A method and system for assessing the health of a bridge

Abstract

The application discloses a kind of bridge and tunnel health assessment method and system, comprising: obtaining the multi-source heterogeneous detection data of bridge and tunnel component, based on global space-time base, space mapping and time sequence penetration are carried out to multi-source heterogeneous detection data, construct accurate alignment component-level multi-modal space-time database;Multi-source heterogeneous detection data is jointly represented using native multi-modal classifier, and comprehensive physical health feature vector is extracted, and the physical engine is dispatched using multi-modal large language model trigger function call mechanism, the physical degradation parameter and residual bearing capacity coefficient of component are calculated by physical information neural network and differentiable structure solver, and the corresponding structure disease grade is output;Based on physical degradation parameter and structure disease grade, feature tracing is carried out in pre-set engineering design specification and historical maintenance knowledge graph using graph retrieval enhancement technology, and bridge and tunnel health assessment result is generated.

Description

Technical Field

[0001] This invention relates to the field of bridge and tunnel health assessment technology, and more specifically, to a bridge and tunnel health assessment method and system. Background Technology

[0002] Current bridge and tunnel health assessment methods mostly rely on manual visual inspections, discrete non-destructive testing, and calculations based on empirical mechanical formulas under pre-set ideal conditions. These methods heavily depend on prior expert knowledge and are constrained by the silo effect of discrete and fragmented data, making it difficult to accurately quantify the deep physical coupling between apparent defects and internal damage in complex and ever-changing disease evolution scenarios. Existing solutions based on multimodal models and multi-agent technologies improve the automation level of disease identification by fusing multi-source data such as images and radar. However, forcibly converting high-dimensional physical signals into text communication easily leads to severe information loss. Simply relying on the probability generation mechanism of language models for assessment lacks underlying mechanical support, resulting in results that are difficult to conform to real-world physical laws. This leads to obvious black-box limitations and blind reasoning. Furthermore, existing methods struggle to trace and match physical calculation results with engineering design specifications and historical maintenance plans, resulting in assessment reports lacking engineering interpretability and decision-making persuasiveness.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a method and system for assessing the health of bridges and tunnels, which at least solves the technical problem of severe loss of high-dimensional physical feature information caused by cross-modal text conversion in the prior art.

[0005] According to one aspect of the present invention, in order to achieve the above-mentioned objective, a bridge and tunnel health assessment method is provided, comprising:

[0006] Acquire multi-source heterogeneous detection data of bridge and tunnel components, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on a global spatiotemporal base, and construct a precisely aligned component-level multimodal spatiotemporal database;

[0007] A native multimodal classifier is used to jointly represent multi-source heterogeneous detection data, extract comprehensive physical health feature vectors, and use a multimodal large language model to trigger function call mechanism to schedule the physical engine. The physical degradation parameters and residual bearing capacity coefficient of the components are calculated through physical information neural network and differentiable structural solver, and the corresponding structural defect level is output.

[0008] Based on physical degradation parameters and structural damage levels, graph retrieval enhancement technology is used to trace features in a pre-set engineering design specification and historical maintenance knowledge graph to generate bridge and tunnel health assessment results.

[0009] Furthermore, the multi-source heterogeneous detection data includes: one-dimensional ground-penetrating radar waveforms, two-dimensional disease images, and semi-structured historical detection text.

[0010] Furthermore, based on a global spatiotemporal base, spatial mapping and temporal penetration are performed on multi-source heterogeneous detection data, including:

[0011] By utilizing the spatial anchoring components within the global spatiotemporal base, a global building information model or geographic information system mapping engine is established. This maps the linear coordinates of ground-penetrating radar, the geographic labels of UAV images, and the manually detected station offsets to physical structure nodes in the same three-dimensional spatial coordinate system, thereby achieving absolute spatial anchoring.

[0012] Furthermore, by utilizing the time modeling component within the global spatiotemporal base, a deep learning architecture based on the state space model is introduced to perform long-sequence modeling of historical detection records spanning multiple years on physical structure nodes, extract temporal evolution patterns, and transform discrete detection data into a continuous health state evolution sequence.

[0013] Furthermore, a native multimodal classifier is used to jointly characterize the multi-source heterogeneous detection data, including:

[0014] The acquired one-dimensional ground-penetrating radar waveform is decomposed into multi-scale time and frequency using discrete wavelet transform. High-frequency detail coefficients that characterize abrupt changes are extracted. The high-frequency detail coefficients are then reconstructed and spliced ​​using feature dimensions to form a frequency domain sequence of the one-dimensional ground-penetrating radar waveform.

[0015] The frequency domain sequence of one-dimensional ground-penetrating radar waveform and the pixel matrix of two-dimensional disease image are directly mapped to a unified high-dimensional shared latent variable space through their respective feature extraction networks without natural language parsing and text encoding, generating multimodal original feature tensor identifiers.

[0016] In a high-dimensional latent variable space, a fusion module based on a cross-modal cross-attention mechanism is constructed to perform cross-modal cross-attention calculation, which enables the anomalous reflection signal of radar waves to interact with the crack features on the image surface at a low level, thereby uncovering the physical coupling relationship between internal defects and surface cracks.

[0017] Furthermore, the physics engine is scheduled using a multimodal large language model-triggered function call mechanism, including:

[0018] Encapsulate the differentiable structure solver as an external executable tool interface;

[0019] After analyzing the comprehensive physical health feature vector using a multimodal large language model, when it is determined that mechanical quantitative verification is required, the comprehensive physical health feature vector is transformed into an equivalent cross-sectional loss rate or a local stiffness degradation coefficient with physical meaning.

[0020] The equivalent cross-sectional loss rate or local stiffness degradation coefficient is called as a physical boundary condition and material property input into the underlying physical mechanism calculation network.

[0021] Furthermore, the physical degradation parameters and residual bearing capacity coefficients of the components are calculated using a physical information neural network and a differentiable structural solver, including:

[0022] Construct a cascaded physics engine that includes an upstream physical information neural network and a downstream differentiable structure solver;

[0023] In the upstream prediction stage, the physical information neural network receives the physical boundary conditions transmitted by the multimodal large language model, predicts the structural stiffness reduction coefficient of the output component, and adds the preset force balance equation as a physical penalty term to the loss function of the physical information neural network, applying gradient penalty to the prediction features that violate the common sense of mechanics.

[0024] In the downstream calculation stage, the structural stiffness reduction factor is used as a known condition and input into the downstream end-to-end differentiable structural solver for forward finite element calculation. The resistance effect and action effect of the section under the current load combination are calculated, and a quantitative residual bearing capacity coefficient is output. Furthermore, the differentiable structural solver supports the backpropagation of the calculated gradient to support the joint training of the upstream physical information neural network.

[0025] Furthermore, the output structural defect level includes:

[0026] By introducing structural stiffness reduction factor, stiffness reduction minimum threshold and code allowable bearing capacity lower limit threshold, a rigid classification judgment rule is established.

[0027] When the calculation results satisfy the condition that the structural stiffness reduction coefficient is less than or equal to the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold of the allowable bearing capacity specified in the code, it is determined that the structure only has surface non-structural damage and the bearing capacity coefficient has no attenuation, and the output is a first-level defect.

[0028] When the calculation results satisfy the condition that the structural stiffness reduction factor is greater than the minimum stiffness reduction threshold and the remaining bearing capacity factor is greater than or equal to the lower limit threshold of the allowable bearing capacity specified in the code, it is determined that the structural durability index is damaged but the remaining bearing capacity has not fallen below the working threshold, and the output is a level two defect.

[0029] When the calculation result satisfies the condition that the remaining bearing capacity coefficient is within the range between the lower limit threshold of the allowable bearing capacity in the code and the preset proportional threshold of the lower limit threshold of the allowable bearing capacity in the code, it is determined that the weakening of the effective section of the structure has led to a reduction in the remaining bearing capacity, and the output is a level three defect.

[0030] When the calculation result satisfies the preset proportional threshold that the remaining bearing capacity coefficient is less than the lower limit threshold of the allowable bearing capacity in the code, it is determined that the remaining bearing capacity of the structure has decreased and there is a risk of component instability or fracture, and the output is a level four defect.

[0031] Furthermore, graph retrieval enhancement technology is used to perform feature tracing within a pre-built engineering design specification and historical maintenance knowledge graph, including:

[0032] Extract the structural bearing capacity attenuation index from the output of the differentiable structure solver, as well as the topological information on the stress importance of the component in the global spatiotemporal base;

[0033] After normalizing the numerical physical parameters, they are input into a multilayer perceptron and encoded into physical feature vectors. Then, a graph convolutional network is used to aggregate and encode the topological importance information of the graph structure to generate topological feature vectors.

[0034] The physical feature vector and the topological feature vector are concatenated, and then nonlinearly mapped to the same metric space as the pre-trained knowledge graph nodes through an alignment projection network to generate a high-dimensional joint query vector.

[0035] The high-dimensional joint query vector is matched in a subgraph of a knowledge graph containing industry standard text, historical reinforcement drawings, and expert decision-making logic. The regulatory restrictions and treatment experience nodes of similar diseases that match the current attenuation index are retrieved and then fed back as contextual prompts to the multimodal large language model.

[0036] Furthermore, bridge and tunnel health assessment results are generated, including:

[0037] Construct a structured prompt word instruction template that includes role definition, feature slots, tiered execution instructions, and output format constraints;

[0038] The multimodal original feature tensor identifier, the quantitative physical degradation parameters and residual bearing capacity coefficients output by the differentiable structure solver, and the specification clauses recalled by the graph retrieval enhancement module are dynamically filled into the corresponding slots of the structured prompt word instruction template to assemble and generate high-dimensional context prompt words.

[0039] High-dimensional contextual prompts are input into a multimodal large language model, enabling the multimodal large language model to perform a step-by-step reasoning chain consisting of disease quantification identification, mechanical mechanism deduction, and standard requirement benchmarking, and explicitly generate structured labels corresponding to each reasoning step in the final output;

[0040] The output includes a digital decision dossier for structural health monitoring containing quantitative evidence of mechanical attenuation, citations of clause codes, and specific construction and reinforcement techniques.

[0041] According to one embodiment of the present invention, a bridge and tunnel health assessment system is also provided, characterized in that it includes:

[0042] The acquisition module is used to acquire multi-source heterogeneous detection data of bridge and tunnel components, and to perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on a global spatiotemporal base to construct a precisely aligned component-level multimodal spatiotemporal database.

[0043] The output module is used to jointly characterize the multi-source heterogeneous detection data using a native multimodal classifier, extract a comprehensive physical health feature vector, and use a multimodal large language model to trigger function call mechanism to schedule the physical engine. It then calculates the physical degradation parameters and remaining bearing capacity coefficient of the component through a physical information neural network and a differentiable structure solver, and outputs the corresponding structural defect level.

[0044] The generation module is used to generate bridge and tunnel health assessment results by using graph retrieval enhancement technology to trace features in a pre-set engineering design specification and historical maintenance knowledge graph based on the physical degradation parameters and the structural damage level.

[0045] In this embodiment of the invention, a native multimodal classifier directly maps multi-source heterogeneous detection data to a high-dimensional shared latent variable space for cross-modal interactive fusion without text encoding, eliminating feature loss caused by traditional text conversion and fully preserving the physical coupling information of apparent defects and internal damage. A multimodal large language model serves as a cognitive central trigger function call mechanism to drive the physical engine of the embedded physical information neural network and differentiable structure solver to perform quantitative mechanical calculations, achieving deep collaboration between AI cognitive reasoning and underlying mechanical mechanism calculations, overcoming the black-box defect of pure data-driven models that violate physical laws. Furthermore, graph retrieval enhancement technology accurately matches source-tracing normative clauses and historical treatment experience in the knowledge graph, combined with hard-level grading of disease severity, ultimately generating assessment results with reference to normative clauses and mechanical interpretation, improving the physical rigor, standard compliance, and decision credibility of bridge and tunnel structural health assessments, thereby solving the technical problem of severe loss of high-dimensional physical feature information caused by cross-modal text conversion in existing technologies. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0047] Figure 1 This is a flowchart of a bridge and tunnel health assessment method according to one embodiment of the present invention;

[0048] Figure 2This is a flowchart of a bridge and tunnel health assessment method according to another embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the feature fusion process based on native multimodal classification and cross-modal interaction in a bridge and tunnel health assessment method according to one embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the physical and mechanical calculation process of a bridge and tunnel health assessment method according to one embodiment of the present invention, which embeds PINN and a differentiable structure solver.

[0051] Figure 5 This is a schematic diagram of the standardized tracing and decision generation process based on graph retrieval enhancement and thought chain in a bridge and tunnel health assessment method according to one embodiment of the present invention.

[0052] Figure 6 This is a structural block diagram of a bridge and tunnel health assessment device according to one embodiment of the present invention. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] According to an embodiment of the present invention, a bridge and tunnel health assessment embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0056] This method embodiment can be executed in an electronic device or similar computing device that includes a memory and a processor. Taking operation on a terminal as an example, the terminal may include one or more processors (processors may include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), digital signal processing (DSP) chips, microcontroller units (MCUs), field-programmable gate arrays (FPGAs), neural network processors (NPUs), tensor processors (TPUs), artificial intelligence (AI) type processors, etc.) and a memory for storing data. Optionally, the terminal may also include transmission devices, input / output devices, and display devices for communication functions. Those skilled in the art will understand that the above structural description is merely illustrative and does not limit the structure of the terminal. For example, the terminal may include more or fewer components than described above, or have a different configuration than described above.

[0057] The memory can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the bridge and tunnel health assessment in this embodiment of the invention. The processor executes various functional applications and data processing by running the computer program stored in the memory, thereby realizing the aforementioned bridge and tunnel health assessment. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to a mobile terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0058] The transmission device is used to receive or send data via a network. Specific examples of the network mentioned above may include a wireless network provided by the mobile terminal's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0059] Display devices can be, for example, touchscreen liquid crystal displays (LCDs) and touch displays (also referred to as "touchscreens" or "touch displays"). The LCD allows users to interact with the user interface of the mobile terminal. In some embodiments, the mobile terminal has a graphical user interface (GUI), which allows users to interact with the GUI through finger contact and / or gestures on a touch-sensitive surface. Optional human-computer interaction functions include: creating web pages, drawing, word processing, creating electronic documents, playing games, video conferencing, instant messaging, sending and receiving emails, call interfaces, playing digital video, playing digital music, and / or web browsing, etc. Executable instructions for performing the above human-computer interaction functions are configured / stored in one or more processor-executable computer program products or readable storage media.

[0060] Figure 1 This is a flowchart of a bridge and tunnel health assessment according to one embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0061] Step S110: Obtain multi-source heterogeneous detection data of bridge and tunnel components. Based on a global spatiotemporal base, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data to construct a precisely aligned component-level multimodal spatiotemporal database. The specific content is as follows:

[0062] In step S110, multi-source heterogeneous detection data of bridge and tunnel components are first acquired. Because bridge and tunnel structures are subject to multiple uncertainties such as dynamic loads and material aging in actual operating environments, data acquired by a single type of sensor cannot fully reflect the true service status of the structure. Therefore, complementary information needs to be obtained from multiple detection methods. Specifically, the acquired multi-source heterogeneous detection data includes at least: one-dimensional waveform data collected by ground-penetrating radar equipment, which can penetrate the concrete overlay and reflect hidden defects such as internal steel reinforcement corrosion, voids, and detachment; two-dimensional defect image data collected by drones or handheld camera equipment, which can visually record surface damage such as crack distribution, spalling, and water seepage; and semi-structured historical detection text data formed by manual inspections or historical inspection reports, which includes qualitative descriptions and quantitative records of each inspection. These three types of data differ significantly in physical properties, data format, and information dimensions, and are typical examples of cross-modal heterogeneous data.

[0063] After acquiring multi-source heterogeneous detection data, in order to eliminate coordinate deviations caused by inconsistent spatial reference systems between different sensor data and overcome temporal discontinuities in historical detection records due to non-uniform sampling, it is necessary to perform spatial mapping and temporal penetration on the aforementioned multi-source heterogeneous detection data based on a global spatiotemporal base. The global spatiotemporal base refers to a functional multidimensional data fusion architecture or hardware / software co-processing module that integrates a three-dimensional absolute spatial geometric reference frame and a one-dimensional continuous-time state evolution sequence, serving as a unified spatiotemporal coordinate system for subsequent cross-modal data alignment and state inference. Structurally, this global spatiotemporal base is composed of a spatial anchoring component and a temporal modeling component.

[0064] At the spatial mapping level, a global building information model or geographic information system mapping engine is established using the spatial anchoring component within the global spatiotemporal base. The linear coordinates recorded by ground-penetrating radar, the geographic location tags carried by UAVs, and the station offsets recorded during manual inspection are all input into this spatial anchoring component. Based on the coordinate mapping principle, this component maps these heterogeneous local sensing coordinates to physical structure nodes in the same three-dimensional spatial coordinate system through spatial transformation, achieving absolute spatial anchoring. After this spatial mapping process, heterogeneous data from different detection methods obtain a unified coordinate reference in the three-dimensional spatial dimension, enabling precise spatial correspondence between apparent defect images, internal radar reflection signals, and historical inspection records at the same physical location.

[0065] At the temporal penetration level, a deep learning architecture based on a state-space model is introduced using the temporal modeling component within the global spatiotemporal base. This architecture performs long-sequence modeling of historical detection records spanning multiple years on spatially anchored physical structure nodes. By transforming discrete historical health feature sequences into the form of continuous-time system state equations, the state-space model can capture the evolution of structural degradation trends at different time scales while maintaining linear computational complexity. This eliminates temporal discontinuity errors caused by unequal detection intervals and sparse data records, transforming the originally discrete and non-uniform historical detection records into a continuous and smooth health state evolution sequence.

[0066] After the aforementioned spatial mapping and temporal penetration processing, the multi-source heterogeneous data from ground-penetrating radar, disease images, and historical detection texts achieve precise alignment across the entire lifecycle, ensuring location traceability, temporal continuity, and state comparability. The system thus constructs a precisely aligned component-level multimodal spatiotemporal database. This database serves as a unified data foundation for subsequent cross-modal feature fusion and physical mechanics calculations, providing high-quality input with both spatial and temporal alignment for the joint representation of the native multimodal classifier.

[0067] Step S120: A native multimodal classifier is used to jointly represent the multi-source heterogeneous detection data, extract the comprehensive physical health feature vector, and the physical engine is scheduled using the multimodal large language model trigger function call mechanism. The physical degradation parameters and residual bearing capacity coefficient of the components are calculated through the physical information neural network and the differentiable structural solver, and the corresponding structural defect level is output. The specific content is as follows:

[0068] In step S120, after completing global spatiotemporal mapping and long-term time-series penetration to obtain a highly aligned component-level multimodal spatiotemporal database, the system inputs this database as the sole data source into the native multimodal classifier to eliminate the high-dimensional feature loss caused by converting physical signals into text descriptions before inference in traditional methods. The dedicated native multimodal classifier refers to a deep neural network architecture that directly performs multidimensional shared space mapping and underlying mathematical alignment on multi-source heterogeneous original feature tensors end-to-end. Its essential difference from conventional multimodal fusion networks lies in the fact that conventional multimodal fusion typically employs a discrete heterogeneous model, first converting visual signals and radar waveforms into natural language text descriptions through independent image recognition models or signal analysis software, and then feeding the text into a large model. This process of converting physical features into textual language causes severe loss of high-dimensional information features.

[0069] In the specific process of joint representation using a native multimodal classifier, the system first preprocesses the one-dimensional ground-penetrating radar waveform, performing multi-scale time-frequency decomposition on the original time-domain waveform sequence through Haar discrete wavelet transform. In the multi-scale decomposition system, homogeneous concrete with normal structure often exhibits a smooth background signal, while interfaces with defects such as internal voids, delamination, and steel corrosion can cause transient abrupt scattering of radar waves. Therefore, after decomposition, the system discards low-frequency approximation coefficients and specifically calculates and extracts high-frequency detail coefficients representing abrupt anomalies. The extracted high-frequency detail coefficients, rich in internal defect reflection characteristics, are then reconstructed and spliced ​​in the feature dimension to form the final frequency domain sequence of the one-dimensional ground-penetrating radar waveform. Subsequently, the system directly maps the frequency domain sequence of the one-dimensional ground-penetrating radar waveform, which has completed absolute spatiotemporal mapping, and the pixel matrix of the two-dimensional defect image to a unified high-dimensional shared latent variable space through their respective feature extraction networks without natural language parsing or text encoding, generating multimodal original feature tensor identifiers.

[0070] Building upon this foundation, the system constructs a fusion module based on a cross-modal cross-attention mechanism within a high-dimensional latent variable space. Specifically, the visual feature matrix extracted from the image and the waveform feature matrix extracted from the radar are transformed by learnable projection weight matrices to calculate the query matrix, key matrix, and value matrix. Through mathematical constraints, cross-modal interactive computation of visual surface crack features and radar internal anomalous reflection wave features is achieved at the underlying level. This process enables low-level mathematical interaction between the anomalous reflection signal matrix of the radar wave and the crack feature tensor of the image surface, allowing the system to autonomously discover the complex implicit physical coupling between internal defects and surface cracks. Ultimately, it outputs a multimodal integrated physical health feature vector that eliminates the dimensionality reduction loss of high-dimensional features.

[0071] After extracting the comprehensive physical health feature vector that integrates apparent defects and internal damage, the system uses a multimodal large language model as the cognitive and routing hub to receive this feature vector, identify and assess the requirements, and trigger a function call mechanism to schedule the backend physics engine. Specifically, the system first encapsulates the differentiable structure solver as an external executable tool interface. After the multimodal large language model analyzes the comprehensive physical health feature vector, when it is determined that mechanical quantitative verification is required, the built-in projection mapping layer automatically converts the nonlinear components in the feature vector into physically meaningful equivalent cross-sectional loss rates or local stiffness degradation coefficients. Subsequently, the external executable tool interface is called, and the converted degradation coefficients are input as physical boundary conditions and material properties into the underlying physical mechanism calculation network.

[0072] In the physics engine, the system employs an end-to-end cascaded architecture that deeply couples deep learning with classical mechanics, clearly divided into an upstream physical information neural network and a downstream differentiable structural solver. In the upstream prediction stage, the physical information neural network receives physical boundary conditions from a multimodal large language model and predicts the structural stiffness reduction coefficients of the output components. To ensure that the AI ​​decoding parameters strictly conform to real-world civil physics laws, the system constructs the core physical penalty term from the classical governing equations extracted from the "General Specifications for Highway Bridge and Culvert Design," forming a joint loss function with the data-driven term. This applies a high gradient penalty to prediction features that violate mechanical equilibrium, ensuring the physical interpretability of the output stiffness reduction coefficients. In the downstream calculation stage, the structural stiffness reduction coefficients decoded by the aforementioned physical constraints are seamlessly input into the downstream end-to-end differentiable structural solver for forward finite element calculations. This calculates the resistance and action effects of the section under the current most unfavorable load combination in real time, ultimately outputting a quantitative residual bearing capacity coefficient. Furthermore, since the solver with a differentiable structure is completely differentiable, the physical calculation process is no longer a black box that blocks gradients. The residual gradients generated by the downstream calculations can flow back into the upstream physical information neural network to update the network weights, thus realizing a deep closed loop between the pure data-driven model and the hard-core physical calculation rules.

[0073] Based on the calculated physical degradation parameters and residual bearing capacity coefficient, the system performs a four-level automated rigid classification in strict accordance with the rigid thresholds of engineering standards. The rigid classification criteria incorporate structural stiffness reduction coefficients, minimum stiffness reduction thresholds, and the lower limit threshold of allowable bearing capacity specified in the code. The specific judgment process is as follows: When the calculation result satisfies that the structural stiffness reduction coefficient is less than or equal to the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold of the allowable bearing capacity specified in the code, the structure is judged to have only surface non-structural damage and the bearing capacity coefficient has no attenuation, and the output is Level 1 defect; when the calculation result satisfies that the structural stiffness reduction coefficient is greater than the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold of the allowable bearing capacity specified in the code, the structure's durability index is judged to be damaged but the remaining bearing capacity has not substantially fallen below the working threshold, and the output is Level 2 defect; when the calculation result satisfies that the remaining bearing capacity coefficient is in the range between the lower limit threshold of the allowable bearing capacity specified in the code and the preset proportional threshold of the lower limit threshold of the allowable bearing capacity specified in the code, the effective section of the structure is judged to be weakened, resulting in a clear reduction in the remaining bearing capacity, and the output is Level 3 defect; when the calculation result satisfies that the remaining bearing capacity coefficient is less than the preset proportional threshold of the lower limit threshold of the allowable bearing capacity specified in the code, the remaining bearing capacity of the structure is judged to be significantly reduced and there is a risk of component instability or fracture, and the output is Level 4 defect. The above four-level hard physical classification ensures high-precision alignment between AI deduction and classical mechanics mechanisms.

[0074] Step S130: Based on physical degradation parameters and structural damage levels, feature tracing is performed using graph retrieval enhancement technology within a pre-set engineering design specification and historical maintenance knowledge graph to generate bridge and tunnel health assessment results. The specific content is as follows:

[0075] In step S130, after obtaining the quantitative physical degradation parameters and rigid structural defects level of the components through cascaded calculations of the upstream physical information neural network and the downstream differentiable structural solver, the system needs to transform the above-mentioned cold numerical calculation results into evaluation conclusions with engineering specifications and mechanical interpretations. To this end, the system uses graph retrieval enhancement technology to perform precise feature tracing in a pre-set engineering design specification and historical maintenance knowledge graph. First, in the knowledge base construction stage, the system uses industry standard texts, material structure relationships, historical reinforcement design drawings, and expert maintenance rules as entity nodes, and subordinate and logical associations as edges to construct a multi-dimensional knowledge graph of the entire life cycle of bridges and tunnels, providing underlying semantic support for subsequent retrieval. This knowledge graph covers the knowledge system of core design specification clauses such as the "General Specifications for Highway Bridge and Culvert Design", typical defect treatment schemes, and expert decision-making logic.

[0076] After completing the pre-construction of the knowledge graph, the system enters the feature encoding and joint query vector generation stage. Specifically, the system first extracts the structural bearing capacity attenuation index output by the differentiable structure solver, including the stiffness reduction coefficient and the residual bearing capacity coefficient, and simultaneously extracts the topological information of the force importance of the component in the global spatiotemporal base. For numerical physical parameters, the system performs maximum-minimum normalization processing and then inputs them into a multilayer perceptron for nonlinear up-dimensional encoding to generate physical feature vectors. For graph structure topological importance information, the system uses a graph convolutional neural network to aggregate the force transmission features of first-order and second-order neighbor nodes to generate topological feature vectors. Subsequently, the system concatenates the physical feature vectors and topological feature vectors, and introduces a learnable alignment projection network to nonlinearly map them to the same metric space as the pre-trained knowledge graph nodes, generating high-dimensional joint query vectors. This alignment projection network is optimized using a contrastive learning loss function during the training phase to narrow the distance between the physical vectors of severely attenuated bearing capacity and the corresponding maintenance specification text vectors in the feature space, eliminating the semantic gap between the engineering physical space and the semantic space of the knowledge graph.

[0077] After the above mapping, the high-dimensional joint query vector has the same mathematical basis and metric as the knowledge graph nodes. The system then performs precise subgraph similarity matching on the high-dimensional joint query vector in the knowledge graph, which contains industry standard text, historical reinforcement drawings, and expert decision-making logic, aiming to locate knowledge clusters highly correlated with the current structural physical attenuation index. Specifically, the system uses cosine similarity to calculate high-scoring matching nodes in the knowledge graph, recalls the main regulatory provisions and high-scoring historical treatment experience nodes that precisely correspond to the current structural bearing capacity attenuation index, and reorganizes the recalled regulatory restrictions and optimal treatment experience nodes for similar defects into a logically rigorous structured context, eliminating the fragmentation of the original data and providing strong legal and engineering experience support for subsequent decision generation.

[0078] After completing graph feature retrieval and context reorganization, the system drives the multimodal large language model's thought chain reasoning engine to execute generative decision-making. The system pre-constructs a structured prompt word instruction template containing role definitions, feature slots, tiered execution instructions, and output format constraints. The system dynamically fills the corresponding slots in the structured prompt word instruction template with the multimodal original feature tensor identifiers, the quantitative physical degradation parameters and residual bearing capacity coefficients output by the differentiable structure solver, and the specification clauses recalled by the graph retrieval enhancement module, assembling and generating high-dimensional context prompt words. These high-dimensional context prompt words are then input into the multimodal large language model, forcing it to execute a tiered thought chain reasoning process consisting of disease quantification identification, mechanical mechanism deduction, and specification requirement alignment. The reasoning process strictly follows the logical tree sequence for sequential analysis: First, it characterizes and quantitatively identifies defects, clearly indicating detection findings such as crack width and radar anomaly depth; second, it determines the underlying mechanical attenuation, listing the calculated stiffness reduction coefficient and residual bearing capacity coefficient; finally, it compares against regulatory requirements and thresholds, referencing specific legal provisions for verification, and explicitly generates structured labels for each reasoning step in the final output. The final automatic output includes a digital decision-making dossier for structural health monitoring containing quantitative evidence of mechanical attenuation, precise clause coding references, and specific construction and reinforcement techniques. This achieves a closed-loop generation from raw physical monitoring parameters to a maintenance decision plan with high standardization, traceability, and operational guidance, significantly improving the engineering credibility and decision-making persuasiveness of the assessment results.

[0079] Based on steps S110 to S130 above, in this embodiment of the invention, multi-source heterogeneous detection data is directly mapped to a high-dimensional shared latent variable space for cross-modal interactive fusion without text encoding, using a native multimodal classifier. This eliminates feature loss caused by traditional text conversion and fully preserves the physical coupling information of apparent defects and internal damage. A multimodal large language model is used as a cognitive center trigger function call mechanism to drive the physical engine of the embedded physical information neural network and differentiable structure solver to perform quantitative mechanical calculations. This achieves deep collaboration between AI cognitive reasoning and underlying mechanical mechanism calculations, overcoming the black-box defect of pure data-driven models that violate physical laws. Furthermore, graph retrieval enhancement technology accurately matches source-tracing normative clauses and historical treatment experience in the knowledge graph, combined with hard-level grading of disease severity, ultimately generating assessment results with reference to normative clauses and mechanical interpretation basis. This improves the physical rigor, standard compliance, and decision credibility of bridge and tunnel structural health assessment, thereby solving the technical problem of severe loss of high-dimensional physical feature information caused by cross-modal text conversion in existing technologies.

[0080] The bridge and tunnel health assessment in this embodiment of the invention uses multi-source heterogeneous detection data, including: one-dimensional ground-penetrating radar waveforms, two-dimensional defect images, and semi-structured historical detection text.

[0081] Furthermore, based on the global spatiotemporal base, spatial mapping and temporal penetration of multi-source heterogeneous detection data are performed, including: using the spatial anchoring component inside the global spatiotemporal base to establish a global building information model or geographic information system mapping engine, mapping the linear coordinates of ground penetrating radar, the geographic labels of UAV images, and the station offsets of manual detection to physical structure nodes under the same three-dimensional spatial coordinate system to achieve absolute spatial anchoring; and using the temporal modeling component inside the global spatiotemporal base to introduce a deep learning architecture based on the state space model to perform long-sequence modeling of historical detection records spanning multiple years on physical structure nodes, extracting temporal evolution patterns, and transforming discrete detection data into a continuous health state evolution sequence.

[0082] Furthermore, a native multimodal classifier is used to jointly characterize the multi-source heterogeneous detection data, including: performing multi-scale time-frequency decomposition on the acquired one-dimensional ground-penetrating radar waveform through discrete wavelet transform, extracting high-frequency detail coefficients representing abrupt anomalies, reconstructing and splicing the high-frequency detail coefficients in terms of feature dimensions to form a frequency domain sequence of the one-dimensional ground-penetrating radar waveform; mapping the frequency domain sequence of the one-dimensional ground-penetrating radar waveform and the pixel matrix of the two-dimensional disease image directly to a unified high-dimensional shared latent variable space through their respective feature extraction networks without natural language parsing and text encoding, generating multimodal original feature tensor identifiers; in the high-dimensional latent variable space, a fusion module based on a cross-modal cross-attention mechanism is constructed to perform cross-modal cross-attention calculation, prompting the anomalous reflection signal of the radar wave to interact with the crack features on the image surface at a low level, and discovering the physical coupling relationship between internal defects and surface cracks.

[0083] Furthermore, the physics engine is scheduled using a multimodal large language model trigger function call mechanism, including: encapsulating the differentiable structure solver as an external executable tool interface; after the multimodal large language model analyzes the comprehensive physical health feature vector and determines that mechanical quantitative verification is required, converting the comprehensive physical health feature vector into a physically meaningful equivalent cross-sectional loss rate or local stiffness degradation coefficient; calling the external executable tool interface to input the equivalent cross-sectional loss rate or local stiffness degradation coefficient as physical boundary conditions and material properties into the underlying physical mechanism calculation network.

[0084] Furthermore, the physical degradation parameters and residual bearing capacity coefficients of the components are calculated through a physical information neural network and a differentiable structural solver. This includes: constructing a cascaded physical engine containing an upstream physical information neural network and a downstream differentiable structural solver; in the upstream prediction stage, the physical information neural network receives the physical boundary conditions transmitted by the multimodal large language model, predicts and outputs the structural stiffness reduction coefficient of the component, and adds the preset force balance equation as a physical penalty term to the loss function of the physical information neural network, applying gradient penalties to prediction features that violate mechanical common sense; in the downstream calculation stage, the structural stiffness reduction coefficient is used as a known condition and input to the downstream end-to-end differentiable structural solver for forward finite element calculation, calculating the resistance effect and action effect of the section under the current load combination, outputting a quantitative residual bearing capacity coefficient, and the differentiable structural solver supports the backpropagation of the calculated gradient to support the joint training of the upstream physical information neural network.

[0085] Furthermore, the structural distress level is output, including: introducing a structural stiffness reduction factor, a minimum stiffness reduction threshold, and a lower limit threshold for allowable bearing capacity as specified in the code, establishing a rigid classification judgment rule; when the calculation result satisfies that the structural stiffness reduction factor is less than or equal to the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold for allowable bearing capacity as specified in the code, the structure is judged to have only surface non-structural damage and no reduction in bearing capacity coefficient, and the output is level one distress; when the calculation result satisfies that the structural stiffness reduction factor is greater than the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold for allowable bearing capacity as specified in the code, the structure is judged to have only surface non-structural damage and no reduction in bearing capacity coefficient, and the output is level one distress; when the calculation result satisfies that the structural stiffness reduction factor is greater than the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold for allowable bearing capacity as specified in the code, the output is level one distress. When the load-bearing capacity is below the lower limit threshold, the structure's durability index is deemed damaged, but the remaining load-bearing capacity has not fallen below the working threshold, and the output is classified as Level II distress. When the calculation result satisfies that the remaining load-bearing capacity coefficient is within the range between the lower limit threshold of the standard allowable load-bearing capacity and the preset proportional threshold of the lower limit threshold of the standard allowable load-bearing capacity, the effective section of the structure is deemed weakened, resulting in a reduction in the remaining load-bearing capacity, and the output is classified as Level III distress. When the calculation result satisfies that the remaining load-bearing capacity coefficient is less than the preset proportional threshold of the lower limit threshold of the standard allowable load-bearing capacity, the remaining load-bearing capacity of the structure is deemed reduced and there is a risk of component instability or fracture, and the output is classified as Level IV distress.

[0086] Furthermore, graph retrieval enhancement technology is used to perform feature tracing in a pre-set engineering design specifications and historical maintenance knowledge graph. This includes: extracting the structural bearing capacity attenuation index output by the differentiable structure solver, as well as the topological information of the component's stress importance in the global spatiotemporal base; normalizing the numerical physical parameters and inputting them into a multilayer perceptron to encode them into physical feature vectors, and using a graph convolutional network to aggregate and encode the graph structure's topological importance information to generate topological feature vectors; concatenating the physical feature vectors and topological feature vectors, and using an alignment projection network to nonlinearly map them to the same metric space as the pre-trained knowledge graph nodes to generate a high-dimensional joint query vector; performing subgraph matching on the high-dimensional joint query vector in a knowledge graph containing industry standard text, historical reinforcement drawings, and expert decision-making logic to retrieve regulatory restrictions and treatment experience nodes for similar defects that match the current attenuation index, and feeding them back as contextual prompts to the multimodal large language model.

[0087] Furthermore, the generation of bridge and tunnel health assessment results includes: constructing a structured prompt word instruction template containing role definitions, feature slots, tiered execution instructions, and output format constraints; dynamically filling the corresponding slots of the structured prompt word instruction template with multimodal original feature tensor identifiers, quantitative physical degradation parameters and residual bearing capacity coefficients calculated by the differentiable structural solver, and standard clauses recalled by the graph retrieval enhancement module, and assembling them to generate high-dimensional contextual prompt words; inputting the high-dimensional contextual prompt words into the multimodal large language model, enabling the multimodal large language model to perform tiered thinking chain reasoning consisting of disease quantification identification, mechanical mechanism deduction, and standard requirement benchmarking, and explicitly generating structured labels corresponding to each reasoning step in the final output; and outputting a digital decision dossier for structural health testing containing quantitative mechanical attenuation evidence, clause code references, and specific construction and reinforcement processes.

[0088] Another method for assessing the health of bridges and tunnels according to one embodiment of the present invention, such as Figures 2-5 As shown, the method includes the following steps:

[0089] Step 201: Acquire multi-source heterogeneous detection data of bridge and tunnel components, perform three-dimensional spatial absolute mapping and long-sequence temporal penetration based on a global spatiotemporal base, and construct a precisely aligned multimodal spatiotemporal database. In this embodiment of the invention, the global spatiotemporal base refers to a multi-dimensional data fusion architecture that integrates a three-dimensional absolute spatial geometric reference frame and a one-dimensional continuous-time state evolution sequence, used as a unified spatiotemporal coordinate system for cross-modal data alignment and state inference;

[0090] By establishing a global BIM or GIS mapping engine, the local coordinates of heterogeneous sensing data such as ground-penetrating radar and UAVs are mapped to global three-dimensional coordinates; and a state space model is introduced to perform continuous time modeling of the long-period health feature sequence of structural nodes, capturing degradation trends with linear complexity and eliminating temporal fault errors caused by discrete detection.

[0091] Step 202: Employing an end-to-end native multimodal classifier, this approach directly maps and aligns the multi-source heterogeneous original feature tensors in a multi-dimensional shared space at the underlying layer, enabling cross-modal interaction and extracting comprehensive physical health features. Unlike conventional discrete methods that first convert visual signals and radar waveforms into text before fusing them, this architecture avoids the severe loss of high-dimensional information caused by converting physical features to text.

[0092] The frequency domain feature matrix of one-dimensional ground-penetrating radar and the visual feature matrix of two-dimensional disease images are directly mapped to a unified high-dimensional latent variable space. A fusion network based on cross-modal cross-attention mechanism is constructed to perform low-level cross-modal feature fusion. Through mathematical constraints, high-fidelity physical coupling between external apparent cracks and internal radar anomaly waves is directly achieved at the low level, and a multi-modal comprehensive physical health feature vector is output.

[0093] like Figure 3 As shown, the frequency domain feature matrix of the one-dimensional ground-penetrating radar and the visual feature matrix of the two-dimensional disease image are extracted. Specifically, the acquired one-dimensional ground-penetrating radar waveform is processed using Haar Discrete Wavelet Transform (DWT) to transform the original time-domain waveform sequence. Multi-scale time-frequency decomposition is performed. The original one-dimensional time-domain waveform is discretized and filtered using Haar wavelet basis functions. In the multi-scale decomposition system, the background signal of homogeneous concrete is smooth, while defects such as voids, gaps, and steel corrosion will cause transient abrupt scattering. High-frequency detail coefficients of layers 1 to L are specifically extracted to form a final frequency domain sequence rich in defect reflection characteristics.

[0094] Visual features of ground-penetrating radar frequency domain sequences and disease images are extracted and directly mapped to a unified high-dimensional latent variable space to generate native multimodal identifiers without text intermediary. A cross-modal cross-attention network is constructed, and the interaction between the visual query matrix and the radar key-value matrix at the bottom layer is used to discover the physical coupling relationship between apparent cracks and internal abnormal reflected waves. Finally, a multimodal comprehensive physical health feature vector with high-dimensional reduction loss is output and input into the back-end evaluation network.

[0095] By using a multimodal large language model as the cognitive center, the system identifies and evaluates needs and triggers a function call mechanism, transforming feature vectors into physical boundary conditions.

[0096] The multimodal physical health feature vectors are input into a large model perceptual encoder and mapped to the evaluation space using semantic alignment. Prompt words guide the identification of disease intent and trigger function calls, outputting structured JSON instructions to schedule an external physics engine. A projection mapping layer then transforms the nonlinear components of the feature vectors into physical boundary parameters such as equivalent cross-sectional loss and stiffness degradation, which are directly input into the physical information neural network and differentiable solver.

[0097] Step 204: Embed a Physical Information Neural Network (PINN) and a differentiable structural solver into the physics engine. Based on boundary conditions, calculate physical degradation parameters and residual bearing capacity coefficients in real time, and output the structural defect level. In this embodiment, the physics engine adopts a cascaded architecture that deeply couples deep learning and classical mechanics, consisting of an upstream PINN and a downstream differentiable solver, forming an end-to-end data and gradient path.

[0098] Classical governing equations (such as the Euler-Bernoulli beam equation) from the "General Specifications for Design of Highway Bridges and Culverts" are extracted to construct physical penalty terms, which are combined with data-driven terms to form a joint loss function. Gradient penalty is used to ensure that the solution parameters are physically interpretable. Boundary conditions are input into a differentiable solver to calculate the remaining bearing capacity coefficient in real time, and the structural defect level is output by comparing it with the rigid threshold of defects from level one to four.

[0099] like Figure 4 As shown, the computational process of embedding PINN and a differentiable structural solver in this embodiment of the invention is as follows: receiving physical boundary conditions such as the equivalent cross-sectional loss rate from the large language model parsing, initializing the parameters of the differentiable structural solver; extracting the normative mechanical control equations, constructing a joint loss function of data-driven terms and physical penalty terms; calculating the resistance and action effect under the most unfavorable load combination through forward propagation, and outputting the equivalent stiffness reduction coefficient; applying a high gradient penalty to the decoding features that violate mechanical equilibrium, and ensuring physical interpretability through backpropagation; outputting the residual bearing capacity coefficient, and giving a hard classification result by comparing with the rigidity thresholds of grades one to four defects.

[0100] Step 205: Input the calculated physical degradation parameters and disease levels into the graph retrieval enhancement module, and perform multi-level association and precise feature tracing in the engineering specification text and historical maintenance knowledge graph;

[0101] Using industry standard texts, material structure relationships, historical reinforcement drawings, and expert maintenance rules as entity nodes and subordinate and logical associations as edges, a multi-dimensional knowledge graph of the entire life cycle of bridges and tunnels is constructed. The topological importance, stiffness reduction coefficient, and residual bearing capacity coefficient of components in the cognitive environment of the large model are extracted. The graph embedding network is used to generate joint query vectors for subgraph matching, recalling relevant regulations and treatment experience, and reorganizing them into a structured context.

[0102] Step 206: Combining the reasoning ability of the large language model with the underlying quantitative mechanical evidence and the high-level normative guidance, automatically generate a comprehensive assessment report and structural treatment strategies.

[0103] like Figure 5 As shown, the process for standard tracing and decision generation based on graph retrieval enhancement and thought chain in this embodiment of the invention specifically includes:

[0104] First, the system integrates industry standards, material structure relationships, historical reinforcement drawings, and expert maintenance rules to construct a multi-dimensional knowledge graph for bridges and tunnels. At the same time, it utilizes a large language model in conjunction with GPU-accelerated calculations to extract key parameters such as component topological importance, stiffness reduction factor, and remaining bearing capacity in real time.

[0105] Subsequently, the extracted physical feature parameters are transformed into a high-dimensional joint query vector through feature mapping via a graph embedding network. This vector serves as the core of the retrieval process, performing precise subgraph similarity matching within the multidimensional knowledge graph to locate knowledge clusters highly correlated with the current structural physical decay indicators.

[0106] During the context reorganization phase, the system accurately recalls and matches key legal regulations, normative requirements, and high-scoring historical case handling experience nodes, integrating them into a logically rigorous structured context. This process eliminates the fragmentation of the original data, providing robust legal and engineering experience for decision generation.

[0107] Finally, through the step-by-step deduction of disease cause analysis, standard requirements benchmarking, and comprehensive decision-making, the system ultimately outputs a comprehensive decision dossier containing authoritative references and targeted reinforcement processes, realizing a closed-loop generation from original monitoring parameters to automated and professional treatment solutions.

[0108] Based on the above technical solution, the following detailed embodiments are provided.

[0109] For the input multi-source heterogeneous detection data of bridge and tunnel components, the following calculation and deduction steps are performed:

[0110] Step 1: This embodiment first addresses the problem of spatial alignment difficulties and temporal discontinuities in multi-source defect data under complex physical environments. It acquires two-dimensional local defect images captured by a UAV and one-dimensional waveform sequences from ground-penetrating radar. The local point cloud coordinates of a single detection are... By using a global BIM mapping engine and employing an iterative nearest-point algorithm for spatial registration, the optimal spatial rotation matrix is ​​obtained. With translation vector Transform it to the global three-dimensional coordinate system In the middle, a defect location map of all components is achieved:

[0111] (1)

[0112] Subsequently, for historical detection records spanning multiple years (temporal discontinuities caused by non-uniform sampling), a selective state-space model was used to analyze the historical health feature sequences at physical structure nodes. Modeling continuous-time systems:

[0113] (2)

[0114] (3)

[0115] in, To hide the health state vector, This represents the hidden health state vector of the system at the next time step. Let A represent the output state of the system at time t, B be the input projection, C be the output matrix, and D be the feedforward matrix. To perform efficient high-level sequence tensor operations within modern deep learning frameworks, a zero-order hold (ZOH) is used to discretize the continuous system described above. The discretization time step is defined as... The system introduces a discretized state transition matrix. With discretized input projection matrix The strict mathematical conversion relationship between these two systems and the matrices of continuous systems is as follows:

[0116] (4)

[0117] in The state transition matrix is ​​discretized, where I is the identity matrix, and the result is discretized into time steps. Update steps:

[0118] (5)

[0119] To dynamically filter redundant noise from historical detections and retain critical degraded nodes, the system incorporates a dynamic state gating mechanism based on the Hadamard product during state transitions. Definition It is a Sigmoid non-linear activation function. This is an element-wise multiplication operation for matrices or vectors. The system performs this operation based on the current input. Dynamic calculation of gating coefficient , To represent the gate weight matrix, To represent the gate bias parameters, and thus achieve efficient updates of the discrete state:

[0120] (6)

[0121] This outputs a smooth characteristic that contains long-term degradation patterns. C is the mapping matrix from the hidden state to the output space, constructing a multimodal spatiotemporal database that is aligned in both time and space.

[0122] Step 2: After completing the global spatiotemporal mapping and long-term time-series penetration described above, the system obtains a highly aligned component-level multimodal spatiotemporal database. To ensure the continuity and fidelity of feature engineering, the subsequent native multimodal classifier will directly use this database as the sole data source input.

[0123] After obtaining unified data, this embodiment employs a native multimodal classifier to eliminate the high-dimensional feature loss caused by converting data to text before inference in traditional large language models for processing disease data. The visual appearance feature matrix extracted from the image is then used. (Characterizing surface cracks, spalling, etc.) and waveform frequency domain feature matrix extracted by radar (Characteristics of internal voids, de-voids, etc.) are mapped to the latent variable space of the same dimension to construct a cross-modal attention mechanism.

[0124] , , (7)

[0125] in, , , To query the matrix, key matrix, and value matrix, , , This is a learnable weight matrix optimized for backpropagation in the network. The system implements mathematical interaction fusion of visual features and internal detection radar waves at the bottom layer, calculates cross-modal attention weights and outputs a preliminary fusion matrix, then introduces residual connections and layer normalization to ensure gradient stability, finally outputting a comprehensive physical health feature vector. :

[0126] (8)

[0127] (9)

[0128] in, This represents the implicit space feature dimension. Through rigorous underlying mathematical constraints, the system can autonomously discover and preserve the complex implicit physical coupling relationship between minute surface cracks and large internal voids.

[0129] Step 3: Parameter Prediction and Physical Constraints of Upstream PINN. After the multimodal large language model triggers the function call, it transforms the features into physical boundary conditions and inputs them into the upstream PINN. PINN, as the front-end parameter prediction network, outputs the structural stiffness reduction coefficient as its forward propagation target. .

[0130] To ensure that the AI ​​decoding parameters strictly conform to the laws of real-world civil physics, a physical information neural network is introduced, and its joint loss function is defined as a weighted sum of data-driven terms, physical constraint terms, and boundary condition penalty terms:

[0131] (10)

[0132] Specifically, the classic Euler-Bernoulli Beam deflection differential equation from the "General Specifications for Highway Bridge and Culvert Design" is extracted to construct the core physical penalty term:

[0133] (11)

[0134] in, To design the initial bending stiffness, The deflection curve of the beam is shown below. For external load distribution, is the structural stiffness reduction factor output by the network decoding, and N is the number of sampling points used to calculate the physical constraint loss. This imposes hard boundary conditions such as zero deflection at the support (e.g.) If PINN predicts... This causes the calculated deflection to violate the mechanical equilibrium state, resulting in the network being subjected to a huge penalty gradient during training, forcing the output parameters to be guaranteed. This ensures the physical credibility of the output parameters, thereby forcibly guaranteeing their physical interpretability.

[0135] Step 4: Forward calculation of the downstream differentiable structure solver. PINN does not directly output the load-bearing capacity, but rather outputs... The solution is passed downstream to a differentiable structure solver. This solver is a finite element calculus operator written in a differentiable tensor framework. It is based on the input... Real-time calculation of the section resistance effect of the component under the most unfavorable load combination Actual effect Among them, the resistance effect With material design strength and the equivalent section modulus after reduction due to disease Directly related. This leads to the determination of the structure's residual bearing capacity coefficient. :

[0136] (12)

[0137] in, This represents the structural importance coefficient. Crucially, because the downstream solver is fully differentiable, the physics computation is no longer a black box blocking gradients. During model optimization, the residual gradients generated by the downstream computation can flow back through the solver and into the upstream PINN to update network weights, thus achieving a perfect closed loop between the purely data-driven model and the hard-core physics computation rules.

[0138] Based on the rigid threshold output corresponding to the structural defect level, the established rigid grading judgment rule introduces a structural stiffness reduction factor. Stiffness reduction minimum threshold and the lower limit threshold of the allowable bearing capacity. Based on the calculation results, the system strictly adheres to the rigid thresholds of engineering standards and implements a four-level automated hard classification:

[0139] Level 1 defect determination: When the calculation results satisfy the structural stiffness reduction factor. And the remaining bearing capacity coefficient When the structure is determined to have only surface non-structural damage and no decrease in bearing capacity coefficient, the output is a level one defect.

[0140] Secondary defect determination: When the calculation results satisfy the structural stiffness reduction factor. And the remaining bearing capacity coefficient When the structural durability index is damaged but the remaining load-bearing capacity has not substantially fallen below the working threshold, the output is a level two defect.

[0141] Level 3 Damage Determination: When the calculation result satisfies the condition that the remaining bearing capacity coefficient is within the specified range... When the weakening of the effective cross section of the structure leads to a clear reduction in the remaining bearing capacity, the output is a level three defect;

[0142] Level IV Defect Judgment: When the calculation result satisfies the aforementioned remaining bearing capacity coefficient... If the remaining load-bearing capacity of the structure is significantly reduced and there is a risk of component instability or fracture, the output will be classified as Level 4 defect.

[0143] To ensure the rigor of the evaluation results and eliminate ambiguities in the qualitative descriptions of the claims, this invention has performed standard engineering quantification on the control parameters in the above-mentioned grading rules. Among them, the stiffness reduction minimum threshold... The range of values ​​is limited to (In this embodiment, 0.05 is preferred), used as the critical point between the degradation of the microscopic surface material and substantial durability damage. The specified lower limit threshold of allowable bearing capacity... The safety reserve requirement for the ultimate limit state of structural bearing capacity is set according to the "Specifications for Maintenance of Highway Bridges and Culverts" (JTG5120) and the "Standards for Technical Condition Assessment of Highway Bridges" (JTG / TH21), and its value range is limited to [value missing]. (In this embodiment, 0.85 is preferred).

[0144] Step 5: Finally, this embodiment utilizes graph retrieval enhancement technology and the thought chain logic of a large language model to achieve highly interpretable engineering-level tracing and decision support.

[0145] To ensure that numerical physical indicators and structured topological information can be accurately retrieved within the same metric space as natural language text, the system constructs a heterogeneous feature joint encoding and spatial alignment module.

[0146] First, the system extracts the bearing capacity coefficients output by the differentiable solver. Stiffness reduction rate Since it is a low-dimensional continuous scalar, the system uses maximum-minimum normalization to map it to the [0,1] interval, and then inputs it into a multilayer perceptron for nonlinear dimensionality upscaling to generate physical feature vectors. , is the dimension of the physical node feature vector.

[0147] Simultaneously, the topological connectivity graph of this component is extracted from the global BIM, and a graph convolutional neural network is used to aggregate the force transmission features of first-order and second-order neighbor nodes to generate a topological importance vector. , is the dimension of the feature vector of the topological node.

[0148] The above features are concatenated into a composite vector. To bridge the semantic gap between the engineering physical space and the knowledge graph semantic space, the system introduces a learnable alignment projection matrix. Through the formula:

[0149] (13)

[0150] in, The bias parameters in the feature alignment process are mapped to the final high-dimensional joint query vector. (Where D is the dimension of the knowledge graph node vector). The alignment projection matrix is ​​optimized using a contrastive learning loss function during the training phase to narrow the distance between the physical vector with severe load-bearing capacity decay and the corresponding maintenance specification text vector in the feature space.

[0151] After the above mapping It has the ability to connect with the knowledge graph node pool The exact same mathematical foundation and metric are used. Subsequently, the system uses cosine similarity to calculate high-scoring matching nodes in the knowledge graph:

[0152] (14)

[0153] in, This is the embedding vector of the nth candidate node in the knowledge graph. It accurately recalls matching clauses from the "General Specifications for Design of Highway Bridges and Culverts" and high-scoring historical reinforcement experiences. The large language model uses this as a structured context, forcing reasoning to follow a hierarchical thought chain logic.

[0154] 1. Disease characteristics and quantitative identification (indicating crack width and radar anomaly depth).

[0155] 2. Determination of bottom-layer mechanical attenuation (list the calculated results) and (numerical value).

[0156] 3. Regulatory requirements and threshold comparison (referencing specific legal provisions related to the recall).

[0157] The process of automatically generating bridge and tunnel health assessment reports and decision-making schemes using the multimodal large language model is achieved by constructing structured prompt word instruction templates and driving an explicit chain of thought reasoning pipeline. The system does not rely on uncontrollable free text generation; instead, it standardizes and fills slots with multimodal features, physics engine-calculated values, and graph retrieval recall terms.

[0158] Ultimately, a digital twin medical examination decision file is automatically generated, which includes specific construction techniques such as carbon fiber bonding and external prestressing reinforcement, and has a complete mechanical evidence chain, thereby enhancing the engineering credibility of the results.

[0159] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0160] This invention also provides a bridge and tunnel health assessment device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0161] Figure 6 A bridge and tunnel health assessment system according to one embodiment of the present invention includes:

[0162] The acquisition module 301 is used to acquire multi-source heterogeneous detection data of bridge and tunnel components, and to perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on a global spatiotemporal base to construct a precisely aligned component-level multimodal spatiotemporal database.

[0163] Output module 302 is used to jointly characterize the multi-source heterogeneous detection data using a native multimodal classifier, extract a comprehensive physical health feature vector, and use a multimodal large language model to trigger function call mechanism to schedule the physical engine. It then calculates the physical degradation parameters and remaining bearing capacity coefficient of the component through a physical information neural network and a differentiable structure solver, and outputs the corresponding structural defect level.

[0164] The generation module 303 is used to generate bridge and tunnel health assessment results by using graph retrieval enhancement technology to trace features in a pre-set engineering design specification and historical maintenance knowledge graph based on the physical degradation parameters and the structural damage level.

[0165] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0166] According to one embodiment of the present invention, an electronic device is also provided, comprising: a memory storing an executable program; and a processor for running the program, wherein the program performs the aforementioned bridge and tunnel health assessment during runtime.

[0167] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0168] Step S1: Obtain multi-source heterogeneous detection data of bridge and tunnel components, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on the global spatiotemporal base, and construct a precisely aligned component-level multimodal spatiotemporal database.

[0169] Step S2: The native multimodal classifier is used to jointly represent the multi-source heterogeneous detection data, extract the comprehensive physical health feature vector, and use the multimodal large language model to trigger the function call mechanism to schedule the physical engine. The physical information neural network and differentiable structural solver are used to calculate the physical degradation parameters and remaining bearing capacity coefficient of the component, and output the corresponding structural disease level.

[0170] Step S3: Based on physical degradation parameters and structural damage levels, feature tracing is performed in a pre-set engineering design specification and historical maintenance knowledge graph using graph retrieval enhancement technology to generate bridge and tunnel health assessment results.

[0171] According to one embodiment of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, it controls the device where the storage medium is located to perform the above-mentioned bridge and tunnel health assessment.

[0172] Optionally, in this embodiment, the storage medium may be configured to store a computer program for performing the following steps:

[0173] Step S1: Obtain multi-source heterogeneous detection data of bridge and tunnel components, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on the global spatiotemporal base, and construct a precisely aligned component-level multimodal spatiotemporal database.

[0174] Step S2: The native multimodal classifier is used to jointly represent the multi-source heterogeneous detection data, extract the comprehensive physical health feature vector, and use the multimodal large language model to trigger the function call mechanism to schedule the physical engine. The physical information neural network and differentiable structural solver are used to calculate the physical degradation parameters and remaining bearing capacity coefficient of the component, and output the corresponding structural disease level.

[0175] Step S3: Based on physical degradation parameters and structural damage levels, feature tracing is performed in a pre-set engineering design specification and historical maintenance knowledge graph using graph retrieval enhancement technology to generate bridge and tunnel health assessment results.

[0176] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0177] According to one embodiment of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, performs the above-described bridge and tunnel health assessment.

[0178] Optionally, in this embodiment, the above-mentioned computer program product can be configured as a computer program that performs the following steps:

[0179] Step S1: Obtain multi-source heterogeneous detection data of bridge and tunnel components, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on the global spatiotemporal base, and construct a precisely aligned component-level multimodal spatiotemporal database.

[0180] Step S2: The native multimodal classifier is used to jointly represent the multi-source heterogeneous detection data, extract the comprehensive physical health feature vector, and use the multimodal large language model to trigger the function call mechanism to schedule the physical engine. The physical information neural network and differentiable structural solver are used to calculate the physical degradation parameters and remaining bearing capacity coefficient of the component, and output the corresponding structural disease level.

[0181] Step S3: Based on physical degradation parameters and structural damage levels, feature tracing is performed in a pre-set engineering design specification and historical maintenance knowledge graph using graph retrieval enhancement technology to generate bridge and tunnel health assessment results.

[0182] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0183] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0184] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0185] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0186] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0187] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0188] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the health of bridges and tunnels, characterized in that, include: Acquire multi-source heterogeneous detection data of bridge and tunnel components, perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on a global spatiotemporal base, and construct a precisely aligned component-level multimodal spatiotemporal database; A native multimodal classifier is used to jointly characterize the multi-source heterogeneous detection data, extract the comprehensive physical health feature vector, and use the multimodal large language model to trigger the function call mechanism to schedule the physical engine. The physical degradation parameters and residual bearing capacity coefficient of the component are calculated through the physical information neural network and the differentiable structure solver, and the corresponding structural defect level is output. Based on the physical degradation parameters and the structural damage level, feature tracing is performed in a pre-set engineering design specification and historical maintenance knowledge graph using graph retrieval enhancement technology to generate bridge and tunnel health assessment results.

2. The bridge and tunnel health assessment method according to claim 1, characterized in that, The multi-source heterogeneous detection data includes: one-dimensional ground-penetrating radar waveforms, two-dimensional disease images, and semi-structured historical detection text.

3. The bridge and tunnel health assessment method according to claim 2, characterized in that, Based on the global spatiotemporal base, spatial mapping and temporal penetration are performed on the multi-source heterogeneous detection data, including: Using the spatial anchoring components inside the global spatiotemporal base, a global building information model or geographic information system mapping engine is established to map the linear coordinates of ground penetrating radar, the geographic labels of UAV images, and the manually detected station offsets to physical structure nodes in the same three-dimensional spatial coordinate system, thereby achieving absolute spatial anchoring. Furthermore, by utilizing the time modeling component within the global spatiotemporal base, a deep learning architecture based on a state-space model is introduced to perform long-sequence modeling of historical detection records spanning multiple years on the physical structure nodes, extract temporal evolution patterns, and transform discrete detection data into a continuous health state evolution sequence.

4. The bridge and tunnel health assessment method according to claim 3, characterized in that, The native multimodal classifier is used to jointly characterize the multi-source heterogeneous detection data, including: The acquired one-dimensional ground-penetrating radar waveform is decomposed into multi-scale time and frequency using discrete wavelet transform to extract high-frequency detail coefficients that characterize abrupt changes and anomalies. The high-frequency detail coefficients are then reconstructed and spliced ​​using feature dimensions to form a frequency domain sequence of the one-dimensional ground-penetrating radar waveform. The frequency domain sequence of the one-dimensional ground-penetrating radar waveform and the pixel matrix of the two-dimensional disease image are directly mapped to a unified high-dimensional shared latent variable space through their respective feature extraction networks without going through natural language parsing and text encoding, thus generating multimodal original feature tensor identifiers. In a high-dimensional latent variable space, a fusion module based on a cross-modal cross-attention mechanism is constructed to perform cross-modal cross-attention calculation, which enables the anomalous reflection signal of radar waves to interact with the crack features on the image surface at a low level, thereby uncovering the physical coupling relationship between internal defects and surface cracks.

5. The bridge and tunnel health assessment method according to claim 4, characterized in that, The physics engine is scheduled by triggering the function call mechanism using the multimodal large language model, including: The differentiable structure solver is encapsulated as an external executable tool interface; After analyzing the comprehensive physical health feature vector, the multimodal large language model determines that mechanical quantitative verification is required, and then transforms the comprehensive physical health feature vector into an equivalent cross-sectional loss rate or a local stiffness degradation coefficient with physical meaning. The external executable tool interface is invoked to input the equivalent cross-sectional loss rate or the local stiffness degradation coefficient as physical boundary conditions and material properties into the underlying physical mechanism calculation network.

6. The bridge and tunnel health assessment method according to claim 5, characterized in that, The physical degradation parameters and the remaining bearing capacity coefficient of the component are calculated using the physical information neural network and the differentiable structural solver, including: Construct a cascaded physics engine that includes an upstream physical information neural network and a downstream differentiable structure solver; In the upstream prediction stage, the physical information neural network receives the physical boundary conditions transmitted by the multimodal large language model, predicts the structural stiffness reduction coefficient of the output component, and adds the preset force balance equation as a physical penalty term to the loss function of the physical information neural network to apply gradient penalty to the prediction features that violate the common sense of mechanics. In the downstream calculation stage, the structural stiffness reduction factor is used as a known condition and input into the downstream end-to-end differentiable structural solver for forward finite element calculation. The resistance effect and action effect of the section under the current load combination are calculated, and a quantitative residual bearing capacity coefficient is output. Furthermore, the differentiable structural solver supports the backpropagation of the calculated gradient to support the joint training of the upstream physical information neural network.

7. The bridge and tunnel health assessment method according to claim 6, characterized in that, The output of the structural defect level includes: By introducing structural stiffness reduction factor, stiffness reduction minimum threshold and code allowable bearing capacity lower limit threshold, a rigid classification judgment rule is established. When the calculation results satisfy the condition that the structural stiffness reduction coefficient is less than or equal to the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold of the allowable bearing capacity in the code, it is determined that the structure only has surface non-structural damage and the bearing capacity coefficient has no decay, and the output is a first-level defect. When the calculation result satisfies the condition that the structural stiffness reduction coefficient is greater than the minimum stiffness reduction threshold and the remaining bearing capacity coefficient is greater than or equal to the lower limit threshold of the allowable bearing capacity specified in the code, it is determined that the structural durability index is damaged but the remaining bearing capacity has not fallen below the working threshold, and the output is a level two defect. When the calculation result satisfies the condition that the remaining bearing capacity coefficient is within the range between the lower limit threshold of the standard allowable bearing capacity and the preset ratio threshold of the lower limit threshold of the standard allowable bearing capacity, it is determined that the weakening of the effective section of the structure has led to a reduction in the remaining bearing capacity, and the output is a level three defect. When the calculation result satisfies the preset proportional threshold that the remaining bearing capacity coefficient is less than the lower limit threshold of the allowable bearing capacity in the specification, it is determined that the remaining bearing capacity of the structure has decreased and there is a risk of component instability or fracture, and the output is a level four defect.

8. The bridge and tunnel health assessment method according to claim 7, characterized in that, The graph retrieval enhancement technology is used to perform feature tracing within the pre-set engineering design specifications and the historical maintenance knowledge graph, including: Extract the structural bearing capacity attenuation index output by the differentiable structure solver, as well as the topological information on the stress importance of the component in the global spatiotemporal base; After normalizing the numerical physical parameters, they are input into a multilayer perceptron and encoded into physical feature vectors. Then, a graph convolutional network is used to aggregate and encode the topological importance information of the graph structure to generate topological feature vectors. The physical feature vector and the topological feature vector are concatenated, and then nonlinearly mapped to the same metric space as the pre-trained knowledge graph nodes through an alignment projection network to generate a high-dimensional joint query vector. The high-dimensional joint query vector is matched in a subgraph of a knowledge graph containing industry standard text, historical reinforcement drawings, and expert decision-making logic. Regulatory restrictions and treatment experience nodes for similar diseases that match the current attenuation index are retrieved and sent back to the multimodal large language model as contextual prompts.

9. The bridge and tunnel health assessment method according to claim 8, characterized in that, The bridge and tunnel health assessment results are generated, including: Construct a structured prompt word instruction template that includes role definition, feature slots, tiered execution instructions, and output format constraints; The multimodal original feature tensor identifier, the quantitative physical degradation parameter and residual bearing capacity coefficient calculated by the differentiable structure solver, and the specification clauses recalled by the graph retrieval enhancement module are dynamically filled into the corresponding slots of the structured prompt word instruction template to assemble and generate high-dimensional context prompt words. The high-dimensional contextual prompts are input into the multimodal large language model, enabling the multimodal large language model to perform a step-by-step reasoning chain consisting of disease quantification identification, mechanical mechanism deduction, and standard requirement benchmarking, and to explicitly generate structured labels corresponding to each reasoning step in the final output; The output includes a digital decision dossier for structural health monitoring containing quantitative evidence of mechanical attenuation, citations of clause codes, and specific construction and reinforcement techniques.

10. A bridge and tunnel health assessment system, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous detection data of bridge and tunnel components, and to perform spatial mapping and temporal penetration on the multi-source heterogeneous detection data based on a global spatiotemporal base to construct a precisely aligned component-level multimodal spatiotemporal database. The output module is used to jointly characterize the multi-source heterogeneous detection data using a native multimodal classifier, extract a comprehensive physical health feature vector, and use a multimodal large language model to trigger function call mechanism to schedule the physical engine. It then calculates the physical degradation parameters and remaining bearing capacity coefficient of the component through a physical information neural network and a differentiable structure solver, and outputs the corresponding structural defect level. The generation module is used to generate bridge and tunnel health assessment results by using graph retrieval enhancement technology to trace features in a pre-set engineering design specification and historical maintenance knowledge graph based on the physical degradation parameters and the structural damage level.

Images