection weights and biases between corresponding network layers. Characterization in a given parameter Below, the mapping rule for absolute energy to be converted into the dimensionless microscopic stiffness degradation amplitude.

[0045] The S220 cloud computing platform constructs a continuously differentiable spatiotemporal smoothing operator to eliminate gradient singularities in the damage-triggered localization model.

[0046] To locate the stiffness degradation amplitude generated by the aforementioned operators in the spatiotemporal domain, traditional mechanical models typically use the Dirac function to locate spatial points and the standard Heaviside step function to locate temporal trigger points. However, the physical information neural network used for solving the equations relies on automatic differentiation mechanisms to calculate the partial derivative residuals of the governing equations, which requires the boundary conditions to possess high-order continuous differentiability within the spatiotemporal domain. Directly introducing these two types of discontinuous singular functions often leads to situations where the derivative exhibits singular values ​​at the event trigger points and is zero in other regions, resulting in computational failures such as gradient explosion or gradient vanishing.

[0047] In this embodiment, the cloud computing platform introduces a globally continuously differentiable narrowband Gaussian kernel function to replace the Dirac function to form a spatial smoothing operator, and introduces a Sigmoid-type function to replace the standard step function to form a time smoothing operator.

[0048] Specifically, it is recorded as the coordinates of the event occurrence. The spatial smoothing operator centered on is This spatial smoothing operator is constructed using an exponential decay function and introduces parameters. The spatial transition zone characterizing the radius of influence of local stiffness degradation is determined based on the average microcrack propagation length of the bridge material under test. It is denoted by the event trigger time. The time smoothing operator centered on is This time smoothing operator is constructed using a log-odds smoothing curve and introduces parameters. The time transition region characterizing the rate of stiffness degradation is determined by pre-setting the transient response time constant of the material during brittle fracture. These two operators mathematically satisfy the requirement of derivative continuity for automatic differentiation operations, and physically reflect the objective law that structural damage exhibits a spatial influence transition region and an extremely short time buffer zone.

[0049] The S230 cloud computing platform integrates nonlinear mapping operators, spatiotemporal smoothing operators, and dissipative physics laws to construct parameterized cross-scale physical mapping functionals.

[0050] After defining the operators mentioned above, the cloud computing platform introduces the relaxation phenomenon, which characterizes stress redistribution in structural dynamics, i.e., the nonlinear dissipation process caused by the redistribution of internal forces after damage occurs. By comprehensively superimposing all high-frequency transient discrete events within the monitoring period, a complete mathematical expression describing the local stiffness degradation of the structure is established.

[0051] The specific formula for this cross-scale physical mapping functional is: ; In the formula, This represents the spatial coordinates under the constraint of the unknown parameter θ. Time node The corresponding time-varying local stiffness degradation of the structure; This represents the total number of events recorded in the set of high-frequency transient discrete events; For the introduced nonlinear dissipation relaxation term, This is the preset system structure dissipation relaxation constant.

[0052] By constructing this cross-scale physical mapping functional, the cloud computing platform transforms discrete high-frequency acoustic emission characteristics into continuously differentiable three-dimensional physical surface variables in the spatiotemporal domain. This reconstruction not only provides time-varying local stiffness boundary conditions with clear physical meaning for the dynamic partial differential equations, but its global smoothness also provides numerical stability guarantees for subsequent neural network models to participate in the equation solution.

[0053] In step S200 above, when constructing the physical information neural network on the cloud computing platform, the optimization of network parameters relies on the automatic differentiation engine built into the deep learning framework. To ensure that the aforementioned cross-scale physical mapping functional can effectively adapt to this automatic differentiation mechanism, the cloud computing platform needs to further establish the derivative propagation path for each smoothing operator, thereby avoiding computational dead zones caused by singularities. This derivative analysis and singularity elimination process may specifically include: In the computational graph of the physical information neural network, the S240 cloud computing platform analyzes the derivative propagation path of the spatial smoothing operator to eliminate the singularity of the spatial gradient.

[0054] The dynamic partial differential governing equations typically contain second-order derivative terms with respect to spatial coordinates. If an ideal Dirac function is used to characterize the spatial location of microscopic damage, its derivative tends to infinity at the damage center. This can cause gradient explosion during chained differentiation in the automatic differentiation engine, leading to divergent updates of network weights.

[0055] In this embodiment, the cloud computing platform embeds the aforementioned spatial smoothing operator nodes in the form of a narrowband Gaussian kernel function into the computation graph. During backpropagation, the automatic differentiation engine calculates its first-order spatial partial derivatives with respect to spatial coordinates in analytical form. Since the spatial smoothing operator is constructed based on an exponential decay form, the solution of its first-order spatial partial derivatives not only preserves the spatial coordinates of the acoustic emission event and the radius parameter of the local stiffness degradation effect, but the expression of this partial derivative is also continuous and bounded throughout the real number domain. Based on this, the cloud computing platform can stably calculate the second-order spatial residual of the partial differential equation containing local damage terms, thereby guiding the network to learn the physical deformation field within the damage transition region.

[0056] The S250 cloud computing platform analyzes the derivative propagation path of the time smoothing operator in the computation graph to overcome the gradient vanishing phenomenon in the time dimension.

[0057] When calculating the time evolution of the system state, if the standard step function is used to characterize the transient moment of damage occurrence, the derivative of this function is always zero in all regions outside the triggering moment. This leads to a significant decay or even zeroing of the time residual gradient with respect to the microscopic physical mechanism parameters during joint optimization iterations, causing the closed-loop inversion mechanism of the outer loop to fail.

[0058] As a preferred approach, the cloud computing platform employs a Sigmoid-type time smoothing operator and analyzes its time partial derivatives to maintain effective gradient propagation. Since this time smoothing operator is constructed based on a logarithmic probability curve, its first-order partial derivative with respect to time is mathematically represented as the product of the operator itself and its mirror-complementary term. This partial derivative result is constrained by both the absolute trigger time of the acoustic emission event and the time transition parameter of the stiffness degradation rate. Through this partial derivative analysis, the automatic differentiation engine can continuously output non-zero smooth gradient signals within a very short time buffer band near the event trigger time. This mechanism ensures that the update step size of the microscopic parameters on the time axis is not zero, effectively driving the convergence of the inversion algorithm.

[0059] The S260 cloud computing platform solidifies the differentiation logic of the aforementioned continuously differentiable operators into the underlying constraints of the physical information neural network, establishing a global gradient backpropagation link without computational dead zones.

[0060] Since both spatial and temporal smoothing operators possess analytic, bounded, high-order continuous derivatives in the spatiotemporal domain, the cloud computing platform compiles them, along with dissipative relaxation terms, into custom physical operation nodes in the computational graph. Specifically, relying on automatic differentiation technology based on the backpropagation chain rule, the cloud computing platform performs element-wise operations on the multidimensional output matrices corresponding to the spatial smoothing operator, temporal smoothing operator, and nonlinear dissipative relaxation terms through tensor multiplication, combining them to form a local stiffness degradation gradient tensor.

[0061] In any iteration of the alternating dual-ring optimization, whether calculating the fitted gradient with respect to network weights or the physical residual gradient with respect to microscopic mechanism parameters, the automatic differentiation engine can complete backpropagation along this continuous and smooth operator link. This mechanism resolves the mathematical contradiction that high-frequency transient discrete events cannot be directly substituted into macroscopic continuous dynamic equations. This not only protects the solution process of partial differential equations from singular boundaries but also ensures the topological continuity of the joint loss function throughout the optimization space, laying the algorithmic foundation for subsequent robust physical mechanism parameter inversion.

[0062] In step S300 above, after the reconstruction of the cross-scale physical mapping functional is completed on the cloud computing platform, in order to drive the network to learn the structural evolution state that conforms to the physical causal relationship, it is necessary to establish a time-varying boundary modulation neural network and combine the joint loss function. This process may specifically include: The S310 cloud computing platform initializes the basic topology of the physical information neural network and establishes the network input / output dimensions and the source of training samples.

[0063] In this embodiment, the physical information neural network specifically adopts a multi-layer fully connected feedforward neural network model. For the bridge monitoring business scenario, the input data dimension of this model is defined as a two-dimensional feature vector, corresponding to the continuous spatial coordinate positions of the bridge structure under test. With time nodes After the input feature data enters the network, it flows through multiple fully connected hidden layers for nonlinear mapping. Adjacent hidden layers are connected by a dense weight matrix in a fully connected form. The number of layers is typically set to 3 to 5, depending on the span of the bridge structure under test and the monitoring time span. The data flow finally reaches the output layer of a single node. This output layer corresponds to a specific physical state, namely, the predicted value of the macroscopic dynamic displacement field of the structure at the spatiotemporal coordinate point (x, t), denoted as... ,in Let be the trainable weight matrix of the neural network.

[0064] Since subsequent partial differential equations and residual calculations involve second-order derivatives of the displacement field, a hyperbolic tangent function (Tanh) is used as a nonlinear activation function in each hidden layer to ensure the second-order differentiability of the computation graph throughout the entire domain. During network training, the training sample set in the input space consists of two parts: the first part is the spatiotemporal coordinates of the actual sensor deployment locations, with corresponding real physical quantity labels derived from the previously extracted low-frequency continuous field observation set; the second part is the coordinates of unlabeled collocation points randomly generated throughout the spatiotemporal domain of the bridge using the Latin hypercube sampling algorithm, used to support the constraint calculations of subsequent continuous physical equations.

[0065] The S320 cloud computing platform extracts the network output results and performs differential transformation to match the physical dimensions of multi-source sensor data, thereby constructing data fitting residual terms.

[0066] Since fiber optic strain gauges and piezoelectric accelerometers do not directly measure displacement fields, the cloud computing platform needs to convert the network-output displacement field prediction values... A physical domain differential transformation is performed. Based on the fundamental relationships of solid mechanics, the cloud computing platform uses an automatic differential engine to perform a second derivative of the predicted displacement field with respect to time to obtain the predicted acceleration field; simultaneously, it performs a second derivative of the predicted displacement field with respect to space and multiplies it by the distance from the measuring point to the neutral axis of the cross section to obtain the predicted strain field.

[0067] To quantify the deviation between predicted and measured values, the cloud computing platform correlates the predicted field with a low-frequency continuous field observation set, constructing a data fitting residual term in the form of a summed mean square error. Specifically, this data fitting residual term consists of two summed parts: the first part is the mean square error between the predicted acceleration field output by the network and the measured acceleration values ​​in the low-frequency continuous field observation set, averaged according to the total number of acceleration sensors; the second part is the mean square error between the predicted strain field output by the network and the measured strain values ​​in the low-frequency continuous field observation set, averaged according to the total number of fiber Bragg grating strain gauges. By minimizing this residual term, the network can be guided to approximate the measured macroscopic response.

[0068] The S330 cloud computing platform substitutes cross-scale physical mapping functionals into the structural dynamics control equations to construct time-varying modulated physical law-constrained residual terms.

[0069] This is the key mechanism that distinguishes this invention from traditional pure data-driven models. In order to ensure that the network output not only conforms to discrete measured data, but also follows the evolutionary laws of materials mechanics, the cloud computing platform directly embeds the aforementioned constructed cross-scale physical mapping functional as a time-varying local stiffness boundary condition into the dynamic partial differential equations used to describe the macroscopic vibration behavior of the beam structure.

[0070] The specific formula for the residual term constrained by this physical law is as follows: ; In the formula, This represents the residual term constrained by physical laws, and its calculation is based on the mean square error integral of an unlabeled set of randomly matched points. The linear mass of the bridge structure to be tested is pre-calibrated based on the bridge design drawings. This represents the distributed stiffness of the structure in its initial, undamaged state. This represents the excitation function of the external environment, such as the distribution of traffic flow loads or wind loads; Indicates unknown parameters Under constraints, spatial coordinates Time node The corresponding time-varying local stiffness degradation of the structure, where The vector of parameters for the microscopic physical mechanism to be inverted and calibrated.

[0071] According to the above formula, when a microscopic damage triggers an acoustic emission event, the stiffness parameter in the equation will undergo dynamic spatial localization and temporal decay. The network tends to adjust its weights. To adapt to this equation variation driven by micro-mechanisms, the physical causal ingot is mathematically characterized as a result of micro-crack propagation causing an instantaneous decrease in local stiffness and altering the global mode.

[0072] Based on the set balance weights, the S340 cloud computing platform fuses the data fitting residuals with the physical law constraint residuals to establish the basic joint loss function of the physical information neural network.

[0073] Since the data terms involve the absolute magnitude of measured values, while the physical constraint terms represent the structural residuals of partial differential equations, their dimensions and orders of magnitude differ significantly. The cloud computing platform employs a dimensionless dynamic weight allocation strategy, combining the two weighted terms to form the overall joint optimization objective. In this embodiment, the cloud computing platform multiplies the data fitting residual term with the first dynamic balance weight coefficient, multiplies the physical constraint residual term with the second dynamic balance weight coefficient, and algebraically sums the products to construct the basic joint loss function. For the implementation of the adaptive balancing strategy for the weight coefficients of the aforementioned loss functions with different dimensions, those skilled in the art can use learning rate annealing or neural tangent kernel (NTK) dynamic weighting algorithms commonly used in network training; these are well-known techniques in the field of deep learning and will not be elaborated upon here. This basic joint loss function initially establishes the gradient direction for network optimization, taking into account both the apparent fitting of macroscopic measured data and the underlying constraints of microscopic damage mechanisms.

[0074] In step S300 above, after establishing the basic joint loss function, the cloud computing platform faces the non-convexity defect in the parameter optimization process. To prevent pure mathematical fitting from causing the microscopic parameter inversion results to violate the true laws of mechanics, the cloud computing platform needs to establish a physical prior deviation prevention penalty mechanism. This process may specifically include: Based on fundamental principles of solid mechanics, the S350 cloud computing platform determines the reasonable physical boundaries of structural parameters during the evolution process.

[0075] In multi-parameter joint optimization iterations, the solution space corresponding to the basic joint loss function typically exhibits highly non-convex characteristics. If relying solely on gradient descent to blindly search for minima, the optimization algorithm may exploit extreme parameter combinations to reduce mathematical residuals. For example, cross-scale physical mapping functionals may output negative stiffness degradation, which corresponds to a non-physical stiffness enhancement effect at the physical level; or it may output extremely large stiffness degradation exceeding the initial structural limits, which in practical engineering means a complete loss of local bearing capacity, inconsistent with the service status characterized by monitoring data.

[0076] Based on the aforementioned physical contradictions, the cloud computing platform establishes two rigid physical boundary conditions for the local stiffness degradation: first, the local stiffness degradation caused by damage is always greater than or equal to zero; second, there is a safety upper limit for this local stiffness degradation, which cannot exceed the set proportional limit of the initial distributed stiffness of the structure.

[0077] S360, a cloud computing platform, uses a one-sided truncation operator to construct physical prior boundary penalty terms, transforming the established physical boundary conditions into numerical constraint rules.

[0078] To seamlessly integrate the aforementioned mechanical principles into the automatic differential calculation graph of the physical information neural network, the cloud computing platform employs a linear rectified function (ReLU operator) to construct the penalty term. Since the output of the linear rectified function is zero when the input is less than zero and increases linearly when the input is greater than zero, this one-sided suppression characteristic effectively isolates the physical reasonable domain from the physical failure domain.

[0079] The specific formula for the physical prior boundary penalty term is as follows: ; In the formula, The physical prior boundary penalty term is calculated by discrete accumulation based on the spatiotemporal domain Latin hypercube random collocation set generated in advance during the network training phase. and These correspond to the continuous spatial coordinate positions and time nodes of the bridge structure under test, respectively. This represents the operation of a linear rectifier function, whose algorithm is to take the maximum value between the input value and zero; Indicates unknown parameters Under constraints, spatial coordinates Time node The corresponding time-varying local stiffness degradation of the structure, where The vector of parameters representing the microscopic physical mechanism to be inverted and calibrated; This represents the distributed stiffness of the structure in its initial, undamaged state. This is a preset structural safety degradation limit threshold.

[0080] In this embodiment, the structural safety degradation limit threshold is... The value of is set to a fixed constant between (0, 1). As a preferred method, the cloud computing platform pre-calibrates this parameter based on the finite element ultimate bearing capacity analysis results of the bridge under test, typically setting it to 0.3 to 0.5. When the inverted stiffness degradation is between zero and the safety upper limit, the input to the linear rectifier function is negative, the penalty term output is zero, and it does not interfere with normal optimization; once the inversion parameter exceeds the limit, the linear rectifier function is activated and outputs a positive residual, thereby increasing the total loss value.

[0081] The S370 cloud computing platform introduces physical prior boundary penalty terms into the basic joint loss function, combining them to form a global joint loss function.

[0082] After obtaining the aforementioned penalty mechanism, the cloud computing platform globally fuses it with the data fitting residual term and the physical constraint residual term. Specifically, the cloud computing platform introduces a third dynamic balance weight coefficient. The cloud computing platform multiplies the physical prior boundary penalty term with the third dynamic balance weight coefficient, and then sums it into the aforementioned basic joint loss function, thereby establishing a global joint loss function to guide the network's iterative updates.

[0083] To ensure the absolute priority of physical hard constraints in the optimization process, the third dynamic balance weight coefficient is set to a constant scalar value much larger than the first and second dynamic balance weight coefficients in the early stages of iteration. Through this penalty mechanism, the cloud computing platform endows the network with the ability to judge the physical rationality of the results. In the subsequent alternating double-loop optimization process, the mechanism parameter vector is involved. The update step size is limited to the solution space allowed by physical logic, which ensures the convergence stability and engineering reliability of the microscopic parameter closed-loop inversion.

[0084] In step S400 above, after establishing the global joint loss function, directly performing joint differentiation can easily lead to gradient coupling conflicts because the joint loss function simultaneously couples the trainable weight matrix of the neural network and the parameter vector of the microscopic physical mechanism. The cloud computing platform employs a dual-loop alternating optimization algorithm for decoupling and dimensionality reduction, as detailed in the appendix. Figure 6 The specific execution process of the inner loop iterative algorithm may include: When the inner loop iteration begins, the S410 cloud computing platform freezes the microscopic physical mechanism parameter vector output by the current outer loop iteration step, establishing deterministic physical boundary conditions.

[0085] In the In the context of the next outer loop iteration, the cloud computing platform will use the microscopic physical mechanism parameter vector (denoted as...) The parameter vector is set to a non-differentiable constant state. In practice, the cloud computing platform stops the automatic differentiation engine from tracking the gradient of the parameter vector of this microscopic physical mechanism, thus halting its differentiation. Through this freezing operation, the time-varying local stiffness degradation of the structure output by the cross-scale physical mapping functional becomes a known and definite numerical field within the current inner loop iteration cycle. This is equivalent to providing a set of definite, smooth physical boundaries that evolve with microscopic damage for the partial differential equations, thereby reducing the originally complex mechanism-data coupling optimization problem to a problem of predicting continuous dynamic responses under specific boundary constraints.

[0086] The S420 cloud computing platform performs forward computation based on the frozen microscopic physical mechanism parameter vector and the current network weights to obtain the global joint loss function value.

[0087] In this embodiment, the cloud computing platform batches the spatiotemporal coordinates corresponding to the low-frequency continuous field observation set and the coordinates of the Latin hypercube random collocations into the physical information neural network. The network is based on the current... The trainable weight matrix of the inner loop iteration performs forward propagation, outputting the current corresponding macroscopic dynamic displacement field prediction value. Based on the aforementioned established global joint loss function calculation logic, the system extracts this prediction value and integrates the data fitting residual term, the physical law constraint residual term, and the physical prior boundary penalty term, thereby outputting the current global joint loss value in scalar form.

[0088] The S430 cloud computing platform uses an automatic differentiation engine to calculate the partial derivatives of the global joint loss function with respect to the network weights and performs gradient descent update operations.

[0089] After obtaining the scalar form of the global joint loss value, the cloud computing platform initiates the backpropagation mechanism of the computational graph. Relying on the automatic differentiation engine, the system calculates the partial derivative gradient matrix of the global joint loss function with respect to the trainable weight matrices of each hidden layer node in the neural network, given the microscopic physical mechanism parameter vector. Subsequently, the cloud computing platform calls a preset gradient descent optimizer to iteratively update the network weights. Specifically, the cloud computing platform subtracts the product of the preset inner loop learning rate step size parameter and the aforementioned partial derivative gradient matrix from the current step's trainable weight matrix of the neural network to obtain the updated trainable weight matrix for the next iteration.

[0090] For the adaptive adjustment principle of first-order momentum and second-order momentum estimation involved in the gradient descent optimizer (such as using the Adam optimization algorithm), those skilled in the art can consult the standard algorithm library documentation in the field of deep learning. It is a well-known technology in this field and will not be elaborated here.

[0091] As a preferred approach, the cloud computing platform calculates the difference in the global joint loss function between two adjacent inner loop iterations after each network weight update. This inner loop iteration process is repeated multiple times until the difference is less than a preset inner loop convergence tolerance (e.g., set to 10). -4 (A very small positive real number on the order of magnitude), the cloud computing platform can guide the network to complete the fitting under given microscopic damage constraints, outputting a continuous macroscopic response prediction field that approximates the measured macroscopic data and satisfies the physical consistency of the equations. This process establishes the response mapping relationship of the network model to the current microscopic assumptions, providing a reliable observational residual benchmark for the subsequent back-inference of the outer loop mechanism parameters.

[0092] In step S400 above, after the inner loop iteration converges and establishes the network's response mapping relationship to the current microscopic hypothesis, the cloud computing platform needs to switch to the outer loop iterative control flow, using the macroscopic observation residuals to perform closed-loop inversion calibration of the acoustic emission microscopic parameters. The specific execution process of this outer loop iterative algorithm may include: When the outer loop iteration is initiated, the S440 cloud computing platform freezes the trainable weight matrix of the neural network output by the inner loop convergence, transforming the physical information neural network into a nonlinear mechanical analytical operator.

[0093] During the outer loop iteration phase, the cloud computing platform sets the trainable weight matrix of the converged neural network to a non-differentiable state. Specifically, the system freezes the model by severing the gradient tracking link between the automatic differentiation engine and the hidden layer parameters. After this operation, the deep learning model, originally responsible for fitting and approximating the network, is transformed into a fixed-parameter nonlinear partial differential equation solver. This solver is then used in subsequent calculations to establish a cross-scale mapping channel from microscopic mechanism parameters to the macroscopic dynamic response of the structure, providing a stable evaluation benchmark for optimizing the mechanism parameters.

[0094] The S450 cloud computing platform uses the current microscopic physical mechanism parameter vector as the independent variable to perform forward computation and extract the outer loop objective function value, which includes data residuals and physical boundary penalties.

[0095] The cloud computing platform uses the microscopic physical mechanism parameter vectors left over from the previous iteration as the current input source. Based on this, the system re-invokes the cross-scale physical mapping functional to generate the local stiffness degradation field and uses the frozen network operators to calculate the current macroscopic dynamic displacement field predictions. The system extracts the data fitting residual term and the physical prior boundary penalty term from the global joint loss function and combines them to construct an outer-loop objective function for optimizing the mechanism parameters.

[0096] Specifically, the outer-loop objective function is composed of two weighted components. The cloud computing platform calculates the data fitting residual term based on the currently given frozen network weights and multiplies it by the first dynamic equilibrium weight coefficient. Simultaneously, it calculates the physical prior boundary penalty term based on the current microscopic physical mechanism parameter vector and multiplies it by the third dynamic equilibrium weight coefficient. Summing the products of these two components yields the outer-loop objective function used to guide the iterative update of the outer loop.

[0097] Since the residual term constrained by physical laws tends to a minimum during the inner loop solution, the outer loop objective function focuses on the fitting approximation error between macroscopic low-frequency continuous observation data and microscopic damage energy evolution characteristics. The construction of this outer loop objective function utilizes the macroscopic vibration phenomenon of the bridge as a verification posterior and applies engineering common sense constraints based on the physical prior anti-deviation penalty mechanism.

[0098] The S460 cloud computing platform, based on the backpropagation chain rule of the automatic differentiation engine, calculates the partial derivative matrix of the outer loop objective function with respect to the parameter vector of the microscopic physical mechanism, and performs reverse updates.

[0099] After obtaining the scalar value of the outer-loop objective function, the cloud computing platform removes the restriction on the differentiation of the parameter vectors of the microscopic physical mechanism. Relying on the automatic differentiation engine, the system calculates the gradient direction of the outer-loop objective function with respect to each undetermined coefficient in the cross-scale physical mapping functional. Thanks to the aforementioned globally continuously differentiable spatiotemporal smoothing operator, the data residual gradient of the macroscopic low-frequency measurement points can propagate backward along the network graph to the functional coefficient nodes representing microscopic high-frequency acoustic emissions.

[0100] The cloud computing platform invokes the optimizer to perform gradient descent update operations on the microphysical mechanism parameter vector. Specifically, the cloud computing platform calculates the partial derivative gradient matrix of the aforementioned outer-loop objective function with respect to the current microphysical mechanism parameter vector. Then, it subtracts the product of the preset outer-loop learning rate step size parameter and this partial derivative gradient matrix from the current step's microphysical mechanism parameter vector to obtain the updated microphysical mechanism parameter vector for the next iteration. As a preferred approach, to maintain the stability of parameter inversion, this step size parameter is typically set to an empirical constant (e.g., 10) that is orders of magnitude lower than the inner-loop learning rate. -5 (magnitude).

[0101] Through the aforementioned outer loop update process, the system utilizes the macroscopic structural vibration residuals to achieve reverse adaptive correction of the evolution law of high-frequency local characteristics of acoustic emission, establishing a physical closed-loop feedback link from macroscopic manifestations to microscopic mechanisms of the bridge.

[0102] In step S400 above, after completing a single outer loop iteration update, the cloud computing platform needs to perform a global state evaluation of the current dual-loop alternating optimization system to determine whether to continue into the next alternating loop or terminate the optimization process. The execution process of this global joint convergence determination may specifically include: S470, cloud computing platform extracts microscopic physical mechanism parameter vectors in adjacent outer loop iteration steps, and calculates the physical mechanism parameter update norm.

[0103] The cloud computing platform retrieves the microscopic physical mechanism parameter vector output from the previous outer loop iteration that has just been updated, as well as the initial input microscopic physical mechanism parameter vector before this update. The system performs a difference operation on these two high-dimensional vectors and calculates the Euclidean norm of the difference vector. This norm represents the absolute change magnitude of the microscopic mechanism parameters in this round of outer loop updates, i.e., the physical mechanism parameter update norm.

[0104] S480 and cloud computing platforms extract the inner loop network weights corresponding to the outer loop iteration steps and calculate the network weight update norm.

[0105] Simultaneously with the update of the outer loop parameters, the cloud computing platform retrieves the inner loop convergence network weights associated with the latest outer loop iteration, as well as the inner loop convergence network weights associated with the previous outer loop iteration. Since the weight distribution of the neural network reflects the stability of the macroscopic continuous dynamic response mapping relationship, the cloud computing platform also performs difference operations on the two sets of dense weight matrices. Specifically, the cloud computing platform flattens and concatenates the weight matrices and bias vectors of all hidden layers into a one-dimensional long tensor, and then solves for the Euclidean norm of the reconstructed difference tensor to extract the scalar form of the network weight update norm.

[0106] The S490 cloud computing platform integrates the physical mechanism parameter update norm and the network weight update norm, and performs global joint convergence condition judgment based on the set control strategy.

[0107] To ensure that the macroscopic state approximation and the microscopic mechanism inversion can reach a synchronous and stable state, the cloud computing platform has constructed a global convergence judgment criterion that includes dual condition constraints.

[0108] Specifically, the cloud computing platform compares the physical mechanism parameter update norm and the network weight update norm obtained from the aforementioned calculations, extracts the maximum value between the two, and compares this maximum value with the preset global joint tolerance threshold.

[0109] In this embodiment, the global joint tolerance threshold is typically set to an extremely small positive real constant. As a preferred approach, this tolerance threshold is preset based on the double-precision floating-point truncation error of the computing hardware and the allowable accuracy for actual bridge structural damage identification; for example, a value of 10. -6 When the maximum norm value is less than or equal to the tolerance threshold, it indicates that the update magnitude of the network parameters and micro-mechanism parameters is extremely small. At this time, the cloud computing platform determines that the alternating optimization iteration of the two loops meets the convergence condition and terminates the entire deep learning training loop; conversely, if any norm value exceeds the set threshold, the cloud computing platform transmits the updated micro-physical mechanism parameter vector back to the inner loop, unlocks the differentiable state of the network weights, and re-enters the next round of the inner loop forward approximation process.

[0110] Through this synchronous convergence determination mechanism, the cloud computing platform avoids unsteady outputs caused by overfitting or underfitting of micro parameters in the network, providing mathematical criteria for generating high-fidelity bridge damage diagnosis results.

[0111] In step S500 above, after the alternating optimization iteration of the dual loops satisfies the global joint convergence condition, all objective functions of the physical information neural network have been reduced to the minimum value allowed by physical constraints. At this point, the cloud computing platform uses the converged parameter system to perform reverse engineering calculations for the bridge structure to extract quantitative damage indicators with practical engineering guidance significance. The specific process of damage localization and stiffness reduction analysis may include: The S510 cloud computing platform extracts the converged microscopic physical mechanism parameter vector and high-frequency transient discrete event set, and substitutes them into the maximum stiffness reduction of the local structure of the cross-scale physical mapping functional analysis structure.

[0112] At the end of the iteration, the cloud computing platform solidifies the globally optimal microscopic physical mechanism parameter vector (denoted as ). To assess the maximum physical damage to the structure caused by each microacoustic emission event, the system ignores the nonlinear dissipation relaxation term during time evolution and focuses on the spatial center point at the moment the event is triggered.

[0113] The cloud computing platform retrieves the first high-frequency transient discrete event set extracted from early edge computing nodes. The system calculates the absolute energy and spatial location parameters of the sub-event. It then substitutes this absolute energy and the optimal mechanism parameter vector into the previously established nonlinear mapping operator. Since the spatial smoothing operator takes the theoretical peak value at the damage center coordinates, the system calculates the maximum stiffness reduction caused by the subacoustic emission event at its occurrence coordinate center point using algebraic multiplication.

[0114] Specifically, the system extracts the constant peak value of the spatial smoothing operator at the core coordinate point and multiplies this constant peak value with the output of the aforementioned nonlinear mapping operator to obtain the reduced peak value of a single event. By performing this traversal calculation on all valid acoustic emission events within the monitoring period, the cloud computing platform generates a set of discrete stiffness reduction peak values ​​characterizing the degree of local microscopic damage.

[0115] The S520 and cloud computing platform, combined with a spatial smoothing operator, extend the discrete stiffness reduction peak to the bridge entity spatial coordinate system, reconstructing the continuous stiffness distribution matrix of the structure.

[0116] To provide bridge maintenance personnel with an intuitive macroscopic stiffness assessment map, the cloud computing platform needs to restore isolated microscopic reduction peaks to a continuous macroscopic mechanical field. Using the initial non-destructive distributed stiffness map of the bridge under test as a benchmark, the cloud computing platform extracts the local stiffness degradation influence radius parameter from the spatial smoothing operator.

[0117] Based on this radius of influence parameter, the cloud computing platform calculates the stiffness reduction effect of each damage event on the surrounding continuous cross-sections. The system performs a spatial superposition operation on the weakening influence domains caused by all events and subtracts them from the initial undamaged distributed stiffness, thereby generating a continuous stiffness distribution matrix of the bridge entity space under the current service state. This matrix represents the true residual stiffness of each key cross-section of the bridge, realizing a physical equivalent restoration from microscopic energy to macroscopic parameters.

[0118] The S530 cloud computing platform uses a spatial extremum search algorithm to locate the three-dimensional coordinates of the damage concentration area based on the reconstructed continuous stiffness distribution matrix.

[0119] After obtaining the continuous stiffness distribution matrix, the cloud computing platform invokes a gridded spatial extremum search algorithm. The system traverses the entire three-dimensional spatial coordinate system of the bridge structure, searching for minimum regions in the continuous stiffness distribution matrix where stiffness values ​​are below a set health threshold. As a preferred approach, this set health threshold is a predetermined percentage of the initial undamaged distributed stiffness of the bridge under test (e.g., set to 90% of the initial undamaged distributed stiffness). The cloud computing platform then extracts the set of spatial coordinate nodes containing these minimum values.

[0120] For the specific optimization and filtering mechanism of the spatial extremum search algorithm, those skilled in the art can consult relevant literature on standard numerical computation and matrix analysis. The local gradient method and multi-point traversal method are well-known techniques in this field and will not be elaborated here.

[0121] These locked minimum node sets represent the three-dimensional coordinates of macroscopic damage indicating densely intersecting cracks and substantial material degradation within the bridge structure. Through this mapping and inversion mechanism, the cloud computing platform overcomes the limitations of poor physical interpretability in conventional data models, and can output the spatial location coordinates of the damage and its actual reduction coefficients to the infrastructure management platform with transparent mechanical causal logic.

[0122] In step S500 above, after completing damage localization and stiffness reduction analysis, the cloud computing platform utilizes a physically consistent full-field continuous response reconstruction, achieved through a globally joint convergent physical information neural network, without being limited by the physical deployment location of discrete sensors. This reconstruction process may specifically include: Based on the safety assessment requirements of the bridge structure, the S540 cloud computing platform sets the target spatiotemporal resolution and generates a high-density spatiotemporal prediction grid within the entire bridge domain.

[0123] Due to limitations in deployment costs and construction conditions, sensor acquisition modules typically only provide discrete observation data at a limited number of structural nodes. To obtain continuous mechanical state data at any location on the bridge under test, the cloud computing platform divides the physical geometric boundaries of the bridge and the entire monitoring period into a uniform spatiotemporal prediction grid according to preset spatial and temporal steps. As a preferred approach, the spatial step is set to the centimeter level, and the time step to the millisecond level. The cloud computing platform extracts the spatial coordinates and time nodes of all intersecting nodes in this grid, and stitches them together to construct a dimension... The input feature tensor, where This represents the total number of nodes in the generated spatiotemporal prediction grid.

[0124] The S550 cloud computing platform inputs the spatiotemporal prediction grid into the converged physical information neural network, performs forward computation to output the continuous displacement response across the entire field.

[0125] In this embodiment, the cloud computing platform loads the globally optimal network weights saved at the termination of the dual-loop alternating optimization iteration. The system inputs the aforementioned constructed input feature tensors into the network in batches. Since the weights of this physical information neural network have simultaneously absorbed the numerical boundaries of macroscopic low-frequency observation data and the constraints of partial differential equations driven by microscopic damage during the training phase, its forward computation process possesses physical consistency. After multiple layers of nonlinear mapping, the network directly outputs continuous displacement field data corresponding to the high-density grid. This output data specifically uses... The displacement prediction vector is presented. This output not only covers the measurement gaps of the original sensors, but also conforms to the energy transfer laws of structural dynamics in terms of spatiotemporal evolution logic.

[0126] The S560 cloud computing platform is based on the full-field continuous displacement response and relies on the automatic differential engine to perform physical differentiation, thereby analyzing and reconstructing the full-field continuous strain field and the full-field continuous acceleration field.

[0127] Bridge fatigue life assessment and load-bearing capacity verification typically rely on continuous strain distribution characteristics. The cloud computing platform utilizes the built-in automatic differentiation engine of a deep learning framework to perform two consecutive partial derivative calculations on the full-field continuous displacement response output from the aforementioned forward calculation with respect to spatial coordinates. Subsequently, the system retrieves the cross-sectional geometric information from the design parameters of the bridge under test and extracts the vertical distance parameter from the measurement point corresponding to the predicted grid node to the neutral axis of the cross-section. The system algebraically multiplies this distance parameter with the aforementioned second-order spatial partial derivative matrix, converting it into a full-field continuous strain field based on the fundamental assumptions of mechanics of materials.

[0128] Using the same differential principle, the cloud computing platform performs second-order partial derivative calculations on the full-field continuous displacement response with respect to time nodes to obtain the full-field continuous acceleration field of the bridge structure. The cloud computing platform then performs multi-dimensional alignment and encapsulation of the reconstructed displacement field, strain field, and acceleration field, and outputs them as standard time-series structural state data.

[0129] Through the aforementioned continuous response reconstruction process across the entire field, the system expands the data coverage of traditional structural health monitoring systems. The cloud computing platform, relying on algorithmic models deeply integrated with physical laws, provides a high-resolution benchmark mechanical response field for subsequent bridge condition early warning and infrastructure operation and maintenance management.

[0130] Specific application examples: Implementation scenario definition: The implementation focuses on a simply supported steel box girder bridge with a span of 30 meters (bridge length). At the mid-span of the main girder of the bridge ( ) and quarter point ( Fiber optic strain gauges and piezoelectric accelerometers are installed at the location. Four piezoelectric ceramic acoustic emission probes are evenly distributed along the longitudinal direction of the bridge.

[0131] Monitoring and Operation Process: Physical event occurred: The bridge vibrated under the load of moving vehicles. During the monitoring period... The micro-fatigue cracks in the mid-span base plate propagate instantaneously.

[0132] Edge computing preprocessing (corresponding to S100): Accelerometers and strain gauges continuously output low-frequency continuous responses. Simultaneously, an acoustic emission probe captures the elastic waves propagating from the crack, and the edge computing node extracts the trigger time of this event. Spatial positioning coordinates m, and the absolute energy is calculated by integration. .

[0133] Boundary functional construction (corresponding to S200): The cloud computing platform substitutes event parameters into the cross-scale physical mapping functional. In the physical governing equations, when Previously, the stiffness remained at its initial value. ;exist After that, A time-smoothing operator is formed nearby. With spatial smoothing operator Local stiffness depression region co-modulated.

[0134] Dual-loop iterative inversion (corresponding to S300-S400): The system initiates dual-loop iteration under strong traffic noise interference. In the inner loop, the neural network... Under the boundary constraints, instead of blindly fitting noise, a smooth displacement field is output; in the outer loop, the mechanism parameters are calibrated in reverse using the measured macroscopic vibration residuals. This accurately yielded the macroscopic stiffness reduction ratio corresponding to the microcrack propagation at that moment.

[0135] Diagnostic output (corresponding to S500): The system ultimately outputs continuous spatiotemporal displacement and strain fields of the bridge, and accurately locates the damage at mid-span. At this point, the stiffness reduction ratio is .

[0136] Experimental verification and effect comparison: To verify the effectiveness of this invention, a numerical simulation experiment of the aforementioned bridge was constructed. Artificial parameters were added to the generated ideal theoretical continuous observation data (acceleration, strain). Gaussian white noise was used to simulate the harsh service environment.

[0137] Comparison method settings: Baseline Method A (pure data-driven method): directly fits noisy observation data using only a multilayer perceptron, without physical equation constraints.

[0138] Baseline Method B (Traditional Physical Neural Network Method, Standard PINN): Employs a method that includes a constant stiffness term. The partial differential equations are used to constrain the high-frequency dynamic boundary of unfused acoustic emission.

[0139] The method of the present invention (acoustic emission modulation of dual-loop closed-loop PINN).

[0140] Comparison of noise immunity and response reconstruction performance (with appendix) Figure 3 ): See attached document Figure 3 The figure shows the displacement reconstruction capabilities of different methods in an environment with 10% noise.

[0141] The horizontal axis in the figure represents the time of bridge monitoring operation, in seconds; the vertical axis represents the mid-span displacement response caused by the load on the mid-span of the bridge, in millimeters.

[0142] The dark gray vertical dashed line in the figure marks the acoustic emission trigger time (i.e., This time point corresponds to the instantaneous propagation of the microcrack and is the key dividing line for initiating the time-varying boundary modulation of local stiffness in this invention. The solid black line (unmarked) in the figure represents the actual physical state. That is, the real, noise-free displacement evolution trajectory of the bridge when it is subjected to moving loads and undergoes microscopic damage.

[0143] The black dashed line in the figure represents the prediction result of the purely data-driven method (baseline method A). It can be seen that, due to noise interference and the lack of physical equation constraints, the curve exhibits a large number of high-frequency spikes, resulting in severe overfitting and displaying violent non-physical oscillations.

[0144] The solid gray line in the figure represents the prediction result of the traditional physical neural network method (baseline method B). Using a solid gray line effectively avoids visual overlap with other black curves. Although this method has good noise resistance (smooth curves), it lacks a transient mechanism for sudden damage in its physical control equations, making it unable to track... Subsequently, the phase shift and amplitude changes caused by the sudden drop in stiffness resulted in the predicted trajectory deviating significantly from the actual physical state after the damage occurred, and the prediction results were severely delayed.

[0145] The black lines marked with hollow circles in the figure represent the prediction results of the method of this invention. Due to the accurate introduction of the time-varying physical boundary triggered by acoustic emission and the reliance on the double-loop optimization mechanism, the result highly coincides with the actual physical state. The reconstructed displacement field not only perfectly filters out noise but also accurately tracks the dynamic phase changes caused by damage, reducing the root mean square error (RMSE) of the displacement by an entire order of magnitude compared to the baseline method B.

[0146] Comparison of quantitative damage inversion results (with appendix) Figure 4 ): See attached document Figure 4 The figure illustrates the diagnostic and analytical capabilities of different methods for determining the location and extent of actual damage within a structure.

[0147] The horizontal axis in the figure represents the spatial coordinates of the bridge along its longitudinal span, in meters, covering the entire length of the bridge under test (0 to 30 meters); the vertical axis represents the ratio of the current remaining stiffness of the structure to its initial undamaged stiffness, i.e., the stiffness retention ratio. The lower the value, the more severe the damage in that area.

[0148] The solid black line (unmarked) in the figure represents the actual stiffness distribution. In the figure, it is shown as a distinct dip at the mid-span (spatial coordinate 15m), indicating that a substantial stiffness reduction has occurred at that point (the retention ratio is 0.85).

[0149] The black dotted line in the figure represents the inversion result of the traditional physical neural network method (baseline method B). Due to the lack of local damage features and spatial smoothing operator localization constraints, this method cannot identify localized minor damage. Its inversion result is a horizontal straight line, which means that the method incorrectly spreads the local damage in the middle of the bridge evenly across the entire bridge, causing the diagnosis to fail completely.

[0150] The black line marked with a cross in the figure represents the structural continuous stiffness distribution matrix reconstructed by the method of this invention. It can be seen that the method of this invention relies on the spatial prior conditions provided by acoustic emission (i.e., the spatial smoothing operator). In the outer ring inversion, the spatial coordinates of the damage occurrence were accurately located. The analysis successfully determined that the stiffness retention rate at this location was 0.85 (i.e., a reduction of 15%). The inverted stiffness concavity amplitude was completely consistent with the actual distribution.

Claims

1. A bridge multi-source data damage fusion diagnostic system driven by physical information, characterized in that, include: The sensor acquisition module is deployed on the bridge structure under test to acquire structural monitoring signals under the service condition of the structure; An edge computing node is physically connected to the sensing and acquisition module and is used to decouple the structure monitoring signal according to the physical properties of the signal, and extract the low-frequency continuous field observation set and the high-frequency transient discrete event set. The cloud computing platform establishes a data transmission link with the edge computing node for: A continuously differentiable cross-scale physical mapping functional containing the parameter vector of the microscopic physical mechanism is constructed. The high-frequency transient discrete event set is converted into time-varying local stiffness boundary conditions in the partial differential equation of structural dynamics through the cross-scale physical mapping functional, and the time-varying dynamic partial differential equation is reconstructed. Initialize the physical information neural network model, obtain the low-frequency continuous field observation set, the reconstructed time-varying dynamic partial differential equation, and the set physical prior boundary conditions, and combine them to construct the joint loss function of the physical information neural network model; The system performs a double-loop alternating optimization iteration. During the inner loop calculation, the network weights are updated through gradient descent, and during the outer loop calculation, the microscopic physical mechanism parameter vector is updated through residual backpropagation. When the convergence condition is met, the full-field response reconstruction result is output, and the converged microscopic physical mechanism parameter vector is called to calculate the local stiffness reduction matrix and damage space coordinates of the structure.

2. The system according to claim 1, characterized in that, The sensing and acquisition module is equipped with a heterogeneous sensor group and cross-scale clock synchronization hardware. The heterogeneous sensor group includes piezoelectric accelerometers, fiber optic strain gauges and piezoelectric ceramic acoustic emission probes. The edge computing node is used for: The cross-scale clock synchronization hardware allocates a unified global time reference for monitoring channels with different sampling frequencies. A time-domain interpolation alignment operation is performed on the original acceleration time series and the original strain time series with timestamps, and the low-frequency continuous field observation set is constructed by combining the spatial deployment coordinates; The edge computing node calculates the dynamic floating threshold in real time based on the ambient background noise, and extracts a transient time window when the voltage signal output by the piezoelectric ceramic acoustic emission probe crosses the dynamic floating threshold. The three-dimensional spatial coordinates of the occurrence of micro fracture events and the corresponding trigger times are calculated by combining the multi-channel trigger time difference, and the absolute energy is quantified by integral calculation within the transient time window. The high-frequency transient discrete event set is then constructed.

3. The system according to claim 1, characterized in that, The cloud computing platform constructs a continuously differentiable cross-scale physical mapping functional containing parameter vectors of microscopic physical mechanisms, including: A nonlinear mapping operator is established to characterize the amplitude of microscopic damage, wherein the nonlinear mapping operator contains a vector of microscopic physical mechanism parameters to be inverted and calibrated; Construct continuously differentiable spatial and temporal smoothing operators; A nonlinear dissipation relaxation term characterizing the nonlinear dissipation process is introduced. The nonlinear mapping operator, the spatial smoothing operator, the temporal smoothing operator, and the nonlinear dissipation relaxation term are fused to establish a cross-scale physical mapping functional for the time-varying local stiffness degradation of the output structure, which serves as the time-varying local stiffness boundary condition in the partial differential equation of the structural dynamics.

4. The system according to claim 3, characterized in that, The spatial smoothing operator is a globally continuous and differentiable operator constructed based on the exponential decay function, and introduces a spatial transition region parameter that characterizes the radius of influence of local stiffness degradation. The time smoothing operator is a globally continuous differentiable operator constructed based on the log-probability smoothing curve, and introduces a time transition region parameter characterizing the rate of stiffness degradation. The cloud computing platform is also used to: compile the cross-scale physical mapping functionals into custom physical operation nodes in the computation graph, and rely on automatic differentiation technology to enable the fitting gradient of the network weights and the physical residual gradient of the microscopic physical mechanism parameter vector to complete back propagation along continuous and smooth operator links.

5. The system according to claim 3, characterized in that, The cloud computing platform is used to construct the joint loss function of the physical information neural network model, which includes: A physical domain differential transformation is performed on the displacement field prediction value output by the physical information neural network model to construct a data fitting residual term that represents the sum of the mean square errors between the predicted field and the low-frequency continuous field observation set; The cross-scale physical mapping functional is embedded as a time-varying local stiffness boundary condition into the structural dynamics partial differential equation to construct a physical law constraint residual term based on the calculation of the mean square error integral of a random collocation set. A physical prior boundary penalty term is constructed based on the set physical prior boundary conditions; By employing a dimensionless dynamic weight allocation strategy, the data fitting residual term, the physical law constraint residual term, and the physical prior boundary penalty term are weighted and combined to establish a global joint loss function, which serves as the joint loss function.

6. The system according to claim 5, characterized in that, The cloud computing platform constructs a physical prior boundary penalty term based on the set physical prior boundary conditions, including: Two hard physical boundary conditions are set as the set physical a priori boundary conditions, wherein the time-varying local stiffness degradation of the structure, characterized by the time-varying local stiffness boundary conditions, is always greater than or equal to zero and does not exceed a set proportional limit of the initial distributed stiffness of the structure. Using the linear rectification function as a one-sided truncation operator, numerical constraint rules are applied to the time-varying local stiffness degradation of the structure. Once the inversion parameters exceed the limits, the positive residuals are activated and summed into the global joint loss function.

7. The system according to claim 5, characterized in that, The cloud computing platform updates network weights through gradient descent during the inner-loop computation process, including: When the inner loop iteration is started, the gradient tracking node of the automatic differentiation engine for the microscopic physical mechanism parameter vector is blocked, the microscopic physical mechanism parameter vector output by the current outer loop iteration step is frozen, and deterministic physical boundary conditions are established. Based on the frozen microphysical mechanism parameter vector and the current network weights, perform forward computation and use the automatic differentiation engine to calculate the partial derivative gradient matrix of the global joint loss function with respect to the network weights. Based on the preset inner loop learning rate step size parameter and the partial derivative gradient matrix, the gradient descent optimizer is invoked to iteratively update the network weights.

8. The system according to claim 7, characterized in that, The cloud computing platform updates the microscopic physical mechanism parameter vector through residual backpropagation during the outer loop calculation process, including: When the outer loop iteration is started, the network weights output by the inner loop convergence are frozen, and the physical information neural network model is transformed into a nonlinear mechanical analytical operator. Using the current microscopic physical mechanism parameter vector as the independent variable, forward computation is performed to extract the outer loop objective function, which is a weighted combination of the data fitting residual term and the physical prior boundary penalty term. Based on the backpropagation chain rule of the automatic differentiation engine, the partial derivative matrix of the outer loop objective function with respect to the microphysical mechanism parameter vector is calculated, and the gradient descent update operation is performed to obtain the microphysical mechanism parameter vector for the next iteration step.

9. The system according to claim 8, characterized in that, The cloud computing platform extracts the physical mechanism parameter update norm of adjacent outer loop iteration steps and the corresponding inner loop network weight update norm. When the maximum value of the two is less than or equal to the preset global joint tolerance threshold, it is determined that the convergence condition is met and the iteration is terminated. The global joint tolerance threshold is preset based on the double-precision floating-point truncation error of the computing hardware and the allowable accuracy of actual bridge structural damage identification. The cloud computing platform extracts the converged microscopic physical mechanism parameter vector and the high-frequency transient discrete event set, substitutes them into the nonlinear mapping operator characterizing the amplitude of microscopic damage to obtain a discrete set of stiffness reduction peaks, and combines them with a continuously differentiable spatial smoothing operator to reconstruct the continuous stiffness distribution matrix of the structure, which serves as the local stiffness reduction matrix of the structure. The spatial extreme value search algorithm is used to search for the minimum value region in the continuous stiffness distribution matrix of the structure where the stiffness value is lower than the set health threshold, and the three-dimensional coordinates of macroscopic damage are locked as the damage spatial coordinates. The set health threshold is a predetermined percentage of the initial undamaged distributed stiffness of the bridge under test.

10. A physical information-driven method for bridge multi-source data damage fusion diagnosis, characterized in that, The physical information-driven bridge multi-source data damage fusion diagnostic system according to any one of claims 1-9 includes the following steps: Edge computing nodes control the sensor acquisition modules deployed on the bridge structure under test to acquire monitoring signals under the service status of the structure; The edge computing node decouples the structure monitoring signal based on the physical properties of the signal, and extracts the low-frequency continuous field observation set and the high-frequency transient discrete event set. The cloud computing platform receives the low-frequency continuous field observation set and the high-frequency transient discrete event set, constructs a continuously differentiable cross-scale physical mapping functional containing the parameter vector of the microscopic physical mechanism, and converts the high-frequency transient discrete event set into time-varying local stiffness boundary conditions in the partial differential equation of structural dynamics through the cross-scale physical mapping functional, and reconstructs the time-varying dynamic partial differential equation. The cloud computing platform initializes the physical information neural network model, obtains the low-frequency continuous field observation set, the reconstructed time-varying dynamic partial differential equation, and the set physical prior boundary conditions, and combines them to construct the joint loss function of the physical information neural network model; The cloud computing platform performs a dual-loop alternating optimization iteration. During the inner loop calculation process, the network weights are updated through gradient descent, and during the outer loop calculation process, the parameter vector of the microscopic physical mechanism is updated through residual backpropagation. When the convergence condition is met, the cloud computing platform outputs the full-field response reconstruction results and calls the converged microscopic physical mechanism parameter vector to calculate the local stiffness reduction matrix and damage space coordinates of the structure.

Images