Bridge fire risk state assessment method and system based on large model

Abstract

This invention relates to the interdisciplinary field of next-generation information technology and intelligent transportation infrastructure safety monitoring, and discloses a bridge fire risk status assessment method and system based on a large model. The key technical points are: system initialization and physical device-twin base mapping construction; collaborative perception of multimodal physical quantities and spatial visual features; transient thermal resistance network deduction and global synchronization deviation quantification; dimension-aligned graph network feature fusion; large model state reasoning with embedded physical knowledge constraints; safety margin quantification and three-dimensional coordinate-level hardware linkage. It employs digital twin thermal resistance network deduction, graph topology feature fusion, and large model physical knowledge constraint collaboration technology. By transforming the underlying objective thermodynamic mechanism into a dynamic penalty term and injecting it into the large model reasoning layer for decision correction, it eliminates the problem of insufficient physical consistency in purely data-driven models and constructs a hardware-level control closed loop for the entire system from multimodal perception to three-dimensional spatial targeted electromechanical linkage.

Description

Technical Field

[0001] This invention relates to the field of cross-disciplinary technology of next-generation information technology and intelligent transportation infrastructure safety monitoring, and more specifically, to a method and system for assessing bridge fire risk status based on a large model. Background Technology

[0002] In the field of bridge fire risk assessment, existing monitoring systems typically rely on isolated sensor alarms or purely data-driven artificial intelligence models, facing several critical technical challenges: First, large, purely data-driven models lack prior constraints from thermodynamic and mechanical physical mechanisms when dealing with extreme long-tail disaster scenarios, making them prone to decision-making biases that violate objective physical laws, resulting in a lack of engineering reliability and interpretability in risk assessment. Second, limited by the physical deployment blind spots of physical sensor networks, existing methods cannot fully quantify the penetration effect of fire heat into the core layer of the main cable and other structures. Furthermore, multi-source heterogeneous sensing data from visual, temperature, and electrical sources lack a deep feature alignment and spatial fusion mechanism based on the bridge's three-dimensional physical topology, making it difficult to accurately identify hidden damage. In addition, existing systems are mostly limited to one-way data display and passive early warning at the software level, failing to establish a complete hardware-level control loop from underlying multimodal feature perception and twin-space thermal field extrapolation to three-dimensional spatial targeted electromechanical linkage (such as spatial fixed-point precision spraying and physical hard isolation).

[0003] Therefore, the present invention provides a method and system for assessing the fire risk status of bridges based on a large model, thereby improving the above-mentioned technical problems. Summary of the Invention

[0004] This disclosure aims to address the shortcomings of existing technologies by providing a bridge fire risk assessment method and system based on a large model. The invention employs digital twin thermal resistance network deduction, graph topology feature fusion, and large model physical knowledge constraint coordination technology. By transforming the underlying objective thermodynamic mechanism into a dynamic penalty term and injecting it into the large model inference layer for decision correction, it eliminates the problem of insufficient physical consistency in pure data-driven models, accurately quantifies hidden structural damage, and constructs a hardware-level control closed loop for the entire system, from multimodal perception to three-dimensional spatial targeted electromechanical linkage.

[0005] To achieve the above objectives, the present disclosure proposes the following technical solutions:

[0006] In a first aspect, this disclosure proposes a bridge fire risk status assessment method based on a large model, applied to a bridge fire risk status assessment system. The system includes: underlying sensing and electromechanical execution equipment, front-end edge computing nodes, a local digital twin server, a cloud-based main control server, and a programmable logic controller. The underlying sensing and electromechanical execution equipment includes at least an explosion-proof high-definition camera, a distributed fiber Bragg grating temperature sensor, a fiber Bragg grating demodulator, a current transformer, an anemometer, a high-pressure fire monitor, and an LED traffic guidance screen. The method includes:

[0007] The front-end edge computing node and the local digital twin server read the bridge's 3D BIM file and discretize it into a graph topology, and bind the communication addresses of the underlying sensing and electromechanical execution equipment to the corresponding 3D spatial physical nodes in the graph topology.

[0008] The front-end edge computing node schedules the underlying sensing and electromechanical execution devices to collect visual and physical signals, analyze the three-dimensional physical coordinates and transient heat release power of the fire source, and convert the physical signals into measured physical temperature, radial physical temperature gradient and instantaneous current gradient.

[0009] The local digital twin server receives the transient heat release power and the three-dimensional physical coordinates, solves the differential equation of the equivalent thermal resistance network, deduces the transient theoretical temperature of the three-dimensional physical node in the twin space, quantifies the global synchronization error scalar between real perception and theoretical deduction, and extracts the current theoretical highest temperature of the entire network.

[0010] The cloud-based master control server performs standardization processing on the received physical parameters to construct local physical feature column vectors, performs spatial physical information aggregation through graph convolutional neural networks to generate graph embedding vectors, and concatenates the global synchronization error scalar to generate a unified high-order digital feature vector.

[0011] The vision-language big model of the cloud master control server receives the unified high-order digital feature vector, and simultaneously activates the mechanism rule constraint module to convert the theoretical highest temperature of the entire network into a physical knowledge constraint penalty function and inject it into the reasoning and decision-making process of the big model for constraint correction, and outputs a comprehensive health status score.

[0012] The local digital twin server quantifies the real-time load-bearing safety margin coefficient of the fire-affected component. When the real-time load-bearing safety margin coefficient and the comprehensive health status score trigger a preset high-risk threshold, the programmable logic controller analyzes the three-dimensional physical coordinates of the fire source, drives the high-pressure fire monitor to perform three-dimensional spatial targeted and precise spraying fire extinguishing, and sends a traffic closure command to the LED traffic guidance screen.

[0013] As a preferred embodiment of the present invention, the front-end edge computing node resolves the three-dimensional physical coordinates and transient heat release power of the fire source, including:

[0014] The explosion-proof high-definition camera acquires real-time video frame sequences. After the front-end edge computing node performs adaptive histogram equalization, it calls the target detection network to extract the two-dimensional bounding box of the suspected fire source. The spatial intersection-over-union loss function is used to optimize the two-dimensional bounding box, and its distance cost... The calculation formula is:

[0015] ;

[0016] in, The index variable for the two-dimensional coordinate axes of the image, with values... shaft or axis; The center points of the predicted bounding box and the center points of the ground truth bounding box are on the coordinate axis. Normalized Euclidean distance in the direction; This is the penalty coefficient for angular deflection; The distance cost attenuation weighting index;

[0017] Based on the pre-calibrated camera intrinsic and extrinsic parameter matrix, the front-end edge computing node projects the optimized two-dimensional bounding box into the three-dimensional physical coordinates of the fire source. Simultaneously, it matches the preset fire growth coefficient. The transient heat release power is output based on the fire growth model. :

[0018] ;

[0019] in, The start timestamp for the system to identify and confirm the fire source; This represents the actual duration of the fire.

[0020] As a preferred embodiment of the present invention, the front-end edge computing node converts physical signals into measured physical temperature, radial physical temperature gradient, and instantaneous current gradient, including:

[0021] The fiber optic demodulator reads data from physical nodes embedded in three-dimensional space. The optical wavelength signal from the sensor is converted into a measured physical temperature by the front-end edge computing node based on the Bragg grating sensing equation. :

[0022] ;

[0023] in, The initial reference temperature for sensor calibration; The center wavelength is read in real time; To calibrate the initial wavelength; This refers to the inherent temperature sensitivity coefficient of the grating material;

[0024] The front-end edge computing node calculates the radial physical temperature gradient of the node's cross-section based on the layered physical layout coordinates of the sensors within the bridge component's cross-section. :

[0025] ;

[0026] in, and They are nodes Measured temperatures of the outer and core layer sensors; The physical distance between the two sensor layers;

[0027] The current transformer acquires a real-time current signal sequence through an analog-to-digital converter module, and obtains a filtered current sequence after noise reduction via a sliding window low-pass filter. The front-end edge computing node outputs the instantaneous current gradient using the discrete-time backward difference method. :

[0028] ;

[0029] in, This is the current filtered current value; This is the filtered current value from the previous sampling period; The sampling time interval for analog-to-digital conversion in the system hardware.

[0030] As a preferred embodiment of the present invention, the local digital twin server solves the equivalent thermal resistance network differential equation to derive the transient theoretical temperature and quantifies the global synchronization error scalar, including:

[0031] Combining the flow field data collected by the anemometer, the local digital twin server constructs the node dynamic thermodynamic equilibrium mechanism, solves the differential equation of the equivalent thermal resistance network with convective heat transfer correction, and derives the node dynamic thermodynamic equilibrium mechanism. Transient theoretical temperature in twin space :

[0032] ;

[0033] in, For nodes The equivalent heat capacity of the component; This represents the theoretical rate of change of temperature over time. Based on transient heat release power and coordinates Calculated fire source pairs of nodes The applied radiative heat flux density function; The surface convective heat transfer coefficient is calculated based on wind speed data; For convective heat transfer area; The temperature of the smoke in the fire environment; For nodes There exists a set of adjacent nodes for solid heat transfer; For adjacent node indexes; For nodes The theoretical temperature; The physical contact thermal resistance between nodes;

[0034] Based on the measured physical temperature Quantize the global synchronization error scalar :

[0035] ;

[0036] in, Total number of monitoring nodes; Let L2 norm be the square of a vector.

[0037] As a preferred technical solution of the present invention, the theoretical maximum temperature of the entire network is transformed into a physical knowledge constraint penalty function and injected into the inference and decision-making process of the large model for constraint correction, including:

[0038] The mechanism rule constraint module reads the theoretical maximum temperature of the entire network. Injecting physical knowledge into large models to constrain and penalize functions :

[0039] ;

[0040] in, This is a function to find the maximum value. It is a step activation function; This is a preset physical failure temperature threshold for steel. Decode the multi-class fire risk probability distribution vector output by the large model; This is a confidence parameter for classifying a risk level as high-risk.

[0041] The cloud-based main control server incorporates the physical knowledge constraint penalty function into the overall objective optimization function of the large model for constraint correction.

[0042] ;

[0043] in, This is the total loss function; For weakly supervised classification loss items; These are pseudo-labels based on experience. The weight hyperparameters are used to control the physical strength of the constraints; The regularization coefficient is used. This is the network structure regularization loss term.

[0044] As a preferred embodiment of the present invention, the method further includes a degradation assessment method under extreme dense smoke conditions:

[0045] When the explosion-proof high-definition camera's target detection confidence level falls below the judgment threshold for a continuous preset time due to dense smoke obscuring the view, the front-end edge computing node automatically blocks the input of the visual perception channel, and the system enters a purely physical mechanism-driven mode.

[0046] In the pure physical mechanism driven mode, the calculation of the three-dimensional physical coordinates and the transient heat release power is skipped, and the measured physical temperature and the instantaneous current gradient are directly extracted. The radiation heat flux density function in the differential equation of the equivalent thermal resistance network is forced to zero, and the thermal field evolution is carried out solely by the physical conduction heat between adjacent nodes to deduce the transient theoretical temperature. At the same time, the trigger weight of the physical knowledge constraint penalty function is adjusted, thereby outputting the comprehensive health status score under the degraded working condition and triggering the linkage warning.

[0047] Secondly, this disclosure proposes a bridge fire risk status assessment system based on a large model. The system includes: a bottom-level sensing and electromechanical execution device, a front-end edge computing node, a local digital twin server, a cloud main control server, and a programmable logic controller.

[0048] The underlying sensing and electromechanical execution equipment includes at least an explosion-proof high-definition camera, a distributed fiber optic temperature sensor, a fiber optic demodulator, a current transformer, an anemometer, a high-pressure fire monitor, and an LED traffic guidance screen, used to collect multimodal visual and physical sensing data and execute electromechanical linkage commands.

[0049] The front-end edge computing node is communicatively connected to the underlying sensing and electromechanical execution equipment, and is used to schedule the data collected by the underlying equipment, perform two-dimensional boundary box optimization of the fire source, three-dimensional physical coordinate calculation, and convert physical signals into measured physical parameters.

[0050] The local digital twin server is communicatively connected to the front-end edge computing node and is used to construct the bridge graph topology, solve the equivalent thermal resistance network differential equation to deduce the transient theoretical temperature, quantify the global synchronization error scalar, and quantify the real-time load-bearing safety margin coefficient of the fire-affected component.

[0051] The cloud-based main control server is communicatively connected to the local digital twin server. It has a built-in multimodal feature fusion computing unit and a visual-language large model, which is used to perform graph neural network spatial information aggregation to generate a unified high-order digital feature vector. It also embeds a physical knowledge constraint penalty function to perform constraint correction for large model reasoning and decision-making, and outputs a comprehensive health status score.

[0052] The programmable logic controller is communicatively connected to the local digital twin server, the high-pressure fire monitor, and the LED traffic guidance screen. When the real-time load safety margin coefficient and the comprehensive health status score trigger a preset high-risk threshold, the controller analyzes the three-dimensional physical coordinates of the fire source, drives the high-pressure fire monitor to perform three-dimensional spatial targeted and precise spraying fire extinguishing, and issues a traffic closure command to the LED traffic guidance screen.

[0053] Thirdly, embodiments of this disclosure propose an electronic device including at least one processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement a bridge fire risk status assessment method based on a large model.

[0054] Fourthly, a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement a bridge fire risk status assessment method based on a large model.

[0055] Fifthly, a computer program product includes computer program instructions that, when executed by a computing device, cause the computing device to implement a bridge fire risk status assessment method based on a large model.

[0056] In summary, the present invention has the following beneficial effects:

[0057] Firstly, this invention addresses the issues of decision-making bias and insufficient physical consistency in large-scale models under long-tail disaster scenarios. By embedding a physical knowledge constraint penalty function into the inference decision-making of the visual-language large model (VLM), the theoretical maximum temperature of the entire network derived from the digital twin space is coupled with the physical rigidity failure threshold. This design breaks the limitations of purely data-driven models. When the underlying thermodynamic deduction touches the structural failure red line, the system forcibly corrects the low-confidence output of the large model through loss gradient, ensuring that the model decision absolutely follows objective thermodynamic and mechanical laws.

[0058] Secondly, this invention breaks through the blind spots of physical sensor deployment, realizing accurate quantification and panoramic identification of deep hidden damage to bridges; by constructing an equivalent thermal resistance network differential equation that considers convective heat transfer correction, the transient theoretical temperature of the entire network is deduced, and the physical absolute temperature is converted into a radial physical temperature gradient; combined with the global synchronous error quantification mechanism between real perception and theoretical deduction, this system can accurately perceive the penetration effect of fire heat into the core layer of the main cable or the interior of the closed box girder, solving the technical pain point that traditional single-point monitoring cannot reflect the internal degradation of the structure.

[0059] Third, this invention achieves physical spatial depth alignment of multi-source heterogeneous sensing data, constructing a globally unified topological health representation. Addressing the problem of fragmented data from different dimensions such as visual images, temperature gradients, and electrical signals, the bridge entity is discretized into a topological graph structure, and spatial information aggregation is performed along the three-dimensional physical topological edge using a graph convolutional neural network (GCN). This strictly physically aligned feature fusion method fully utilizes the prior information of the bridge's spatial structure, improving the collaborative effectiveness of heterogeneous data in disaster assessment.

[0060] Fourth, this invention achieves a leap from one-way passive early warning to active physical intervention, and establishes a high-reliability control closed loop at the hardware level for the entire system. It not only outputs a comprehensive health score at the software level, but also quantifies the real-time load-bearing safety margin coefficient by combining a material mechanics degradation model. When a high-risk threshold is triggered, the system directly schedules the PLC to analyze the three-dimensional physical coordinates of the fire source and calls the inverse kinematics algorithm to drive the high-pressure fire monitor to perform three-dimensional targeted and precise spraying. At the same time, it issues roadside traffic guidance and barrier lowering commands, which significantly improves the accuracy and response efficiency of emergency intervention under extreme disasters and minimizes the secondary physical damage caused by fire. Attached Figure Description

[0061] Figure 1 A flowchart of a bridge fire risk status assessment method based on a large model provided in an embodiment of the present invention;

[0062] Figure 2 This is a framework diagram of a bridge fire risk status assessment system based on a large model, provided for an embodiment of the present invention. Detailed Implementation

[0063] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0065] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0066] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0067] This disclosure aims to address the problems in bridge fire safety operation and maintenance where purely data-driven large models are prone to decision-making biases that violate objective physical laws in long-tail industrial scenarios, and the difficulty in quantifying hidden structural damage due to blind spots in the underlying physical sensor network. In view of this, this disclosure proposes a bridge fire risk status assessment method and system based on a large model to evaluate the structural safety and health status of bridge physical components and the overall topology under extreme fire conditions. This method employs digital twin thermal resistance extrapolation, graph network feature fusion, and large model mechanism constraint coordination techniques. By synchronously mapping physical thermodynamic equations in virtual space and injecting the underlying rigid failure threshold into a dynamic knowledge constraint penalty term into the inference layer, it addresses the uninterpretability of risk assessment and the complexity of multi-source heterogeneous data fusion in extremely complex disaster scenarios. This ensures the physical consistency of the large model's decision output and constructs a fully hardware-level, highly reliable closed loop from multimodal perception to three-dimensional spatial targeted electromechanical linkage.

[0068] Please refer to Figure 1 , Figure 1 This diagram illustrates the framework of the bridge fire risk assessment system based on a large model, as described in this embodiment. The system architecture, from bottom to top, includes:

[0069] The underlying sensing and electromechanical execution equipment (explosion-proof high-definition cameras, distributed fiber Bragg grating temperature sensors and their supporting fiber Bragg grating demodulators, current transformers, anemometers, high-pressure fire monitors, LED traffic guidance screens), front-end edge computing nodes, local digital twin servers, cloud main control servers (with built-in GPU multimodal feature fusion computing units and vision-language large models), and programmable logic controllers (PLCs).

[0070] The front-end edge computing node communicates with the underlying sensing and electromechanical execution equipment to schedule the data collection of the underlying equipment, perform two-dimensional boundary box optimization of the fire source, three-dimensional physical coordinate calculation, and convert physical signals into measured physical parameters;

[0071] The local digital twin server communicates with the front-end edge computing nodes to construct the bridge graph topology, solve the equivalent thermal resistance network differential equation to deduce the transient theoretical temperature, quantify the global synchronization error scalar, and quantify the real-time load-bearing safety margin coefficient of the fire-affected components.

[0072] The cloud-based main control server communicates with the local digital twin server. It has a built-in multimodal feature fusion computing unit and a vision-language large model, which is used to perform graph neural network spatial information aggregation to generate a unified high-order digital feature vector. It also embeds physical knowledge constraint penalty function to perform constraint correction for large model reasoning and decision-making, and outputs a comprehensive health status score.

[0073] The programmable logic controller (PLC) is connected to the local digital twin server, the high-pressure fire monitor, and the LED traffic guidance screen. When the real-time load safety margin coefficient and comprehensive health status score trigger the preset high-risk threshold, it analyzes the three-dimensional physical coordinates of the fire source, drives the high-pressure fire monitor to perform three-dimensional spatial targeted and precise spraying fire extinguishing, and sends a traffic closure command to the LED traffic guidance screen.

[0074] Please refer to Figure 2 , Figure 2 A flowchart of the bridge fire risk status assessment method based on a large model as described in this disclosure is shown; the overall process mainly includes the following 6 steps:

[0075] Step S1: System initialization and physical device-twin docking construction.

[0076] When the system starts, the front-end edge computing nodes and the local digital twin server perform initialization configuration: reading the bridge's 3D BIM (Building Information Model) file and discretizing it into a structure containing... Graph topology with 1 node ;in, This refers to the set of nodes representing the load-bearing and heat-transfer physical components of a bridge. It is a set of topological edges representing the spatial contact and physical connection relationships between components.

[0077] Perform hardware communication address mapping: Uniquely bind the demodulation channel number corresponding to the distributed fiber Bragg grating temperature sensor deployed on-site, the MAC address of the current transformer, and the IP address of the explosion-proof high-definition camera to the graph topology. The corresponding three-dimensional physical nodes .

[0078] Pre-configure the system's memory database with engineering benchmark constants and algorithm hyperparameter libraries, including: algorithm calibration parameters, a failure critical temperature of 300℃ characterizing the loss of physical load-bearing capacity of high-strength steel, an instantaneous current gradient alarm threshold of 300A / ms characterizing the induction of electrical fires, and a typical vehicle fire growth coefficient matrix set according to fire protection standards. and the initial design bearing capacity parameters of each load-bearing component. .

[0079] Step S2: Cooperative perception of multimodal physical quantities and spatial visual features

[0080] The front-end edge computing nodes schedule the underlying hardware to perform visual and physical signal acquisition, photoelectric conversion, and feature quantization in parallel.

[0081] In the visual perception channel, the bridge deck's explosion-proof high-definition camera continuously collects real-time video frame sequences. The front-end edge computing node performs adaptive histogram equalization through its internal processing chip to suppress ambient light interference. A pre-set target detection network is invoked to extract two-dimensional bounding boxes of suspected fire sources. To reduce positioning errors, the spatial intersection-union ratio (SIoU) loss function is used to optimize the bounding boxes, with its distance cost... The calculation formula is:

[0082] ;

[0083] in, The index variable for the two-dimensional coordinate axes of the image, with values... shaft or axis; The center points of the predicted bounding box and the center points of the ground truth bounding box are on the coordinate axis. Normalized Euclidean distance in the direction; This is the penalty coefficient used to correct the angle deflection of the target bounding box; This is the distance cost attenuation weighting index.

[0084] Based on the pre-calibrated camera intrinsic and extrinsic parameter matrices, the front-end edge computing nodes project the optimized 2D bounding box into the 3D physical coordinates of the fire source. Meanwhile, in accordance with fire protection engineering Fire growth model, matching the fire growth coefficient in loading step S0 Transient heat release power of the output fire source :

[0085] ;

[0086] in, For matching fire growth coefficients; The start timestamp for the system to identify and confirm the fire source; The actual duration of the fire is a variable representing the time it lasts.

[0087] In the physical thermal characteristic sensing channel, the fiber optic grating demodulator in the equipment room reads data embedded in the nodes. The sensor detects the optical wavelength signal and sends it to the front-end edge computing node; the front-end edge computing node converts it into a measured physical temperature based on the physical mechanism of the center wavelength drift of the Bragg grating. :

[0088] ;

[0089] in, The initial reference temperature for sensor calibration; The center wavelength is read in real time; To calibrate the initial wavelength; This is the inherent temperature sensitivity coefficient of the grating material.

[0090] After obtaining the physical absolute temperature at each measuring point, in order to quantify the penetration effect of fire heat into the structure, the front-end edge computing nodes further calculate the radial physical temperature gradient of the node cross-section based on the layered physical layout coordinates of the sensors within the cross-section of bridge components (such as the main cable). :

[0091] ;

[0092] in, and They are nodes Measured temperatures of the outer and core layer sensors; This refers to the physical distance between the two sensor layers.

[0093] Synchronously, in the electrical sensing channel, the current transformer directly and continuously acquires the real-time current signal sequence of key electrical equipment (such as heating cables and lighting cables) attached to the node through an analog-to-digital converter (ADC) module. To eliminate grid glitches, the front-end edge computing node performs sliding window low-pass filtering on the real-time current signal sequence to obtain a filtered current sequence, and uses the discrete-time backward differential method to output the instantaneous current gradient for early warning of electrical fires. :

[0094] ;

[0095] in, This is the current filtered current value; This is the filtered current value from the previous sampling period; The sampling time interval for analog-to-digital conversion in the system hardware.

[0096] Step S3: Transient thermal resistance network derivation and global synchronization deviation quantification

[0097] The local digital twin server receives the output of step S2. and The system constructs a dynamic thermodynamic equilibrium mechanism of nodes in virtual space based on the law of conservation of energy. Specifically, the dynamic temperature rise and fall rate of the node depends on the equivalent heat capacity of the component itself and is determined by the net input of three heat transfer paths: external space fire source radiation heat, surface convection heat transfer and solid internal physical conduction heat transfer.

[0098] Based on the above mechanism, the server combines the on-site environmental flow field data collected by the underlying anemometer to solve the differential equation of the equivalent thermal resistance network, including convective heat transfer correction, and accurately deduce the nodes. Transient theoretical temperature in twin space :

[0099] ;

[0100] in, For nodes The equivalent heat capacity of the component; This represents the theoretical rate of change of temperature over time. Based on transient heat release power and coordinates Calculated fire source pairs of nodes The applied radiative heat flux density function; The surface convective heat transfer coefficient is calculated based on wind speed data; For convective heat transfer area; The temperature of the smoke in the fire environment; For the topological edge set based on step S1 Determined with nodes There exists a set of adjacent nodes for solid heat transfer; For adjacent node indexes; For nodes The theoretical temperature; The physical contact thermal resistance between nodes.

[0101] To identify blind spots or hidden structural damage in the underlying sensing layers, the local digital twin server quantifies the global synchronization error scalar between real-world perception and theoretical deduction. :

[0102] ;

[0103] in, Total number of monitoring nodes; Let L2 norm be the square of a vector.

[0104] Meanwhile, the local digital twin server extracts the current theoretical highest temperature across the entire network: .

[0105] Step S4: Strictly Dimensionally Aligned Graph Network Feature Fusion

[0106] The cloud-based master control server performs Z-Score normalization on the parameters with different dimensions received from steps S2 and S3, and constructs the node based on the local feature dimensions of a single node. Multidimensional local physical feature column vector :

[0107] ;

[0108] in, This is the matrix transpose symbol.

[0109] To effectively utilize the physical topological correlations of the bridge, the cloud-based main control server will generate the initial feature matrix. Input graph convolutional neural network, perform spatial physical information aggregation along the three-dimensional topological edges of the bridge:

[0110] ;

[0111] in, This is the current hidden layer index variable; For the first The hidden state matrix of the nodes in the layer; It is a linear rectification activation function; For topology-based Construct a component adjacency matrix that includes self-connected features; Let be the degree matrix of the adjacency matrix; For the first Layer input feature matrix; For the first Layer learnable weight parameter matrix.

[0112] through After the layer graph convolution is computed, the mean pooling operator (Readout) is called to generate graph embedding vectors representing the topological state of the full bridge. :

[0113] ;

[0114] in, This represents the total number of layers in the convolutional network. For nodes In the The hidden state column vector output by the layer.

[0115] Subsequently, the system will use the global synchronization error scalar. As an independent feature dimension, it is concatenated to Finally, a unified high-order digital feature vector is generated: .

[0116] Step S5: Large Model State Reasoning with Embedded Physical Knowledge Constraints

[0117] To address the issue of insufficient physical consistency of purely data-driven models in long-tail industrial scenarios, the large vision-language model on the cloud-based main control server receives... At the same time, the mechanism rule constraint module within the system architecture is activated.

[0118] This module reads the theoretical maximum temperature extracted from the entire network in step S3. Injecting physical knowledge constraints and penalty functions into the backpropagation and inference decision-making process of large models. :

[0119] ;

[0120] in, To maximize the function and ensure the penalty term is non-negative; This is a step activation function that outputs a scalar of 1 when the internally calculated value is greater than 0, and a scalar of 0 otherwise. The physical failure temperature threshold of the steel applied in step S1; Decode the multi-class fire risk probability distribution vector output by the large model; This is the confidence parameter extracted from the probability distribution vector to classify a risk as "high risk".

[0121] The cloud-based master server incorporates the penalty function into the overall objective optimization function of the large model for inference correction.

[0122] ;

[0123] in, The total loss function for joint optimization; For large model prediction results and empirical pseudo-labels The weakly supervised classification loss term between them; The weight hyperparameters are used to control the physical strength of the constraints; Regularization coefficients to prevent overfitting; The loss term is the network structure regularization term; after forced correction by physical mechanism rules, the large model outputs the final comprehensive health status score. .

[0124] Step S6: Safety margin quantization and 3D coordinate level hardware linkage

[0125] The server in the local control room quantifies the real-time load-bearing safety margin coefficient of the fire-affected components based on a material mechanics degradation model. :

[0126] ;

[0127] in, This refers to the real-time remaining load-bearing capacity of the component calculated based on the mapping relationship of the highest heated temperature. The initial design bearing capacity parameters of the component are loaded in step S1.

[0128] The programmable logic controller (PLC) reads data in real time via industrial fieldbus polling. and Data, when the safety margin coefficient is identified And overall health status score When the risk level is identified as high, the PLC immediately analyzes the three-dimensional physical coordinates of the fire source issued in step S2. The PLC calls the built-in inverse kinematics control algorithm to calculate the horizontal rotation and pitch servo motor deflection angles of the high-pressure fire monitor closest to the three-dimensional physical coordinates, driving the electromechanical equipment to perform three-dimensional spatial targeted and precise spraying fire extinguishing; simultaneously, it sends closure display commands and physical raising and lowering electrical commands to the LED traffic guidance screens at the outer end of the bridge, completing the whole system hardware-level control closed loop from perception and calculation to electromechanical linkage.

[0129] This disclosure also proposes a degradation assessment method under extreme dense smoke conditions: In actual engineering, when the target detection confidence of the explosion-proof high-definition camera is lower than the threshold for N consecutive seconds due to dense smoke obstruction, the front-end edge computing node automatically blocks the input of the visual perception channel.

[0130] At this point, the system enters a purely physical mechanism-driven mode: skipping visual localization in step S2 and... Calculate and directly extract the measured physical absolute temperature of all distributed fiber Bragg gratings in the full bridge from step S2. With instantaneous current gradient .

[0131] In the twin derivation of step S3, the radiative heat flux density function in the differential equation is... Forced to zero, relying solely on physical heat conduction between adjacent nodes. Perform thermal field evolution; under this degraded input, the large model adjusts the trigger weights of the knowledge constraint penalty function. This ensures that the system can still output a conservative health status score even when the user is visually blinded. And trigger a linked warning.

[0132] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A bridge fire risk status assessment method based on a large model, characterized in that, An application is made to a bridge fire risk assessment system. The system includes: underlying sensing and electromechanical execution equipment, front-end edge computing nodes, a local digital twin server, a cloud-based main control server, and a programmable logic controller. The underlying sensing and electromechanical execution equipment includes at least an explosion-proof high-definition camera, a distributed fiber Bragg grating temperature sensor, a fiber Bragg grating demodulator, a current transformer, an anemometer, a high-pressure fire monitor, and an LED traffic guidance screen. The method includes: The front-end edge computing node and the local digital twin server read the bridge's 3D BIM file and discretize it into a graph topology, and bind the communication addresses of the underlying sensing and electromechanical execution equipment to the corresponding 3D spatial physical nodes in the graph topology. The front-end edge computing node schedules the underlying sensing and electromechanical execution devices to collect visual and physical signals, analyze the three-dimensional physical coordinates and transient heat release power of the fire source, and convert the physical signals into measured physical temperature, radial physical temperature gradient and instantaneous current gradient. The local digital twin server receives the transient heat release power and the three-dimensional physical coordinates, solves the differential equation of the equivalent thermal resistance network, deduces the transient theoretical temperature of the three-dimensional physical node in the twin space, quantifies the global synchronization error scalar between real perception and theoretical deduction, and extracts the current theoretical highest temperature of the entire network. The cloud-based master control server performs standardization processing on the received physical parameters to construct local physical feature column vectors, performs spatial physical information aggregation through graph convolutional neural networks to generate graph embedding vectors, and concatenates the global synchronization error scalar to generate a unified high-order digital feature vector. The vision-language big model of the cloud master control server receives the unified high-order digital feature vector, and simultaneously activates the mechanism rule constraint module to convert the theoretical highest temperature of the entire network into a physical knowledge constraint penalty function and inject it into the reasoning and decision-making process of the big model for constraint correction, and outputs a comprehensive health status score. The local digital twin server quantifies the real-time load-bearing safety margin coefficient of the fire-affected component. When the real-time load-bearing safety margin coefficient and the comprehensive health status score trigger a preset high-risk threshold, the programmable logic controller analyzes the three-dimensional physical coordinates of the fire source, drives the high-pressure fire monitor to perform three-dimensional spatial targeted and precise spraying fire extinguishing, and sends a traffic closure command to the LED traffic guidance screen.

2. The bridge fire risk status assessment method based on a large model according to claim 1, characterized in that, The front-end edge computing node resolves the three-dimensional physical coordinates and transient heat release power of the fire source, including: The explosion-proof high-definition camera acquires real-time video frame sequences. After the front-end edge computing node performs adaptive histogram equalization, it calls the target detection network to extract the two-dimensional bounding box of the suspected fire source. The spatial intersection-over-union loss function is used to optimize the two-dimensional bounding box, and its distance cost... The calculation formula is: ； in, The index variable for the two-dimensional coordinate axes of the image, with values... shaft or axis; The center points of the predicted bounding box and the center points of the ground truth bounding box are on the coordinate axis. Normalized Euclidean distance in the direction; This is the penalty coefficient for angular deflection; The distance cost attenuation weighting index; Based on the pre-calibrated camera intrinsic and extrinsic parameter matrix, the front-end edge computing node projects the optimized two-dimensional bounding box into the three-dimensional physical coordinates of the fire source. Simultaneously, it matches the preset fire growth coefficient. The transient heat release power is output based on the fire growth model. : ； in, The start timestamp for the system to identify and confirm the fire source; This represents the actual duration of the fire.

3. The bridge fire risk status assessment method based on a large model according to claim 2, characterized in that, The front-end edge computing node converts physical signals into measured physical temperature, radial physical temperature gradient, and instantaneous current gradient, including: The fiber optic demodulator reads data from physical nodes embedded in three-dimensional space. The optical wavelength signal from the sensor is converted into a measured physical temperature by the front-end edge computing node based on the Bragg grating sensing equation. : ； in, The initial reference temperature for sensor calibration; The center wavelength is read in real time; To calibrate the initial wavelength; This refers to the inherent temperature sensitivity coefficient of the grating material; The front-end edge computing node calculates the radial physical temperature gradient of the node's cross-section based on the layered physical layout coordinates of the sensors within the bridge component's cross-section. : ； in, and They are nodes Measured temperatures of the outer and core layer sensors; The physical distance between the two sensor layers; The current transformer acquires a real-time current signal sequence through an analog-to-digital converter module, and obtains a filtered current sequence after noise reduction via a sliding window low-pass filter. The front-end edge computing node outputs the instantaneous current gradient using the discrete-time backward difference method. : ； in, This is the current filtered current value; This is the filtered current value from the previous sampling period; The sampling time interval for analog-to-digital conversion in the system hardware.

4. The bridge fire risk status assessment method based on a large model according to claim 3, characterized in that, The local digital twin server solves the equivalent thermal resistance network differential equation to derive the transient theoretical temperature and quantifies the global synchronization error scalar, including: Combining the flow field data collected by the anemometer, the local digital twin server constructs the node dynamic thermodynamic equilibrium mechanism, solves the differential equation of the equivalent thermal resistance network with convective heat transfer correction, and derives the node dynamic thermodynamic equilibrium mechanism. Transient theoretical temperature in twin space : ； in, For nodes The equivalent heat capacity of the component; This represents the theoretical rate of change of temperature over time. Based on transient heat release power and coordinates Calculated fire source pairs of nodes The applied radiative heat flux density function; The surface convective heat transfer coefficient is calculated based on wind speed data; For convective heat transfer area; The temperature of the smoke in the fire environment; For nodes There exists a set of adjacent nodes for solid heat transfer; For adjacent node indexes; For nodes The theoretical temperature; The physical contact thermal resistance between nodes; Based on the measured physical temperature Quantize the global synchronization error scalar : ； in, Total number of monitoring nodes; Let L2 norm be the square of a vector.

5. The bridge fire risk status assessment method based on a large model according to claim 4, characterized in that, The theoretical highest temperature across the entire network is transformed into a physical knowledge constraint penalty function and injected into the inference and decision-making process of the large model for constraint correction, including: The mechanism rule constraint module reads the theoretical maximum temperature of the entire network. Injecting physical knowledge into large models to constrain and penalize functions : ； in, This is a function to find the maximum value. It is a step activation function; This is a preset physical failure temperature threshold for steel. Decode the multi-class fire risk probability distribution vector output by the large model; This is a confidence parameter for classifying a risk level as high-risk. The cloud-based main control server incorporates the physical knowledge constraint penalty function into the overall objective optimization function of the large model for constraint correction. ； in, This is the total loss function; For weakly supervised classification loss items; These are pseudo-labels based on experience. The weight hyperparameters are used to control the physical strength of the constraints; The regularization coefficient is used. This is the network structure regularization loss term.

6. The bridge fire risk status assessment method based on a large model according to claim 5, characterized in that, The method also includes a degradation assessment method under extreme dense smoke conditions: When the explosion-proof high-definition camera's target detection confidence level falls below the judgment threshold for a continuous preset time due to dense smoke obscuring the view, the front-end edge computing node automatically blocks the input of the visual perception channel, and the system enters a purely physical mechanism-driven mode. In the pure physical mechanism driven mode, the calculation of the three-dimensional physical coordinates and the transient heat release power is skipped, and the measured physical temperature and the instantaneous current gradient are directly extracted. The radiation heat flux density function in the differential equation of the equivalent thermal resistance network is forced to zero, and the thermal field evolution is carried out solely by the physical conduction heat between adjacent nodes to deduce the transient theoretical temperature. At the same time, the trigger weight of the physical knowledge constraint penalty function is adjusted, thereby outputting the comprehensive health status score under the degraded working condition and triggering the linkage warning.

7. A bridge fire risk status assessment system based on a large model, characterized in that, The system is used to implement the bridge fire risk status assessment method based on a large model as described in any one of claims 1 to 6. The system includes: a bottom-level sensing and electromechanical execution device, a front-end edge computing node, a local digital twin server, a cloud main control server, and a programmable logic controller. The underlying sensing and electromechanical execution equipment includes at least an explosion-proof high-definition camera, a distributed fiber optic temperature sensor, a fiber optic demodulator, a current transformer, an anemometer, a high-pressure fire monitor, and an LED traffic guidance screen, used to collect multimodal visual and physical sensing data and execute electromechanical linkage commands. The front-end edge computing node is communicatively connected to the underlying sensing and electromechanical execution equipment, and is used to schedule the data collected by the underlying equipment, perform two-dimensional boundary box optimization of the fire source, three-dimensional physical coordinate calculation, and convert physical signals into measured physical parameters. The local digital twin server is communicatively connected to the front-end edge computing node and is used to construct the bridge graph topology, solve the equivalent thermal resistance network differential equation to deduce the transient theoretical temperature, quantify the global synchronization error scalar, and quantify the real-time load-bearing safety margin coefficient of the fire-affected component. The cloud-based main control server is communicatively connected to the local digital twin server. It has a built-in multimodal feature fusion computing unit and a visual-language large model, which is used to perform graph neural network spatial information aggregation to generate a unified high-order digital feature vector. It also embeds a physical knowledge constraint penalty function to perform constraint correction for large model reasoning and decision-making, and outputs a comprehensive health status score. The programmable logic controller is communicatively connected to the local digital twin server, the high-pressure fire monitor, and the LED traffic guidance screen. When the real-time load safety margin coefficient and the comprehensive health status score trigger a preset high-risk threshold, the controller analyzes the three-dimensional physical coordinates of the fire source, drives the high-pressure fire monitor to perform three-dimensional spatial targeted and precise spraying fire extinguishing, and issues a traffic closure command to the LED traffic guidance screen.

8. An electronic device, characterized in that, It includes at least one processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the bridge fire risk status assessment method based on a large model as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the bridge fire risk status assessment method based on a large model as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The method includes computer program instructions that, when executed by a computing device, cause the computing device to implement the bridge fire risk status assessment method based on a large model as described in any one of claims 1 to 6.

Images