A large model-based cable bridge vehicle fire intelligent research and judgment system and method

Abstract

The application discloses a cable bridge vehicle fire intelligent research and judgment system and method based on a large model, relates to the technical field of bridge fire prevention, and comprises the following modules: a man-machine interaction and parameter extraction module configured to extract and verify user requirements and working condition parameters through natural language interaction and multi-modal information identification; a pre-disaster prevention module configured to generate a conservative temperature prevention curve and a prevention report based on multi-working condition sampling and a temperature field prediction model; an emergency assessment module configured to perform rapid temperature field and flame form prediction, three-dimensional visual risk mapping, and generate an emergency assessment report; a post-disaster assessment module configured to perform main cable internal temperature inversion and cable surface temperature field prediction, damage classification, and generate a damage assessment report; and a large language model scheduling and management module configured to understand user intent, verify parameters, and schedule the pre-disaster prevention module, the emergency assessment module and the post-disaster assessment module to perform corresponding tasks.

Description

Technical Field

[0001] This invention relates to the field of bridge fire prevention technology, and in particular to an intelligent assessment system and method for vehicle fires on cable-stayed bridges based on a large model. Background Technology

[0002] In recent years, with the widespread application of long-span cable-stayed bridges in important projects such as those spanning seas and rivers, fire safety issues have become increasingly prominent. Bridge-vehicle fires have occurred frequently in recent years, and the high-temperature flames and hot smoke generated by burning vehicles pose a serious threat to the bridge's critical load-bearing components (such as main cables, suspenders, stay cables, and bridge towers). The mechanical properties of steel are extremely sensitive to temperature. When the temperature of the cable exceeds its material critical value, its strength and stiffness will significantly degrade, potentially leading to overall structural failure and catastrophic consequences.

[0003] Currently, the industry generally adopts a phased and fragmented approach to fire safety management of bridge cables, which has significant technical shortcomings at different stages of the fire lifecycle. 1. Pre-disaster preparedness phase: Lack of intelligent proactive defense measures and crude preparedness standards. Existing fire protection designs mainly rely on static preparedness based on standard fire curves, which are insufficient to cope with the complex thermal environment caused by changes in operating parameters such as fire source distance and wind speed in actual fires. Traditional methods use finite element software for multi-condition thermal analysis, which is costly and time-consuming, and cannot achieve rapid and comprehensive risk assessment. This results in preparedness curves that are often too conservative or insufficient, which may lead to economic waste and potential safety hazards.

[0004] 2. Fire Emergency Response Phase: Lack of rapid risk situation awareness and visualization capabilities, resulting in insufficient decision support. When a fire occurs, emergency personnel struggle to quickly predict the flame spread and temperature distribution of the main cable. Current technology cannot predict temperature field and flame morphology within minutes, nor can it intuitively map the prediction results onto a 3D model of the bridge for risk visualization. This leads to a lack of real-time and accurate data support for command and decision-making, hindering the efficient development of evacuation, firefighting, and traffic control plans, and potentially causing the missed window of opportunity for optimal emergency response.

[0005] 3. Post-disaster assessment phase: The assessment process is fragmented and inefficient, making it difficult to quickly guide repairs. Post-disaster assessments face severe challenges in terms of timeliness, accuracy, and intelligence. Although existing methods have addressed some of the problems in post-disaster assessment, from a full lifecycle perspective, the data, models, and processes of the pre-disaster, emergency response, and post-disaster stages are disconnected, failing to form a unified and coherent management loop. From receiving a disaster report to generating the final assessment report, multiple independent systems and manual intervention are required, making the process cumbersome and severely impacting the efficiency of post-disaster recovery and reconstruction.

[0006] In summary, current technologies lack an intelligent, integrated management system capable of covering the entire lifecycle of a fire—from pre-disaster to during and post-disaster. The disconnect between different stages and the insufficient level of intelligence make it difficult for bridge cables to fundamentally shift from a "passive response" to "proactive early warning, rapid response, and accurate assessment" when facing vehicle fire risks. Therefore, developing an integrated, large-model-based method for handling vehicle fires on cable-stayed bridges, achieving seamless integration and intelligent decision-making at each stage, has become an urgent need to ensure the safe operation of bridges. Summary of the Invention

[0007] This solution addresses the problems and needs raised above by proposing an intelligent assessment system and method for vehicle fires on cable bridges based on a large model. The system achieves the aforementioned technical objectives and brings about several other technical benefits by adopting the following technical features.

[0008] One objective of this invention is to propose an intelligent assessment system for cable bridge vehicle fires based on a large model, comprising: a human-computer interaction and parameter extraction module, a large language model scheduling and management module, a pre-disaster preparedness module, an emergency assessment module, and a post-disaster assessment module; wherein, The human-computer interaction and parameter extraction module is configured to extract and verify user needs and operating condition parameters through natural language interaction and multimodal information recognition. The pre-disaster preparedness module is configured to generate conservative temperature preparedness curves and preparedness reports based on multi-condition sampling and temperature field prediction models. The emergency assessment module is configured to perform rapid temperature field and flame pattern prediction, three-dimensional visualization of risk mapping, and generate an emergency assessment report. The post-disaster assessment module is configured to perform internal temperature inversion of the main cable and prediction of the cable surface temperature field, damage classification, and generate a damage assessment report; and The large language model scheduling and management module is configured to understand user intent, verify parameters, and schedule the pre-disaster defense module, emergency assessment module, and post-disaster assessment module to perform corresponding tasks.

[0009] In addition, the intelligent fire assessment system for cable bridge vehicles based on a large model according to the present invention may also have the following technical features: In one example of the present invention, the human-computer interaction and parameter extraction module includes: The multimodal information fusion unit is configured to compare, supplement, and fuse the operating condition parameters described by the user in natural language with the visual parameters automatically identified from on-site images or videos to generate a more complete and accurate set of standardized operating condition parameters. The parameter conflict resolution unit is configured to initiate a follow-up questioning or confirmation process when there is a contradiction between the natural language description parameters and the visual recognition parameters, guiding the user to clarify and ensuring the reliability of the input parameters.

[0010] In one example of the present invention, the pre-disaster preparedness module includes: The first parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized first working condition parameter dataset. The first model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the first working condition parameter dataset, and generate a call instruction for the cable surface temperature field prediction model. The first deep learning unit is configured to load and obtain the corresponding cable surface temperature field prediction model based on the cable surface temperature field prediction model call command, perform temperature field spatiotemporal evolution simulation calculation, and output the cable surface temperature field prediction result. The first report generation unit is configured to integrate the first working condition parameter dataset and the temperature field prediction results to generate a report containing parameter vectors, temperature protection curves for bridge cable vehicle fires, protection indicators, and protection recommendations.

[0011] In one example of the present invention, the first deep learning unit includes: The first fire simulation dataset sub-unit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The first data preprocessing and standardization subunit is configured to map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the operating parameters and temperature values ​​to form an operating parameter set, which is then divided into a training set, a validation set, and a test set. The first adversarial network model subunit is configured to generate a first conditional generative adversarial network model consisting of a generator A and a discriminator A. The generator A is configured to generate a predicted temperature curve value based on a noise vector and normalized working condition parameters. The discriminator A is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The first definition is a weighted loss function subunit, configured as generator A adopts a weighted combined loss function that includes adversarial loss and physical constraints, and trains the first conditional generative adversarial network model through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. The first model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between discriminator A and generator A to train the first conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning efficiency is 2 × 10⁻⁶. -4The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The first model validation and deployment subunit is configured to evaluate the performance of the first conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The mean absolute error (MAE) is used to evaluate the first-condition generative adversarial network model, where the coefficient of determination R0 is... 2 >0.9, with a mean absolute error (MAE) <50℃. After successful verification, the weight file of the fully trained first-condition generative adversarial network model is deployed to a local professional model library for use by large language models.

[0012] In one example of the present invention, the generator A includes a first input layer, three first fully connected hidden layers, and a first output layer; the first input layer is a concatenated vector of a noise vector and normalized working condition parameters, used for data input and feature initialization; the three first fully connected hidden layers are connected sequentially, each followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively; the first output layer includes a first fully connected layer and a Tanh activation function, used to output temperature prediction values ​​for multiple standard altitude points; The discriminator A is a two-stream network structure, including a first conditional stream, a first data stream, a first concatenation layer, two second fully connected hidden layers, and a second output layer. The first conditional stream includes a second fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The first data stream includes a third fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The first concatenation layer concatenates the feature vectors output from the first conditional stream and the first data stream into a fused feature vector. The two second fully connected hidden layers are connected sequentially, with each second fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The second output layer includes a fourth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

[0013] In one example of this invention, the weighted combination loss function L used by generator A total From the counter-loss L adv and physical constraint loss L smooth Weighted according to the following formula: Among them, the resistance loss L adv The calculation formula is: Among them, physical constraint loss L smooth The calculation formula is: In the formula, λ is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data (C) represents the data distribution under actual operating conditions; D(.) represents the output probability of discriminator A; G(.) represents the output of generator A; N represents the length of the temperature sequence; T i This represents the predicted temperature value at the i-th elevation point; The loss function used by discriminator A in the fire temperature prediction model The expression is: In the formula, Discriminator A represents the probability of outputting "true" under given operating conditions.

[0014] In one example of the present invention, the emergency assessment module includes: The second parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized second working condition parameter dataset. The second model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the standardized second working condition parameter dataset, and generate a deep learning model calling instruction for predicting the temperature field and flame morphology of bridge cables. The second deep learning unit is configured to load and train the corresponding temperature field and flame pattern prediction model based on the deep learning model call instruction, perform spatiotemporal evolution simulation calculation of temperature field and flame pattern, and output the prediction results of temperature field and flame pattern of bridge cable. The second report generation unit is configured to map the predicted results of the bridge cable temperature field and flame pattern onto the bridge's three-dimensional model, and perform dynamic risk analysis and visualization; it integrates key operating parameters, bridge cable temperature field distribution map, flame pattern data, risk level assessment and emergency measure recommendations to generate a structured emergency assessment report.

[0015] In one example of the present invention, the second deep learning unit includes: The second fire simulation dataset sub-unit is configured to generate flame morphology data under different standardized working condition parameter combinations through simulation software. The second data preprocessing and standardization subunit is configured to map the non-uniform spatial temperature data output by the simulation model to a preset regular grid through cubic spline interpolation, normalize the temperature data to form a set of operating parameters, and divide the set of operating parameters into a training set, a validation set and a test set. The second adversarial network model subunit is configured to construct a second conditional generative adversarial network model. The second conditional generative adversarial network model adopts a generator-discriminator dual-path system, including a generator B and a structure discriminator B. The generator B is configured to generate flame shape prediction values ​​based on noise vectors and normalized chemical condition parameters. The discriminator B is configured to distinguish between generated flame shapes and real flame shapes and output a discrimination probability value. The second definition is a weighted loss function subunit, configured so that generator B adopts a weighted combined loss function that includes adversarial loss and physical constraints. The second conditional generative adversarial network model is trained through the training set, wherein the adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the flame shape reasonably satisfies the empirical formula. The second model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It uses an alternating iterative update strategy of discriminator B and generator B to train the second conditional generative adversarial network model through the training set. During the training process, the training effect is monitored through the validation set and the training parameters are adjusted. The second model validation and deployment subunit is configured to evaluate the performance of the second conditional generative adversarial network model on a test set, using the coefficient of determination R. 2 The mean absolute error (MAE) is used to evaluate the second-conditional generative adversarial network model, where the coefficient of determination R0 is... 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained second-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

[0016] In one example of the invention, the weighted combination loss function Adversarial loss from generator B and physical constraints Together they constitute, specifically as follows: in, The adversarial loss of generator B, whose goal is to deceive discriminator B, is calculated as follows: in, This is the physical constraint loss, used to ensure the flame shape conforms to the empirical formula, and its calculation formula is as follows: In the formula, λ1 is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data The distribution of real data; D flame (·) represents the output probability of discriminator B; G flame(·) represents the output of generator B; H pred H represents the predicted flame height. empirical The results are calculated using empirical formulas based on the classic fire plume model; P represents the fire source power. The mass combustion rate.

[0017] In one example of the present invention, the second deep learning unit further includes: The third fire simulation dataset sub-unit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The third data preprocessing and standardization subunit is configured to uniformly map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the working parameters and temperature values ​​to form a working parameter set, which is then divided into a training set, a validation set, and a test set. The third adversarial network model subunit is configured to construct a third conditional generative adversarial network model, which consists of a generator C and a discriminator C. The generator C is configured to generate predicted temperature curve values ​​based on noise vectors and normalized working condition parameters. The discriminator C is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The third definition is a weighted loss function subunit, configured so that the generator C adopts a weighted combined loss function that includes adversarial loss and physical constraints. The third conditional generative adversarial network model is trained through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. The third model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between the discriminator C and the generator C to train the third conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning rate is 2×10⁻⁶. -4 The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The third model validation and deployment subunit is configured to evaluate the performance of the third conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The third-condition generative adversarial network model is evaluated using the mean absolute error (MAE) and the coefficient of determination (R²). 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained third-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

[0018] In one example of the present invention, the second report generation unit includes: The spatial mapping subunit is configured to establish a spatial mapping relationship between the bridge's 3D model and the predicted data, and to render the temperature field on the surface of the bridge cables using a heat map. The flame model sub-unit is configured to generate a dynamic flame model based on flame morphology characteristics and uses a particle system to simulate the flame morphology change process. The spatial relationship calculation subunit is configured to calculate the spatial relationship between flames and bridge cables in real time, and to perform licking detection and risk area identification. The risk level sub-unit is configured to display the risk level through color coding; where red represents high risk, yellow represents medium risk, and green represents low risk.

[0019] In one example of the present invention, the post-disaster assessment module includes: The third parameter extraction unit is configured to identify key working condition parameters and main cable protection type in bridge cable vehicle fire scenarios based on a large language model, and obtain a standardized third working condition parameter dataset including fire source power, fire source distance and bridge surface wind speed. The third model scheduling and management unit is configured to call the local knowledge base for parameter verification based on the standardized third working condition parameter dataset, and generate calling instructions for the cable surface temperature field prediction model and the main cable internal temperature inversion model. The third deep learning unit is configured to use the calling command based on the cable surface temperature field prediction model and the main cable internal temperature inversion model to load the bridge main cable internal temperature inversion model and the bridge cable surface temperature field prediction model and train them; according to the verified working condition parameters, the bridge main cable internal temperature inversion model is used to output the predicted value of the time-temperature inside the main cable, and the cable surface temperature field prediction model is used to output the predicted value of the highest temperature on the surface of the main cable. The third report generation unit is configured to compare the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the safety threshold of the main cable material to classify the damage; and to integrate the standardized third working condition parameter dataset, the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the damage level to generate a report containing parameter vectors, the time-temperature curve inside the main cable, the maximum temperature on the main cable surface, the damage level, and repair recommendations.

[0020] In one example of the present invention, the third deep learning unit includes: The first data simulation subunit is configured to build a bridge vehicle fire simulation model, and the simulation model calculates and outputs peak temperature-height data of the main cable surface under different working conditions. The first dataset is divided into sub-units, configured to construct a dataset from the peak temperature-height data output by the simulation model, the experimental data on the surface of the main cable, and the corresponding working parameters. After preprocessing, the dataset is divided into training set, validation set, and test set. The first inversion model training subunit is configured to construct a bridge main cable surface temperature prediction model based on a conditional generative adversarial network. This model has a generator D and a discriminator D. The generator D is configured to generate predicted temperature curves based on noise vectors and normalized working condition parameters. The discriminator D is configured to distinguish between the generated temperature curves and the actual temperature curves and output a discrimination probability value. The bridge main cable surface temperature prediction model using a loss function is trained on the training set. During the training process, the training effect is monitored and the training parameters are adjusted through the validation set. The performance of the bridge main cable surface temperature prediction model is evaluated on the test set to obtain the trained bridge main cable surface temperature prediction model. The first prediction value output subunit is configured to input the operating condition parameters into the trained bridge main cable surface temperature prediction model and output the predicted value of the highest surface temperature of the main cable.

[0021] In one example of the present invention, the third deep learning unit includes: The second data simulation subunit is configured to input operating parameters into the bridge vehicle fire simulation model, and output the time-temperature data of the main cable surface by the simulation model; input the time-temperature data of the main cable surface into the main cable heat transfer simulation model, and output the time-temperature data of the main cable interior. The second dataset is divided into sub-units, configured to construct a dataset from the output time-temperature data inside the main cable, the experimental data inside the main cable, and the corresponding operating parameters. After preprocessing, the dataset is divided into training set, validation set, and test set. The second inversion model training subunit is configured to construct a bridge main cable internal temperature inversion model based on LSTM conditional generative adversarial network. It adopts a weighted combination loss function that includes adversarial loss and physical constraints. The bridge main cable internal temperature inversion model is trained by the training set. During the training process, the training effect is monitored and the training parameters are adjusted by the validation set. The performance of the bridge main cable internal temperature inversion model is evaluated on the test set to obtain the trained bridge main cable internal temperature inversion model. The second prediction output subunit is configured to input operating parameters into the trained bridge main cable internal temperature inversion model and output the predicted time-temperature value inside the main cable.

[0022] In one example of the present invention, the heat transfer simulation model of the main cable is derived using a rectangular finite element model. The rectangular finite element model consists of the main cable steel wires and the air between the steel wires. The internal steel wire dimensions and porosity are consistent with the experimental model. There is surface-to-surface thermal radiation between the steel wires. The radiating surface is defined as a diffuse reflective surface and the emissivity is set. The radiation direction is set to be controlled by the opacity. The opacity of the steel wires and the air is set to opaque and transparent, respectively. The air is set to be non-flowing and the thermal convection of the air is ignored. The main cable structure is regarded as a porous medium of "steel wire-void". By fitting the equivalent thermal conductivity and correcting the equivalent specific heat capacity, a simplified equivalent model of the main cable structure of steel wire-void is established. The input is the time-temperature data of the main cable surface, and the output is the time-temperature data of the main cable interior.

[0023] In one example of the invention, the external temperature for damage grading is a predicted value of the highest temperature on the main cable surface. The internal temperature is measured using the highest surface temperature of the outer steel wire of the main cable. Measured by the predicted time-temperature value inside the main cable; The safety threshold for main cable materials includes the first surface temperature threshold. Second surface temperature threshold and the temperature threshold of the outer steel wire Set the over-threshold time threshold. Cumulative time exceeding the threshold The cumulative duration during which the surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold. when When, it is determined to be undamaged; when ,and If it is, then it is judged as a minor injury; when ,and If it is, then it is judged as a minor injury; when ,and If the injury is moderate, it is considered a moderate injury; if the injury is severe, it is considered a moderate injury. ,and If so, it is determined to be a severe injury.

[0024] Another objective of this invention is to propose a method for analyzing vehicle fires in cable bridges based on a large model, as described above, comprising the following steps: S10: Users input their needs in natural language through the human-computer interaction and parameter extraction module's human-computer interaction interface; S20: The large language model scheduling and management module understands user intent and extracts and verifies standardized working condition parameter sets through multi-turn dialogue; S30: The large language model scheduling and management module schedules the parameter set to the pre-disaster defense module, emergency assessment module, or post-disaster assessment module according to the intent. S40: The invoked pre-disaster preparedness module, emergency assessment module, or post-disaster assessment module performs calculations and analyses, and generates corresponding structured professional reports; S50: The processing system outputs the final structured report to the user.

[0025] Compared with the prior art, the present invention has the following beneficial effects: 1. This application achieves intelligent and integrated management throughout the entire lifecycle, fundamentally changing the traditional phased and fragmented passive response model. It organically integrates the three originally independent technical aspects of "pre-disaster preparedness," "in-disaster emergency response," and "post-disaster assessment" through an intelligent scheduling hub that integrates a large language model, constructing a complete business closed loop. The system can automatically identify scenario intents based on user needs and call the corresponding modules, realizing a seamless operation from proactive prevention and rapid response to precise repair, significantly improving the systematic, proactive, and forward-looking nature of bridge cable fire safety management.

[0026] 2. Significantly improves the intelligence level and decision-making efficiency of each stage. In each core module, this application deeply integrates advanced deep learning models trained with physical constraints, replacing traditional methods that rely on empirical formulas or time-consuming high-fidelity numerical simulations. This reduces core computational tasks such as generating temperature protection curves, performing real-time fire simulations, and retrieving internal temperature histories from hours or even days to minutes or even seconds, achieving true "rapid analysis and real-time response." This provides unprecedentedly efficient and intelligent decision support for engineering design, emergency command, and post-disaster assessment.

[0027] 3. An "evolvable and scalable" intelligent system kernel has been constructed, ensuring the continuous technological leadership and long-term application value. This application enables the system to self-optimize and grow through iterative learning and module upgrade management modules. On the one hand, the system can continuously fine-tune its internal predictive models and knowledge base based on user feedback, becoming more accurate with use; on the other hand, based on a loosely coupled modular architecture, new algorithms and functions can be easily integrated into the system like "building blocks," achieving smooth upgrades without the need for complete overhaul. This effectively solves the pain point of traditional software systems easily becoming rigid and obsolete with technological development, greatly extending the technical lifecycle of this system.

[0028] 4. It provides intuitive and accurate visualization and report generation capabilities, significantly lowering the professional threshold and improving communication efficiency. The system not only offers dynamic, three-dimensional visualization of flame and temperature field risks in the emergency module, enabling non-professionals to intuitively understand the risk situation; more importantly, all three core modules can automatically generate well-structured, richly illustrated, and clearly concluded professional reports. This completely changes the inefficient model that previously relied on experts manually interpreting data and writing reports, freeing engineers from tedious paperwork so they can focus on higher-value decisions, while also greatly facilitating technical communication and collaboration between different departments and units.

[0029] 5. Through a localized and modular system architecture, high security, stability, and maintainability are ensured while pursuing high performance. All core models and data are deployed locally, eliminating the risk of leakage of sensitive engineering data. The high cohesion and low coupling of each functional module not only facilitates expansion but also ensures that the maintenance, updates, and troubleshooting of individual modules do not affect the overall system operation, significantly improving system stability and maintainability and reducing long-term operation and maintenance costs.

[0030] 6. By introducing multimodal information recognition capabilities, the system can directly and objectively extract key parameters (such as flame size, smoke color, and structural burn marks) from on-site images, effectively making up for the omissions, subjectivity, or misjudgments that may exist if relying solely on manual descriptions. This further improves the accuracy and reliability of the input data, laying a more solid data foundation for subsequent intelligent prediction and evaluation.

[0031] 7. This application not only addresses the specific pain points of existing technologies in various stages, such as "slow computation, lack of intuitiveness, fragmented processes, and high barriers to entry," but also constructs an intelligent ecosystem capable of self-learning and continuously adapting to future technological developments by introducing an "evolvable" design concept. This establishes a new and sustainable technological paradigm for the safe operation and maintenance of bridge cables and similar major engineering structures throughout their entire lifecycle, possessing significant engineering application value and long-term economic benefits.

[0032] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0034] Figure 1 This is a schematic diagram of the intelligent fire assessment system for cable bridge vehicles based on a large model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a cable bridge vehicle fire intelligent analysis system based on a large model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the first deep learning unit according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the generator according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the discriminator according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the fire emergency assessment module according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of the second deep learning unit according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a bridge vehicle fire post-disaster assessment module according to an embodiment of the present invention; Figure 9 This is a flowchart of the bridge main cable internal temperature inversion model and the cable surface temperature field prediction model according to an embodiment of the present invention. Figure 10 This is a schematic diagram of the structure of the third deep learning unit according to an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] According to a first aspect of the present invention, a large-scale model-based intelligent analysis system for vehicle fires on cable-stayed bridges, such as... Figure 1 and Figure 2 As shown, it includes: a human-computer interaction and parameter extraction module, a large language model scheduling and management module, a pre-disaster defense module, an emergency assessment module, and a post-disaster assessment module; among which, The human-computer interaction and parameter extraction module is configured to extract and verify user needs and operating condition parameters through natural language interaction and multimodal information recognition. The pre-disaster preparedness module is configured to generate conservative temperature preparedness curves and preparedness reports based on multi-condition sampling and temperature field prediction models. The emergency assessment module is configured to perform rapid temperature field and flame pattern prediction, three-dimensional visualization of risk mapping, and generate an emergency assessment report. The post-disaster assessment module is configured to perform internal temperature inversion of the main cable and prediction of the cable surface temperature field, damage classification, and generate a damage assessment report; and The large language model scheduling and management module is configured to understand user intent, verify parameters, and schedule the pre-disaster defense module, emergency assessment module, and post-disaster assessment module to perform corresponding tasks.

[0037] The processing system works as follows: Users input their needs in natural language through the human-computer interaction and parameter extraction module's interface; the large language model scheduling and management module understands the user's intent and extracts and verifies a standardized set of working condition parameters through multi-turn dialogues; the large language model scheduling and management module schedules the parameter set to the pre-disaster preparedness module, emergency assessment module, or post-disaster assessment module according to the intent; the invoked pre-disaster preparedness module, emergency assessment module, or post-disaster assessment module performs calculations and analyses, and generates corresponding structured professional reports; the processing system outputs the final structured report to the user.

[0038] This processing system achieves intelligent and integrated management throughout the entire lifecycle, fundamentally changing the traditional phased and fragmented passive response model. This application organically integrates three originally independent technical aspects—"pre-disaster preparedness," "disaster emergency response," and "post-disaster assessment"—through an intelligent scheduling hub integrating a large language model, constructing a complete business closed loop. The system can automatically identify scenario intents based on user needs and invoke corresponding modules, achieving seamless operation throughout the entire process from proactive prevention and rapid response to precise repair, significantly improving the systematic, proactive, and forward-looking nature of bridge cable fire safety management.

[0039] This processing system significantly improves the intelligence level and decision-making efficiency of each stage. In each core module, this application deeply integrates advanced deep learning models trained with physical constraints, replacing traditional methods that rely on empirical formulas or time-consuming high-fidelity numerical simulations. This reduces core computational tasks such as generating temperature protection curves, performing real-time fire simulations, and retrieving internal temperature histories from hours or even days to minutes or even seconds, achieving true "rapid analysis and real-time response." This provides unprecedentedly efficient and intelligent decision support for engineering design, emergency command, and post-disaster assessment.

[0040] This processing system constructs an "evolvable and scalable" intelligent system kernel, ensuring the continuous technological leadership and long-term application value. Through iterative learning and module upgrade management modules, this application enables the system to possess self-optimization and growth capabilities. On the one hand, the system can continuously fine-tune its internal predictive models and knowledge base based on user feedback, becoming more accurate with use; on the other hand, based on a loosely coupled modular architecture, new algorithms and functions can be easily integrated into the system like "building blocks," achieving smooth upgrades without the need for complete overhaul. This effectively solves the pain point of traditional software systems easily becoming rigid and obsolete with technological development, greatly extending the technical lifecycle of this system.

[0041] This processing system provides intuitive and accurate visualization and report generation capabilities, significantly lowering the professional threshold and improving communication efficiency. The system not only offers dynamic, three-dimensional visualization of flame and temperature field risks in the emergency module, enabling non-professionals to intuitively understand the risk situation; more importantly, all three core modules can automatically generate well-structured, richly illustrated, and clearly concluded professional reports. This completely changes the inefficient model that previously relied on experts manually interpreting data and writing reports, freeing engineers from tedious paperwork so they can focus on higher-value decision-making, while also greatly facilitating technical communication and collaboration between different departments and units.

[0042] This processing system, through its localized and modular architecture, ensures high security, stability, and maintainability while pursuing high performance. All core models and data are deployed locally, eliminating the risk of sensitive engineering data leakage. The high cohesion and low coupling of each functional module not only facilitates expansion but also ensures that the maintenance, updates, and troubleshooting of individual modules do not affect the overall system operation, significantly improving system stability and maintainability and reducing long-term maintenance costs.

[0043] By introducing multimodal information recognition capabilities, this processing system can directly and objectively extract key parameters (such as flame size, smoke color, and structural burn marks) from on-site images. This effectively compensates for the omissions, subjectivity, or misjudgments that may occur when relying solely on manual descriptions, further improving the accuracy and reliability of the input data and laying a more solid data foundation for subsequent intelligent prediction and evaluation.

[0044] This processing system not only addresses the specific pain points of existing technologies in various stages, such as slow computation, lack of intuitiveness, fragmented processes, and high barriers to entry, but also constructs an intelligent ecosystem capable of self-learning and continuously adapting to future technological developments by introducing an "evolvable" design concept. This establishes a new and sustainable technological paradigm for the safe operation and maintenance of bridge cables and similar major engineering structures throughout their entire lifecycle, possessing significant engineering application value and long-term economic benefits.

[0045] In one example of the present invention, the human-computer interaction and parameter extraction module includes: The multimodal information fusion unit is configured to compare, supplement, and fuse the operating condition parameters described by the user in natural language with the visual parameters automatically identified from on-site images or videos to generate a more complete and accurate set of standardized operating condition parameters. The parameter conflict resolution unit is configured to initiate a follow-up questioning or confirmation process when there is a contradiction between the natural language description parameters and the visual recognition parameters, guiding the user to clarify and ensuring the reliability of the input parameters.

[0046] In one example of the present invention, the pre-disaster preparedness module includes: The first parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized first working condition parameter dataset. The first model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the first working condition parameter dataset, and generate a call instruction for the cable surface temperature field prediction model. The first deep learning unit is configured to load and obtain the corresponding cable surface temperature field prediction model based on the cable surface temperature field prediction model call command, perform temperature field spatiotemporal evolution simulation calculation, and output the cable surface temperature field prediction result. The first report generation unit is configured to integrate the first working condition parameter dataset and the temperature field prediction results to generate a report containing parameter vectors, temperature protection curves for bridge cable vehicle fires, protection indicators, and protection recommendations.

[0047] This fire protection system represents a paradigm shift from "manually driven" to "intelligently automated." Traditional methods rely on engineers manually setting up massive amounts of operational conditions and performing tedious post-processing. This application uses a large language model as the intelligent scheduling hub, integrating multiple discrete steps such as parameter parsing, operational condition combination, model invocation, result integration, and report generation into a fully automated pipeline. Users only need to describe their fire protection requirements in natural language, and the system can automatically complete all the work that previously required weeks of manual labor. This reduces the efficiency of fire protection curve generation from days or even tens of days to minutes, completely solving the core pain point of low efficiency in traditional methods.

[0048] This fire protection system employs a dual approach of "physical laws + data-driven" to significantly enhance the scientific rigor and reliability of the design curve. This application is not simply a data fitting exercise. On one hand, the deep learning model incorporates physical constraints during training, ensuring its predictions conform to fundamental thermodynamic principles. On the other hand, during the design curve generation stage, a multi-condition coverage strategy and conservative envelope extraction are employed, along with stringent physical corrections (such as monotonicity constraints), ensuring that the final design curve covers various potential risk scenarios while meeting engineering conservatism requirements. Its scientific rigor and reliability far surpass traditional results relying on single conditions or empirical formulas.

[0049] This fire protection system ensures the complete localization of core data and models, fundamentally eliminating the risk of sensitive information leakage. All processing, calculation, and storage of sensitive data involving bridge structural parameters, geographic information, etc., are completed within the organization's internal network of server clusters, completely eliminating the risk of data leakage that may result from using cloud services. This feature makes this application particularly suitable for bridge engineering projects with extremely high information security requirements, providing a secure and reliable solution for the application of intelligent technologies in critical infrastructure.

[0050] This fire protection system lowers the professional threshold, empowering frontline designers to produce expert-level results. Through a natural language interface and a fully automated process, this application encapsulates complex fire simulation and fire protection curve generation technologies into an easy-to-use tool. Non-senior experts or frontline designers do not need in-depth knowledge of fire dynamics or numerical simulation to quickly generate professional fire protection schemes that meet regulatory requirements, greatly reducing the technical application threshold and facilitating the promotion and popularization of advanced design methods.

[0051] This fire protection system adopts a modular and standardized system architecture, possessing excellent maintainability and scalability. Each core module of the system is encapsulated using containerization technology and communicates through standard API interfaces. This highly cohesive and loosely coupled architecture design allows any component in the system (such as a new temperature prediction model or an updated design specification knowledge base) to be independently upgraded or replaced without requiring a complete system reconstruction. This ensures that this application can continuously absorb technological advancements, adapt to future changes in specifications and requirements, and possesses a long lifecycle and application value.

[0052] In one example of the present invention, the first parameter extraction unit includes: The parameter entity recognition subunit is configured to utilize the semantic understanding capabilities of a large language model to identify fire scene parameter entities that are implicit or explicit in the user input. The parameter set subunit is configured to initiate inquiry logic for undefined key condition parameters; and to initiate confirmation and correction logic for ambiguous or unreasonable condition parameters, ultimately outputting a complete and error-free standard chemical condition parameter set.

[0053] In one example of the present invention, the standardized operating condition parameter set includes: fire source power, fire source distance, and bridge deck wind speed, wherein the fire source power is a fixed value; the fire source distance parameter set is obtained by sampling in a uniform distribution in [0, s], where s is the maximum fire source distance, and the maximum fire source distance is the distance between the adjacent lane of the emergency lane and the cable; the bridge deck wind speed is sampled according to a Weibull distribution; its probability density function is... The expression is as follows: In the formula, v represents the annual average wind speed of the design area, and k and c are two parameters of the Weibull distribution. k is called the shape parameter, and c is called the scale parameter. When c=1, it is called the standard Weibull distribution. The value of the shape parameter k is adjusted according to local conditions to make the wind speed distribution conform to reality. The smaller the k, the greater the variation in annual average wind speed; the larger the k, the smaller the variation in annual average wind speed.

[0054] This involves arranging and combining three operating parameters to achieve comprehensive coverage of different fire scenarios through this hybrid sampling strategy.

[0055] In one example of the present invention, such as Figure 3 As shown, the first deep learning unit includes: The first fire simulation dataset subunit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The working condition parameters include fire source power, bridge deck wind speed and fire source distance. The first data preprocessing and standardization subunit is configured to map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the operating parameters and temperature values ​​to form an operating parameter set. The operating parameter set is then divided into a training set, a validation set, and a test set. For example, the same processing is performed on the experimental data to form an operating parameter set. The two sets of operating parameter sets are then mixed and divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0056] The first adversarial network model subunit is configured to generate a first conditional generative adversarial network model consisting of a generator A and a discriminator A. The generator A is configured to generate a predicted temperature curve value based on a noise vector and normalized working condition parameters. The discriminator A is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The first definition is a weighted loss function subunit, configured as generator A adopts a weighted combined loss function that includes adversarial loss and physical constraints, and trains the first conditional generative adversarial network model through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. The first model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between discriminator A and generator A to train the first conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning efficiency is 2 × 10⁻⁶. -4 The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The first model validation and deployment subunit is configured to evaluate the performance of the first conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The mean absolute error (MAE) is used to evaluate the first-condition generative adversarial network model, where the coefficient of determination R0 is... 2 >0.9, with a mean absolute error (MAE) <50℃. After successful verification, the weight file of the fully trained first-condition generative adversarial network model is deployed to a local professional model library for use by large language models.

[0057] In one example of this invention, constructing a bridge vehicle fire simulation model specifically includes: The two-dimensional cable plane is simplified into a one-dimensional vertical bar, and any position of the cable can be represented by the height position of the vertical bar, thus simplifying the two-dimensional spatial coordinates into a one-dimensional position. The vehicle is simplified into a cube with a fire source on its surface. Recommended maximum grid sizes are given for different vehicle types: 20cm for a maximum heat release rate of 0-5MW; 25cm for 5-10MW; 30cm for 10-15MW; 35cm for 15-20MW; and 50cm for 20-200MW. The maximum heat release efficiency of the flame is simulated for 8-12 minutes to equivalently replace the highest temperature of the actual flame simulated for 1.5 hours. Temperature measuring points are set at equal intervals along the height direction on the surface of the vertical rod according to the required accuracy.

[0058] In one example of the present invention, the operating parameters include: fire source intensity, bridge deck wind speed, and fire source distance; wherein, the fire source intensity is densely sampled in a uniform distribution in the range of 0-50MW, and sparsely sampled in a uniform distribution in the range of 50-200MW; the bridge deck wind speed is sampled in a uniform distribution in the range of 0-15m / s; and the fire source distance is the distance between the emergency lane and the adjacent lane from the edge of the main cable.

[0059] In one example of this invention, experimental data is obtained using the Chinese invention patent application number 2025111956407, "Real-time Hybrid Test Platform and Implementation Method for Vehicle-Fire-Wind Force Mixed Test of Substructure of Long-Span Bridge," to conduct full-scale or large-scale model fire resistance tests on cables under controlled parameter wind, fire, and force coupling conditions, directly collecting temperature field data at height points. The repeatable, measurable, and high-fidelity fire environment provided by this platform greatly ensures the diversity and reliability of the original data, laying the foundation for a sample set for training the temperature field generation model.

[0060] In one example of the present invention, preprocessing the dataset includes: mapping the non-uniform height-temperature data output by the simulation model to a standard height sampled at equal intervals within 0-35m using linear interpolation; and normalizing the working condition parameter matrix and the temperature value matrix respectively.

[0061] In one example of the present invention, such as Figure 4 and Figure 5 As shown, the generator A includes a first input layer, three first fully connected hidden layers, and a first output layer. The first input layer is a concatenated vector of a noise vector and normalized working condition parameters, used for data input and feature initialization. The three first fully connected hidden layers are connected sequentially, with each first fully connected hidden layer followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively. The first output layer includes a first fully connected layer and a Tanh activation function, used to output the predicted temperature values ​​at multiple standard altitude points. The discriminator A is a two-stream network structure, including a first conditional stream, a first data stream, a first concatenation layer, two second fully connected hidden layers, and a second output layer. The first conditional stream includes a second fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The first data stream includes a third fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The first concatenation layer concatenates the feature vectors output from the first conditional stream and the first data stream into a fused feature vector. The two second fully connected hidden layers are connected sequentially, with each second fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The second output layer includes a fourth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

[0062] Specifically, the generator A has the following structure: the first input layer receives a concatenated vector of a 100-dimensional noise vector and a 3-dimensional normalized operating condition parameter vector; it then passes through a first fully connected hidden layer of 256 nodes, a 512-node layer, and a 256-node layer, with each first fully connected hidden layer followed by a batch normalization layer and a LeakyReLU activation function; the first output layer is a 36-node fully connected layer using the Tanh activation function. The discriminator A has a dual-branch structure: a second fully connected layer of 64 nodes for processing the 3-dimensional operating condition vector; a third fully connected layer of 64 nodes for processing the 36-dimensional temperature curve; the 64-dimensional feature vectors output from the two branches are concatenated into a 128-dimensional vector, which then passes through a fourth fully connected layer of 256 nodes and a 128-node layer, with each fourth fully connected layer followed by a batch normalization layer and a LeakyReLU activation function; finally, a discriminant probability value is output through a Sigmoid activation function.

[0063] In one example of the present invention, The weighted combination loss function L used by generator A total From the counter-loss L adv and physical constraint loss L smooth Weighted according to the following formula: Among them, the resistance loss L adv The calculation formula is: Among them, physical constraint loss L smooth The calculation formula is: In the formula, λ is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data (C) represents the data distribution under actual operating conditions; D(.) represents the output probability of discriminator A; G(.) represents the output of generator A; N represents the length of the temperature sequence; T i This represents the predicted temperature value at the i-th elevation point; The loss function used by discriminator A in the fire temperature prediction model The expression is: In the formula, Discriminator A represents the probability of outputting "true" under given operating conditions.

[0064] In one example of the present invention, the first report generation unit includes: The abnormal data cleaning subunit is configured with a correction method based on statistical distribution and neighboring point interpolation. For example, it judges the maximum temperature at each altitude point. If it exceeds the second highest temperature by more than 25%, it performs interpolation correction by averaging the temperatures of the two nearest altitude points of the curve. It identifies and corrects abnormal temperature points on the distribution curves generated under all working conditions. For each standard altitude point, it statistically analyzes the distribution of temperature values ​​at that point under all working conditions, sets an anomaly detection threshold based on statistical quantiles, and identifies temperature points exceeding the threshold as abnormal values. It then uses the temperature values ​​of the distribution curve at adjacent altitude points to calculate correction values ​​through linear interpolation and replaces the abnormal values. The envelope extraction subunit is configured to extract the outer envelope of all curves along the height direction based on the cleaned temperature distribution curve dataset. The envelope physical correction subunit is configured to apply a monotonic physical constraint to the extracted preliminary outer envelope to ensure that it satisfies the law that the temperature decays with increasing height.

[0065] In one example of the present invention, a monotonic physical constraint is applied to the extracted preliminary outer envelope to ensure that it satisfies the law of temperature decay with increasing altitude, specifically including: Starting from the bridge deck and proceeding upwards, check each point one by one. If the temperature value of the current elevation point is lower than the temperature value of the subsequent elevation points, then correct the temperature value of the abnormal point to the maximum value in the temperature sequence of the subsequent elevation points, thereby ensuring the physical rationality and conservatism of the final fortification curve.

[0066] In one example of the present invention, the generated report is a Word document, the contents of which include: project summary, list of input operating parameters, temperature protection curve, key safety indicator data, risk level assessment, and specific graded protection recommendations.

[0067] In one example of the present invention, the deployment environment of the fire protection method is a local server cluster under the internal network of the organization. All model parameters, calculation data and knowledge base content are stored in local storage devices to ensure offline processing and privacy security throughout the data processing chain. The large language model is an open-source generative pre-trained model that has been fine-tuned with text and code instructions in the field of bridge engineering, and has professional bridge fire protection terminology understanding and logical reasoning capabilities.

[0068] In one example of the present invention, the safety assessment logic for generating the report is as follows: the temperature protection curve is compared with the material critical temperature threshold, and the fire risk level is automatically determined based on the protection code clauses stored in the local knowledge base, and a graded protection recommendation is given.

[0069] In one example of the present invention, such as Figure 6 As shown, the emergency assessment module includes: The second parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized second working condition parameter dataset. The second model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the standardized second working condition parameter dataset, and generate a deep learning model calling instruction for predicting the temperature field and flame morphology of bridge cables. The second deep learning unit is configured to load and train the corresponding temperature field and flame pattern prediction model based on the deep learning model call instruction, perform spatiotemporal evolution simulation calculation of temperature field and flame pattern, and output the prediction results of temperature field and flame pattern of bridge cable. The second report generation unit is configured to map the predicted results of the bridge cable temperature field and flame pattern onto the bridge's three-dimensional model, and perform dynamic risk analysis and visualization; it integrates key operating parameters, bridge cable temperature field distribution map, flame pattern data, risk level assessment and emergency measure recommendations to generate a structured emergency assessment report.

[0070] Specifically, the emergency assessment module is deployed on a local server cluster within an internal network. All model parameters, computational data, and knowledge base content are stored on local storage devices, ensuring offline processing and privacy security throughout the entire data processing chain. The large language model is an open-source generative pre-trained model fine-tuned with text and code instructions from the field of bridge engineering, possessing professional bridge fire prevention terminology understanding and logical reasoning capabilities.

[0071] The parameters that need to be recognized in a large language model include: 1) Wind speed. The language involved may include: "wind level X", "wind level X to X", "wind speed is 5 m / s", "wind speed is 2 to 3 m / s", and may include vague words such as "around" or "approximately", and may include terms such as "sample size" or "at what interval to take a number".

[0072] Table 1. Wind speed related terms and their corresponding wind speed levels As shown in Table 1, when the corresponding wind speed description is input, the large language model will automatically match the corresponding wind level, for example, "wind speed about 5 m / s" → [4, 6], "a little windy" → [1.6-3.3]; when the corresponding wind level description is input, the large language model will automatically match the corresponding wind speed, for example, "around level 6 wind" → [10.8-20%×10.8, 13.8+20%×13.8], "level 7 to 8 wind" → [13.9, 20.7], so as to accurately determine the wind force.

[0073] 2) Location of the fire source / vehicle.

[0074] a. Lateral position, such as the lane where the fire occurred, is used to determine the distance between the burning vehicle and bridge components, serving as input parameters for the machine learning model. Specific positional parameters can be set according to the actual dimensions of the bridge. Language that may be involved includes: "Vehicle distance from component / cable / main cable / cable clamp / air clamp / guardrail / ... (bridge component related terms) is in meters / centimeter," "Vehicle is located in the rightmost lane / middle lane / leftmost lane / first lane / second lane / emergency lane / ... (Lane location information needs to be combined with the lane distribution map of the Shenzhen-Zhongshan Bridge)," and may include vague terms such as "approximately," "about," or "possibly a few to a few meters."

[0075] b. Longitudinal position, i.e., where the vehicle is located on the bridge. This may involve phrases such as "the vehicle is located in section XX" or "the vehicle is located in the middle / side span / ..." of the Shenzhen-Zhongshan Bridge. The height information of the corresponding components on the bridge is obtained through the longitudinal position and used as input parameters for the neural network model.

[0076] 3) Fire intensity, mainly used to infer flame intensity based on vehicle information. The language used may include: "fire source intensity / fire intensity / vehicle fire intensity / heat release rate / ...", "car / truck / truck loaded with XX goods...", and may include terms such as "sample quantity" or "at what interval to take a number", and may include vague terms such as "around" or "approximately".

[0077] If the detailed information of the vehicle cannot be determined, you can refer to the range in Table 2.

[0078] Table 2. Vehicle Heat Release Power Table If detailed vehicle information can be determined, enter the corresponding data.

[0079] Furthermore, in this embodiment, the emergency assessment module also includes a user feedback learning mechanism to optimize subsequent model predictions and report generation effects based on user feedback on the assessment report; and the intelligent assessment system supports historical tracing and comparative analysis of assessment results.

[0080] This emergency assessment module achieves full automation and intelligence in the assessment process, resulting in an order-of-magnitude improvement in emergency response efficiency. By seamlessly integrating multiple stages such as human-computer interaction, intelligent scheduling of large language models, prediction by professional deep learning models, 3D visualization mapping, and report generation, a complete automated intelligent pipeline has been constructed. This completely changes the traditional fragmented and time-consuming model that relies on manual setting of working conditions, driving simulations, and manual interpretation of results. The assessment cycle, which originally took hours or even days, has been shortened to minutes, gaining valuable "golden time" for fire emergency decision-making and solving the core timeliness bottleneck problem in the background technology.

[0081] This emergency assessment module ensures the complete localization of sensitive data and core models, fundamentally eliminating the risk of data leakage. All processing, calculation, and storage of sensitive data, such as bridge structural parameters and geographic information, are completed on a local server cluster within the organization's internal network, completely eliminating the inherent risks of data export and leakage associated with cloud services. This feature enables this application to be used in emergency assessment scenarios for critical infrastructure projects with extremely high information security requirements, providing a secure and reliable underlying guarantee for the deep application of intelligent technologies in key areas.

[0082] This emergency assessment module provides high-fidelity prediction results that deeply integrate physical laws and data-driven approaches, ensuring the scientific rigor and reliability of the assessment conclusions. The deep learning model in this application is not a simple "black box" data fitting; it incorporates physical constraints during training to ensure that the predicted temperature field and flame morphology conform to fundamental physical principles. During the assessment process, uncertainties are covered through multi-condition sampling, and envelope extraction and physical correction strategies are employed, resulting in a physically reasonable cable temperature curve. Its scientific validity and reliability in supporting emergency decision-making far exceed those of traditional methods relying on empirical formulas or single-condition simulations.

[0083] This emergency assessment module features intuitive and easy-to-understand 3D dynamic visualization and interactive capabilities, greatly improving situational awareness and decision-making efficiency in emergency command. By mapping abstract numerical prediction results onto a 3D model of the bridge in real time, and presenting them visually in the form of heat maps, dynamic flames, and color-coded risk areas, complex professional data is transformed into a clear battlefield situation map. It supports interactive operations such as free switching of perspectives and real-time parameter adjustment, enabling commanders without a professional background to quickly understand the impact of fires, greatly reducing the difficulty of decision-making and improving the accuracy of emergency response.

[0084] This emergency assessment module employs a modular and standardized system architecture, endowing the system with excellent maintainability and scalability. Each core module is encapsulated using containerization technology and communicates through standard API interfaces, achieving a highly cohesive and loosely coupled architectural design. This design allows any component in the system, such as an updated prediction model or a new version of the design specification knowledge base, to be independently upgraded, expanded, or replaced without refactoring the entire system. This ensures the continuous evolution of this application, adapting to future technological developments and standard changes, and possessing a long lifecycle and application value.

[0085] As a further improvement of the present invention, the specific process of the multi-round human-computer interaction includes: using the semantic understanding capability of the large language model to identify fire scene parameter entities implicit or explicit in the user input; for key condition parameters that are not clearly defined, initiating follow-up questioning logic; for condition parameters that are ambiguous or exceed a reasonable range, initiating confirmation and correction logic, and finally outputting a standardized second working condition parameter set.

[0086] As a further improvement of the present invention, the standardized second operating condition parameters include fire source power, fire source distance, and bridge surface wind speed; wherein, the fire source power is given a combustion ratio range based on the vehicle type and the fuel type of the goods transported by the vehicle, and a fire source power range is obtained by linear interpolation, and samples are uniformly distributed within the range; the fire source distance is a fixed value; the bridge surface wind speed is sampled uniformly distributed within the range of the identification parameters; the three operating condition parameters of fire source power, fire source distance, and bridge surface wind speed are arranged and combined, and a hybrid sampling strategy is used to achieve comprehensive coverage of possible operating conditions of fire scenarios.

[0087] In this embodiment, the input parameters are required to include: (1) fire intensity, wind speed, and distance range; (2) generate samples based on the input parameters, with the number of samples being the product of the number of each parameter; (3) the location of the ignition point, such as the length direction and width direction, and can automatically obtain the distribution of components around the ignition point, such as what components are there and how high they are, and can correspond to the modeling of the bridge.

[0088] As a further improvement to the present invention, such as Figure 7 As shown, the second deep learning unit includes: The second fire simulation dataset subunit is configured to generate flame morphology data under different standardized working condition parameter combinations through simulation software; wherein, the working condition parameters include fire source power and bridge surface wind speed; The second data preprocessing and standardization subunit is configured to map the non-uniform spatial temperature data output by the simulation model to a preset regular grid (e.g., a 51×51 regular grid) using cubic spline interpolation, normalize the temperature data to form a set of operating parameters, and divide the set of operating parameters into a training set, a validation set, and a test set. The second adversarial network model subunit is configured to construct a second conditional generative adversarial network model. The second conditional generative adversarial network model adopts a generator-discriminator dual-path system, including a generator B and a structure discriminator B. The generator B is configured to generate flame shape prediction values ​​based on noise vectors and normalized chemical condition parameters. The discriminator B is configured to distinguish between generated flame shapes and real flame shapes and output a discrimination probability value. The second definition is a weighted loss function subunit, configured so that generator B adopts a weighted combined loss function that includes adversarial loss and physical constraints. The second conditional generative adversarial network model is trained through the training set, wherein the adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the flame shape reasonably satisfies the empirical formula. The second model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It uses an alternating iterative update strategy of discriminator B and generator B to train the second conditional generative adversarial network model through the training set. During the training process, the training effect is monitored through the validation set and the training parameters are adjusted. The second model validation and deployment subunit is configured to evaluate the performance of the second conditional generative adversarial network model on a test set, using the coefficient of determination R. 2 The mean absolute error (MAE) is used to evaluate the second-conditional generative adversarial network model, where the coefficient of determination R0 is... 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained second-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

[0089] As a further improvement of the present invention, the generation-discrimination dual-path system of the flame morphology prediction model includes a multi-flow generator B and a dual-flow structure discriminator B; In the multi-stream generator B structure, the conditional coding stream is a double fully connected layer used to receive normalized working condition parameters and extract working condition features; the noise coding stream is a double fully connected layer used to receive random noise vectors and extract noise features; the feature fusion layer concatenates the conditional features and noise features, and passes them through two fully connected layers in sequence. Each fully connected layer is followed by a batch normalization layer and a LeakyReLU activation function, and finally outputs a temperature field vector, which is mapped to the range [0,1] through a Sigmoid activation function. In the dual-stream discriminator B, the conditional coding stream is a double fully connected layer used to process the operating condition vector, and the data stream is a double fully connected layer used to process the temperature field vector. The feature vectors output by the dual-stream network structure are concatenated and then passed through two fully connected layers. Each fully connected layer is followed by a LeakyReLU activation function and a Dropout layer. Finally, a discriminant probability value is output through a Sigmoid activation function.

[0090] The generator B comprises a second input layer, three third fully connected hidden layers, and a third output layer. The second input layer is a concatenated vector of a received noise vector and normalized working condition parameters, used for data input and feature initialization. The three third fully connected hidden layers are connected sequentially, with each third fully connected hidden layer followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively. The third output layer comprises a fifth fully connected layer and a Tanh activation function, used to output temperature prediction values ​​for multiple standard altitude points. The discriminator B is a two-stream network structure, including a second conditional stream, a second data stream, a second concatenation layer, two fourth fully connected hidden layers, and a fourth output layer. The second conditional stream includes a sixth fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The second data stream includes a seventh fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The second concatenation layer concatenates the feature vectors output by the second conditional stream and the second data stream into a fused feature vector. The two fourth fully connected hidden layers are connected sequentially, with each fourth fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The fourth output layer includes an eighth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

[0091] For example, in the generator B structure, the conditional encoding stream is a double fully connected layer with 64 and 128 nodes, used to receive normalized fire source power and bridge deck wind speed operating parameters and extract operating features; the second data stream, i.e., the noise encoding stream, is a double fully connected layer with 256 and 512 nodes, used to receive a 100-dimensional random noise vector and extract noise features; the second concatenation layer concatenates the conditional features and noise features, passing them sequentially through two fully connected layers with 1024 and 2048 nodes. Each fully connected layer is followed by a batch normalization layer and a LeakyReLU activation function, ultimately outputting a 2601-dimensional temperature field vector, which is then mapped to the [0,1] range using a Sigmoid activation function. In this embodiment, the operating parameter features (128-dimensional) and noise features (512-dimensional) are concatenated, gradually amplified to a 51×51=2601-dimensional output through three fully connected layers, and finally compressed to 0-1 using a Sigmoid function.

[0092] In the two-stream discriminator B, the conditional encoding stream is a dual fully connected layer with 64 and 128 nodes, used to process the operating condition vector, ensuring that discriminator B also knows the current operating conditions. The data stream is a fully connected layer with 512 and 256 nodes, used to process the temperature field vector. Two fully connected layers compress the features, flattening the two-dimensional temperature field into a one-dimensional vector. The 128-dimensional feature vector output from the two-stream network structure is concatenated to 256 dimensions, and then passed sequentially through the 512-node and 256-node fully connected layers. Each fully connected layer is followed by a LeakyReLU activation function and a Dropout layer, finally outputting a discriminant probability value through a Sigmoid activation function. The expression for the Sigmoid activation function is: , where x is the output value of the fully connected layer.

[0093] By splicing the operating condition features (128 dimensions) and temperature field features (256 dimensions), a value of 0-1 (1=true, 0=false) is finally output, with Dropout added in the middle to prevent overfitting.

[0094] As a further improvement to the present invention, the weighted combination loss function Adversarial loss from generator B and physical constraints Together they constitute, specifically as follows: in, The adversarial loss of generator B, whose goal is to deceive discriminator B, is calculated as follows: in, This is the physical constraint loss, used to ensure the flame shape conforms to the empirical formula, and its calculation formula is as follows: In the formula, λ1 is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data The distribution of real data; D flame (·) represents the output probability of discriminator B; G flame (·) represents the output of generator B; H pred H represents the predicted flame height. empirical The results are calculated using empirical formulas based on the classic fire plume model; P represents the fire source power. The mass combustion rate.

[0095] As a further improvement of this invention, during model training, the training parameters are adjusted by monitoring the training effect through a validation set; wherein the hyperparameter learning rate is 2×10⁻⁶. -4The minimum number of training rounds is 500, and the maximum is 1000.

[0096] To obtain the most realistic flame image data, this embodiment utilizes the "Real-time Mixed Test Platform and Implementation Method for Vehicle-Fire-Wind Force of Substructure of Long-Span Bridge" disclosed in Chinese Patent Application 2025111956407 to conduct a full-scale or large-scale model fire resistance test of the cable under controllable parameter wind, fire, and force coupling conditions, and directly collects flame morphology data.

[0097] As a further improvement to the present invention, such as Figure 7 As shown, the second deep learning unit also includes: The third fire simulation dataset sub-unit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The third data preprocessing and standardization subunit is configured to uniformly map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the working parameters and temperature values ​​to form a working parameter set, which is then divided into a training set, a validation set, and a test set. The third adversarial network model subunit is configured to construct a third conditional generative adversarial network model, which consists of a generator C and a discriminator C. The generator C is configured to generate predicted temperature curve values ​​based on noise vectors and normalized working condition parameters. The discriminator C is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The generator C comprises a third input layer, three fifth fully connected hidden layers, and a fifth output layer. The third input layer is a concatenated vector of a noise vector and normalized working condition parameters, used for data input and feature initialization. The three fifth fully connected hidden layers are connected sequentially, with each fifth fully connected hidden layer followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively. The fifth output layer comprises a ninth fully connected layer and a Tanh activation function, used to output temperature prediction values ​​for multiple standard altitude points. The discriminator C is a two-stream network structure, including a third conditional stream, a third data stream, a third concatenation layer, two sixth fully connected hidden layers, and a sixth output layer. The third conditional stream includes a tenth fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The third data stream includes an eleventh fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The third concatenation layer concatenates the feature vectors output from the third conditional stream and the third data stream into a fused feature vector. The two sixth fully connected hidden layers are connected sequentially, with each sixth fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The sixth output layer includes a twelfth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

[0098] Specifically, the generator C structure takes a concatenated vector of a 100-dimensional noise vector and 3-dimensional normalized operating condition parameters as input. This vector passes through five fully connected hidden layers with 256, 512, and 256 nodes, each followed by batch normalization and LeakyReLU. The output layer is a 36-node fully connected layer with Tanh activation, outputting temperature predictions at 36 standard altitude points. The discriminator C structure employs a dual-branch architecture. The third conditional stream uses a 64-node fully connected layer to process the 3-dimensional operating condition vector, and the third data stream uses a 64-node fully connected layer to process the 36-dimensional temperature curve. The outputs from both branches are concatenated into a 128-dimensional vector, which then passes through twelfth fully connected layers with 256 and 128 nodes (each followed by batch normalization and LeakyReLU), finally outputting the discrimination probability value via a Sigmoid function.

[0099] The third definition is a weighted loss function subunit, configured so that the generator C adopts a weighted combined loss function that includes adversarial loss and physical constraints. The third conditional generative adversarial network model is trained through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. Among them, the weighted combination loss function L used by generator C total From the counter-loss L adv and physical constraint loss L smooth Weighted according to the following formula: Among them, the resistance loss L adv The calculation formula is: Among them, physical constraint loss L smooth The calculation formula is: In the formula, λ is the hyperparameter; E represents the mathematical expectation; PZ (Z) represents the probability distribution of random noise; P data (C) represents the data distribution under actual operating conditions; D(.) represents the output probability of discriminator C; G(.) represents the output of generator C; N represents the length of the temperature sequence; T i This represents the predicted temperature value at the i-th elevation point; The expression for the loss function used by the discriminator C in the fire temperature prediction model is as follows: In the formula, C represents the probability of outputting "true" under given operating conditions.

[0100] The third model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between the discriminator C and the generator C to train the third conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning rate is 2×10⁻⁶. -4 The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The third model validation and deployment subunit is configured to evaluate the performance of the third conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The third-condition generative adversarial network model is evaluated using the mean absolute error (MAE) and the coefficient of determination (R²). 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained third-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

[0101] As can be seen, the temperature field prediction model and the flame morphology prediction model use the same training method, only the input and output are different, which will not be elaborated here.

[0102] In this embodiment, the construction of the bridge vehicle fire simulation model in the second deep learning unit specifically includes: The two-dimensional cable plane is simplified into a one-dimensional vertical bar, and any position of the cable can be represented by the height position of the vertical bar, thus simplifying the two-dimensional spatial coordinates into a one-dimensional position. The vehicle is simplified into a cube with a fire source on its surface. Recommended maximum grid sizes are given for different vehicle types: 20cm for a maximum heat release rate of 0–5MW; 25cm for a maximum heat release rate of 5–10MW; 30cm for a maximum heat release rate of 10–15MW; 35cm for a maximum heat release rate of 15–20MW; and 50cm for a maximum heat release rate of 20–200MW. The maximum heat release rate of the simulated flame is 8–12 min to equivalently replace the highest temperature of the simulated actual flame over 1.5 h; Temperature measuring points are set at equal intervals along the height direction on the surface of the vertical rod according to the required accuracy.

[0103] In this embodiment, Conditional Generative Adversarial Network (cGAN) technology is used to achieve end-to-end generation from engineering condition parameters to a probability map of flame temperature distribution. Designed for bridge fire protection scenarios, the system can generate a corresponding 51×51 grid of two-dimensional time-averaged temperature distribution probability cloud map based on two key input parameters: heat release rate and wind speed (without requiring vehicle location). Since the flame simulated by FDS differs significantly from the actual flame, this embodiment uses spatial temperature field data to represent flame characteristics, ensuring not only realistic generation results but also values ​​closer to the actual temperature distribution of a fire scene.

[0104] As a further improvement of the present invention, the second report generation unit includes: The spatial mapping subunit is configured to establish a spatial mapping relationship between the bridge's 3D model and the predicted data, and to render the temperature field on the surface of the bridge cables using a heat map. The flame model sub-unit is configured to generate a dynamic flame model based on flame morphology characteristics and use a particle system to simulate the flame morphology change process. The spatial relationship calculation subunit is configured to calculate the spatial relationship between the flame and the bridge cable in real time, and to perform licking detection and risk area identification. The risk level sub-unit is configured to display the risk level through color coding; where red represents high risk, yellow represents medium risk, and green represents low risk.

[0105] As a further improvement of the present invention, the bridge 3D model supports multi-dimensional interactive operations, including: Freely switchable perspectives: Supports global overview, close-up views, and multi-angle observation; Real-time parameter adjustment: Allows users to modify operating parameters and view changes in evaluation results instantly; View details of risk areas: Click on a risk area to view detailed temperature data and safety assessment information.

[0106] In one example of the present invention, dynamic risk analysis specifically includes: comparing the predicted temperature field with the safety threshold of the bridge cable material to assess the impact of thermal load; calculating the probability of flame licking the cable based on the flame morphology and performing risk classification; The thermal load impact analysis includes: overlaying the predicted temperature field distribution with the spatial location of the main cable, extracting the temperature values ​​at various points near the main cable, and assessing whether they exceed the critical temperature threshold in the material database. The flame licking detection process includes: calculating the minimum distance between the flame envelope and the main cable spatial model based on the generated flame morphology characteristics; determining the licking probability based on a distance threshold; and classifying the risk by referring to safety specifications in the local knowledge base. Specifically, after obtaining the flame tilt angle, length, and fire source location, the minimum distance between the fire source and the cable, as well as the maximum distance between the flame and the fire source, can be calculated. The minimum distance can be directly obtained by using concentric circles representing these two distances.

[0107] As a further improvement of the present invention, the structured emergency assessment report includes the following: Scene summary: Basic description of the fire; Input parameter list: All input operating parameters and sampling information; 3D visualization results: Screenshots of a 3D model including cable temperature field distribution, flame morphology, and risk areas; Safety assessment conclusions: probability of flame licking the cable and risk classification based on professional assessment using a knowledge base; Emergency response recommendations: Targeted response plans.

[0108] In this embodiment, when the fire situation on a cable-stayed bridge is input into the intelligent assessment system... User input: "A large truck has caught fire at kilometer marker K69+100. The surveillance footage shows it's near the guardrail. The wind is very strong now, around force 7! Please assess the risk immediately!" Large language model output: Received! An emergency assessment is underway. Based on your description and the system knowledge base, I have analyzed the following fire scenario: Event type: Truck fire Fire intensity (estimated): 45 MW to 75 MW Wind speed (estimated): 13.9 m / s to 17.1 m / s Fire source location (estimated): Near K69+100 longitudinally, above the component at a height of h m; distance of fire source from the component to d m. An emergency assessment neural network model has been used for analysis, and the assessment results are as follows: 1. Flame Form: Predicted flame length range: Rice to rice.

[0109] Predicted flame tilt angle range: to .

[0110] 2. 3D model visualization: Dynamic flames have been generated at position K69+100 in the 3D model.

[0111] Flame shape (length, angle) is currently... m to m / to The fire changes in real time within the area to show the most likely fire scenarios.

[0112] Note: The flames have intermittently licked the main cable, as shown in the simulation results.

[0113] Emergency advice: The probability of flame contact with {{component}} is G, the surface steel wire temperature is below 300℃, and no repair is required.

[0114] In this embodiment, by seamlessly integrating multiple processes such as human-computer interaction, intelligent scheduling of large language models, prediction by professional deep learning models, 3D visualization mapping, and report generation, a complete automated intelligent pipeline is constructed. This completely changes the traditional fragmented and time-consuming model that relies on manual setting of operating conditions, driving simulation, and manual interpretation of results. The evaluation cycle, which originally required hours or even days, is shortened to minutes, gaining valuable "golden time" for fire emergency decision-making.

[0115] In one example of the present invention, such as Figure 8 As shown, the post-disaster assessment module includes: The third parameter extraction unit is configured to identify key working condition parameters and main cable protection type in bridge cable vehicle fire scenarios based on a large language model, and obtain a standardized third working condition parameter dataset including fire source power, fire source distance and bridge surface wind speed. The third model scheduling and management unit is configured to call the local knowledge base for parameter verification based on the standardized third working condition parameter dataset, and generate calling instructions for the cable surface temperature field prediction model and the main cable internal temperature inversion model. The third deep learning unit is configured to use the calling command based on the cable surface temperature field prediction model and the main cable internal temperature inversion model to load the bridge main cable internal temperature inversion model and the bridge cable surface temperature field prediction model and train them; according to the verified working condition parameters, the bridge main cable internal temperature inversion model is used to output the predicted value of the time-temperature inside the main cable, and the cable surface temperature field prediction model is used to output the predicted value of the highest temperature on the surface of the main cable. The third report generation unit is configured to compare the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the safety threshold of the main cable material to classify the damage; and to integrate the standardized third working condition parameter dataset, the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the damage level to generate a report containing parameter vectors, the time-temperature curve inside the main cable, the maximum temperature on the main cable surface, the damage level, and repair recommendations.

[0116] Specifically, the bridge vehicle fire post-fire assessment module can be deployed in a local server cluster within an internal network. All unit parameters, calculation data, and knowledge base content are stored on local storage devices, ensuring end-to-end offline data processing and privacy security. Each unit is encapsulated using containerization technology, exchanging data and scheduling processes through application programming interfaces (APIs), achieving high cohesion and low coupling between units. Furthermore, the data transmission component can be wireless or wired. Wireless transmission utilizes LoRa wireless communication technology, suitable for long-distance bridges and environments with heavy vehicle traffic, with a transmission distance of up to 500 meters. Wired transmission uses high-temperature resistant cables for areas with weak wireless signals, ensuring continuous data transmission and avoiding wireless interference. The data transmission component also includes signal boosters, deployed every 500 meters to improve signal strength and reduce data packet loss. In the third parameter extraction unit, the large language model confirms the extraction through multiple rounds of human-computer interaction. Utilizing its semantic understanding capabilities, the model identifies implicit or explicit fire scenario parameter entities in the user input, allowing the user to select the main cable protection type for the bridge based on existing protection types. The large language model can be trained on existing language models to be applicable to bridge fire damage assessment. This means it can be an open-source pre-trained model that has been fine-tuned using text and code instructions from the bridge engineering field, possessing professional bridge fire post-assessment terminology understanding and logical reasoning capabilities. For unclear key condition parameters, follow-up questioning logic is initiated; for ambiguous or unreasonable condition parameters, confirmation and correction logic is initiated, ultimately outputting a complete and error-free set of standardized working condition parameters. For example, keyword extraction can be performed based on the large language model. This extraction method converts language into text, then calibrates the text to extract keywords such as "fire" and "location," determining the main cable protection type and outputting standardized working condition parameters for fire source intensity, bridge deck wind speed, and fire source distance. Standardized operating condition parameters include: fire source intensity, fire source distance, and bridge deck wind speed. Among them, fire source intensity is a range value, which is obtained by sampling the fire source intensity parameter in a uniform distribution within this range; fire source distance is the distance between the fire source location and the cable; bridge deck wind speed is a range value, which is obtained by sampling the bridge deck wind speed in a uniform distribution within this range. The three operating condition parameters are arranged and combined to achieve comprehensive coverage of the possibilities of fire scenarios through this hybrid sampling strategy. In the third model scheduling and management unit, the local knowledge base can be called to verify the working condition parameters. That is, the local knowledge base may include standard data such as bridge professional terminology library, working condition parameter terminology and database, etc. It can verify the working condition parameter dataset based on test data, historical data, etc., remove fuzzy and inaccurate working condition parameters, and generate the first and second dispatch orders applicable to the main cable protection type. The first dispatch order is used to load the bridge main cable internal temperature inversion model. The inversion model outputs the main cable internal time-temperature prediction value based on the working condition parameter dataset. The second dispatch order is used to load the bridge main cable surface temperature prediction model. The prediction model outputs the main cable surface maximum temperature prediction value based on the working condition parameter dataset. Among them, the experimental data can be obtained from the Chinese invention patent "Real-time Mixed Test Platform and Implementation Method of Vehicle-Fire-Wind Force for Substructure of Long-span Bridge" authorized by our research group with the authorization announcement number CN120741035B. That is, to conduct full-scale or large-scale model fire resistance test of cable under controllable parameter wind, fire and force coupling conditions. The repeatable, measurable and high-fidelity fire environment provided by the platform can greatly ensure the diversity and reliability of the original data. This example integrates large language model interaction, knowledge base parameter verification and scheduling, which can transform post-disaster natural language descriptions into structured chemical condition parameter datasets and correct missing / ambiguous parameters, reducing reliance on expert experience and software operation, and improving engineering usability and accessibility. By jointly comparing and fusing the predicted time-temperature values ​​inside the main cable and the predicted maximum surface temperature values ​​of the main cable during the damage analysis stage, the damage assessment can be more closely aligned with the actual heat-transfer-degradation mechanism. In addition, it can automatically generate structured and traceable assessment reports, forming a closed loop from data analysis to decision support: it not only includes parameter ranges, sampling strategies and curve results, but also provides grading criteria, damage probability / confidence levels and targeted repair suggestions, which can be directly used for post-disaster response and subsequent review.

[0117] This post-disaster assessment module inputs working condition parameters into the trained bridge main cable surface temperature prediction model and the bridge main cable internal temperature inversion model, generates an adversarial network model for forward prediction, and can generate the predicted value of the highest surface temperature of the main cable and the predicted value of the internal time-temperature of the main cable in milliseconds. It can quickly obtain stable conclusions under multiple scenarios and multiple uncertain parameters, achieving an order-of-magnitude breakthrough in assessment efficiency. It is suitable for bridge fire post-disaster assessment scenarios and provides a key time window for rapid judgment of bridge safety status and repair decisions. This post-disaster assessment module constructs a main cable heat transfer inversion model based on an LSTM conditional generative adversarial network. It can not only invert the temperature-time change curve inside the main cable, solving the inversion problem that the internal temperature of the main cable cannot be directly measured, but also train it with different operating parameters so that the model can cover unknown operating conditions far beyond the training sample range, rather than the calculation results of a single operating condition. The output predicted values ​​of the maximum surface temperature and internal temperature of the main cable can provide data support and are more universal and safer. The post-disaster assessment module adopts a dual-model calling framework of "internal temperature inversion and surface maximum temperature prediction" to characterize the internal temperature rise process of the main cable and the maximum surface temperature of the cable respectively. It also performs joint comparison and fusion classification in the damage analysis stage to make the damage judgment closer to the real heat-heat transfer-degradation mechanism. This post-disaster assessment module can automatically generate structured and traceable assessment reports, forming a closed loop from data analysis to decision support: it not only includes parameter ranges, sampling strategies and curve results, but also provides grading criteria, damage probability / confidence levels and targeted repair suggestions, which can be directly used for post-disaster response and subsequent review. The post-disaster assessment module provides a new paradigm of "physical mechanism + data-driven" in the bridge main cable surface temperature prediction model; the bridge main cable internal temperature inversion model can ensure that the inverted internal temperature time series change of the main cable conforms to the physical law of "rapid heating-slow climbing-gradual cooling", ensuring a high degree of consistency between the generated sequence and the real data, and achieving high assessment accuracy while greatly improving assessment efficiency. This post-disaster assessment module compares the external and internal temperatures of the main cable with the safety threshold of the main cable material, and compares the cumulative time exceeding the threshold with the threshold time exceeding the threshold. Based on different value ranges, the damage to the main cable is divided into different levels, avoiding the large range of classification caused by using a single parameter for comparison. The damage classification is more clear and provides a reliable reference for the degree of damage to the main cable.

[0118] In one example of the present invention, such as Figure 9 , Figure 10 As shown, the third deep learning unit includes: The first data simulation subunit is configured to build a bridge vehicle fire simulation model, and the simulation model calculates and outputs peak temperature-height data of the main cable surface under different working conditions. The first dataset is divided into sub-units, configured to construct a dataset from the peak temperature-height data output by the simulation model, the experimental data on the surface of the main cable, and the corresponding working parameters. After preprocessing, the dataset is divided into training set, validation set, and test set. The first inversion model training subunit is configured to construct a bridge main cable surface temperature prediction model based on a conditional generative adversarial network. This model has a generator D and a discriminator D. The generator D is configured to generate predicted temperature curves based on noise vectors and normalized working condition parameters. The discriminator D is configured to distinguish between the generated temperature curves and the actual temperature curves and output a discrimination probability value. The bridge main cable surface temperature prediction model using a loss function is trained on the training set. During the training process, the training effect is monitored and the training parameters are adjusted through the validation set. The performance of the bridge main cable surface temperature prediction model is evaluated on the test set to obtain the trained bridge main cable surface temperature prediction model. The first prediction value output subunit is configured to input the operating condition parameters into the trained bridge main cable surface temperature prediction model and output the predicted value of the highest surface temperature of the main cable.

[0119] In one example of the present invention, a bridge vehicle fire simulation model is constructed, and the ambient temperature field on the cable surface is obtained by simulation calculation using the simulation model. The two-dimensional cable plane is simplified into a one-dimensional vertical rod. The arbitrary position of the cable can be represented by the height of the vertical rod. The two-dimensional spatial coordinates are simplified into one-dimensional positions. The vehicle is simplified into a cube with a fire source on its surface. For different vehicle types, the maximum recommended grid size is given: 20cm for a maximum heat release rate of 0-5MW; 25cm for a maximum heat release rate of 5-10MW; 30cm for a maximum heat release rate of 10-15MW; 35cm for a maximum heat release rate of 15-20MW; and 50cm for a maximum heat release rate of 20-200MW. The maximum heat release rate of the simulated flame is used to replace the highest temperature of the simulated actual flame over 1.5 hours for 8-12 minutes; temperature measuring points are set at equal intervals along the height direction on the surface of the vertical rod according to the required accuracy. The operating parameters include fire source intensity, bridge deck wind speed, and fire source distance. The fire source intensity is sampled densely with a uniform distribution in the range of 0-50MW, and sparsely with a uniform distribution in the range of 50-200MW. The bridge deck wind speed is sampled uniformly in the range of 0-15m / s. The fire source distance is the distance between the emergency lane and the adjacent lane from the edge of the main cable. Experimental data on the main cable surface can be obtained by conducting experiments on the main cable under certain operating parameters, thereby obtaining experimental data on the peak temperature-height of the main cable surface. This experimental data, along with the corresponding operating parameters, and the peak temperature-height data output by the simulation model, are combined to construct a dataset. By combining simulation and experimental data, the output results of the bridge main cable surface temperature prediction model are made more accurate, avoiding the problem of model deviation caused by the same type of dataset.

[0120] In one example of the present invention, the generator D includes a fourth input layer, three seventh fully connected hidden layers, and a seventh output layer; the fourth input layer is a concatenated vector of a noise vector and normalized working condition parameters, used for data input and feature initialization; the three seventh fully connected hidden layers are connected sequentially, each followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively; the seventh output layer includes a thirteenth fully connected layer and a Tanh activation function, used to output temperature prediction values ​​for multiple standard altitude points; The discriminator D is a two-stream network structure, including a fourth conditional stream, a fourth data stream, a fourth concatenation layer, two eighth fully connected hidden layers, and an eighth output layer. The fourth conditional stream includes a fourteenth fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The fourth data stream includes a fifteenth fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The fourth concatenation layer concatenates the feature vectors output from the fourth conditional stream and the fourth data stream into a fused feature vector. The two eighth fully connected hidden layers are connected sequentially, with each eighth fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The eighth output layer includes a sixteenth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

[0121] In one example of the present invention, the generator D employs a weighted combination loss function L. total From the counter-loss L adv and physical constraint loss L smooth Weighted according to the following formula: Among them, the resistance loss L adv The calculation formula is: Among them, physical constraint loss L smooth The calculation formula is: In the formula, z is the noise vector; c is the conditional scalar; λ is the hyperparameter; E represents the mathematical expectation in the prediction model; P Z( Z) is the probability distribution of random noise; P data (C) represents the data distribution under actual working conditions; D(.) represents the output probability of the discriminator D; G(.) represents the output of the generator D; N represents the length of the temperature sequence; Ti represents the predicted temperature value at the i-th elevation point; The expression for the loss function used by discriminator D in the cable surface temperature prediction model is as follows: in, The discriminator D represents the probability of outputting "true" under given operating conditions.

[0122] In one example of this invention, a gradient descent-based optimization algorithm and its hyperparameter configuration strategy are employed. The network is trained using an iterative update strategy, sequentially employing a discriminator D and a generator D. The determination coefficient R is then used on a test set. 2 The model was evaluated using mean absolute error (MAE); during training, the training effect was monitored using a validation set to adjust the training parameters, with a hyperparameter learning rate of 2×10⁻⁶. -4 The minimum number of training epochs is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The coefficient of determination R is used on the test set. 2 The model is evaluated using the mean absolute error (MAE) and the coefficient of determination (R²). 2 >0.9, mean absolute error (MAE) <50℃; In this example, a pre-trained conditional generative adversarial network model is used for forward prediction, which can generate a high-precision peak temperature-height curve in milliseconds, improving computational efficiency by several orders of magnitude. By training with different operating parameters, the model can cover unknown operating conditions far beyond the training sample range, rather than the calculation results of a single operating condition, making it more universal and safer. In addition, the large-scale model technology of conditional generative adversarial networks was innovatively applied to the specific engineering scenario of fire protection for bridge cable vehicles. This not only solved the efficiency bottleneck of traditional methods, but more importantly, it provided a new paradigm of "physical mechanism + data-driven", offering a brand-new technical path and solution. In one example of the present invention, such as Figure 9 , Figure 10 As shown, the third deep learning unit includes: The second data simulation subunit is configured to input operating parameters into the bridge vehicle fire simulation model, and output the time-temperature data of the main cable surface by the simulation model; input the time-temperature data of the main cable surface into the main cable heat transfer simulation model, and output the time-temperature data of the main cable interior. The second dataset is divided into sub-units, configured to construct a dataset from the output time-temperature data inside the main cable, the experimental data inside the main cable, and the corresponding operating parameters. After preprocessing, the dataset is divided into training, validation, and test sets. In other words, the experimental data inside the main cable can be obtained by conducting experiments on the main cable under certain operating parameters. This experimental data, along with the corresponding operating parameters, the surface peak temperature-height data output by the simulation model, and the corresponding operating parameters, are combined to construct a dataset. By combining simulation and experimental data, the output results of the bridge main cable internal temperature inversion model are made more accurate, avoiding the problem of model deviation caused by the same type of dataset.

[0123] The second inversion model training subunit is configured to construct a bridge main cable internal temperature inversion model based on LSTM conditional generative adversarial network. It adopts a weighted combination loss function that includes adversarial loss and physical constraints. The bridge main cable internal temperature inversion model is trained by the training set. During the training process, the training effect is monitored and the training parameters are adjusted by the validation set. The performance of the bridge main cable internal temperature inversion model is evaluated on the test set to obtain the trained bridge main cable internal temperature inversion model. The second prediction output subunit is configured to input operating parameters into the trained bridge main cable internal temperature inversion model and output the predicted time-temperature value inside the main cable.

[0124] This example demonstrates a main cable heat transfer inversion model based on an LSTM-based conditional generative adversarial network. It not only inverts the time-temperature change curve within the main cable, solving the problem of inverting the inability to directly measure the internal temperature, but also achieves a significant breakthrough in assessment efficiency through forward prediction using a pre-trained LSTM-based main cable heat transfer inversion model. This model is suitable for post-fire assessment scenarios of bridges, providing a crucial time window for rapid judgment and repair decisions regarding bridge safety. Furthermore, by using time-temperature data from the main cable surface, it simulates and obtains internal time-temperature data to form a dataset. Only the operating parameters of a vehicle fire need to be input to directly generate the internal temperature field of the main cable, addressing the infeasibility of deploying temperature measurement equipment along the entire main cable and avoiding the need for additional temperature measurement equipment or the difficulty in deploying such equipment.

[0125] In one example of the present invention, the heat transfer simulation model of the main cable is derived using a rectangular finite element model. The rectangular finite element model consists of the main cable steel wires and the air between the steel wires. The internal steel wire dimensions and porosity are consistent with the experimental model. There is surface-to-surface thermal radiation between the steel wires. The radiating surface is defined as a diffuse reflective surface and the emissivity is set. The radiation direction is set to be controlled by the opacity. The opacity of the steel wires and the air is set to opaque and transparent, respectively. The air is set to be non-flowing and the thermal convection of the air is ignored. The main cable structure is regarded as a porous medium of "steel wire-void". By fitting the equivalent thermal conductivity and correcting the equivalent specific heat capacity, a simplified equivalent model of the main cable structure of steel wire-void is established. The input is the time-temperature data of the main cable surface, and the output is the time-temperature data of the main cable interior.

[0126] In one example of the present invention, the equivalent thermal conductivity of the simplified equivalent model is... The calculation formula is derived from the rectangular finite element model and is as follows: In the formula, The average heat flux density is denoted as H; H is the thickness of the main cable protective layer; T1 and T2 are the upper and lower boundary temperatures, respectively. The equivalent thermal conductivity was numerically fitted using a fourth-order polynomial to obtain the equivalent thermal conductivity. Formula for calculating the change with temperature: In the formula, , , , For temperature coefficient, It is a constant; The equivalent specific heat capacity is corrected to 60%-80% of the original specific heat capacity based on the porosity of the model.

[0127] In one example of the present invention, the temperature inversion model inside the main cable of the bridge has a generator E and a discriminator E; The generator E employs an LSTM autoregressive structure, specifically including: The fifth input layer is configured to receive the spliced ​​vector [z;c], where c is a conditional scalar defined by operating parameters and z is a noise vector; The seventeenth fully connected layer is configured to generate the initial hidden state and initial cell state of the LSTM. The ninth output layer is configured to concatenate the predicted temperature from the previous time step with the conditional scalar c as input for each time step. The hidden state and cell state are updated through the LSTM gating mechanism. The output layer uses the Sigmoid activation function to generate normalized temperature values, thus obtaining the complete main cable temperature prediction sequence. The discriminator E adopts a conditional LSTM structure, which is configured to concatenate the temperature value with the conditional scalar c at each time step to form an input sequence. The temporal features are extracted by the LSTM encoder, and the final hidden state outputs the probability of authenticity through a fully connected layer and a sigmoid activation function.

[0128] Specifically, the discriminator E adopts a conditional LSTM structure, which includes the following steps: At each time step, the temperature value is concatenated with the conditional scalar c to form the input sequence. Temporal features are extracted by an LSTM encoder, and the final hidden state outputs the probability of authenticity through a fully connected layer and a Sigmoid activation function.

[0129] Specifically, a training dataset was constructed based on 109 sets of main cable heat transfer simulation results, with a time range of 0 to 90 minutes and a sampling time step of 1 minute. After resampling and interpolation, a main cable temperature sequence y with a length of T=91 was obtained. Each set of data corresponds to a set of operating parameters. The operating parameters are used as the conditional scalar c of the sequence-level condition input. After normalization of each sample sequence, a "condition-sequence" sample pair (c,y) is formed. The generator E uses an LSTM autoregression structure with noise and conditions; for each sample, the conditional scalar c and the noise vector z with dimension d_z=32 are concatenated in the feature dimension to form a joint vector [z;c]; First, the joint vector [z;c] is mapped through a fully connected layer to generate the initial hidden state and the initial cell state: in, , For the LSTM hidden unit dimension, the split() function splits the fully connected layer output into hidden state h0 and cell state c0, W. init This is the initialization of the weight matrix, b init It is a bias term; Generator E generates the main cable temperature prediction sequence stepwise in an autoregressive manner. That is, at each time step The temperature value generated at the previous moment With conditional scalar The current inputs that make up the gating mechanism: The current input x t Feeding into LSTM gating mechanism update ,Right now: Among them, the input gate Controls the extent to which the current input is written into the cell state; forget gate Determines which dimensions the memory from the previous moment is retained; Output gate Determine which information in the cell state is visible to the current hidden state; Here, is the Sigmoid activation function; tanh(·) is the hyperbolic tangent function; ⊙ represents the Hadamard element-wise multiplication; W ix W fxW ox W cx W represents the input weight matrix. ih W fh W oh W ch Let b represent the hidden state weight matrix of the previous time step. i b f b o b c Indicates bias; h t-1 c t-1 The previous hidden state and cell state; h t c t The current state and cell state are hidden. Introducing cell state c at each time step t and input gate Forgotten Gate and output gate Three types of gating mechanisms can automatically "remember" key stages (such as the rapid heating period) related to the temperature rise of the main cable and "forget" disturbances that are less related to the final result during the long-term temperature evolution from 0 to 90 minutes, thereby realizing time-series modeling of the entire process of "rapid heating, slow climbing, and gradual cooling" of the main cable temperature. Then, the normalized temperature prediction value for this moment is given by outputting a fully connected layer and sigmoid activation: in, , This is the transpose of the output weight matrix. As a bias, during training, the value is adjusted based on the global minimum-maximum. Denormalized to the physical temperature range via a self-regressive cycle Next, the main cable temperature prediction sequence was obtained: Generator E adopts an autoregressive generation method that uses the output of the previous time step as the input of the next time step, so that generator E naturally satisfies the time smoothness and continuity constraints in numerical terms, and can better reproduce the characteristics of gradual temperature rise and slow temperature drop of the main cable in the solid heat transfer finite element model. The discriminator E uses an LSTM structure to encode the entire time series, concatenating the predicted temperature at each moment with the conditional scalar c to form the input vector. Repeat the concatenation operation along the time dimension to obtain the input sequence. ; input vector The data is fed into the LSTM encoder sequentially. in, This represents the cell state at the previous moment. The hidden state from the previous time step is considered a high-dimensional feature representation of the entire sequence under a given condition. Then, it is passed through two fully connected layers and a sigmoid activation layer to output the "realism" probability. in, The LeakyReLU activation function is used. It is considered as a high-dimensional feature representation of the entire sequence under a given working condition, and W1 is the weight matrix of the first fully connected layer. b1 and b2 are the transpose of the weight matrix of the second fully connected layer, and b1 and b2 are the biases of each fully connected layer.

[0130] Discriminator E introduces conditional scalars explicitly at the input layer. It can learn the physical correspondence between "operating parameters - temperature rise curve inside the main cable". When the heating amplitude or cooling rate of a certain generated sequence does not match the current operating parameters, the discriminator E will give a lower authenticity score, thereby pushing the generator E to move closer to the real COMSOL distribution in adversarial training.

[0131] Generator E introduces a reconstruction term for the main cable temperature sequence on top of the adversarial loss. The weighted combined loss function of adversarial loss and physical constraints can be written as: The discriminator E loss uses the standard binary cross-entropy form: in, denoted by z, representing the mathematical expectation in the inversion model; z is the noise vector; c is the conditional scalar. This represents the probability function of the discriminator E's output being true; This represents the internal temperature sequence function of the main cable generated by generator E; y represents the hyperparameters in the inversion model; y represents the actual main cable temperature sequence sampled from the real data distribution; Pdata represents the data distribution of the actual main cable temperature sequence, representing the overall characteristics of the temperature data obtained from COMSOL simulation or experiment. During the training phase of the heat transfer inversion model for the main cable, the Adam optimization algorithm can be used, with a learning rate set to 2×10. -4 During the evaluation phase of the main cable heat transfer inversion model, the mean square error (MSE), mean absolute error (MAE), and coefficient of determination were calculated using the inversely normalized temperature sequences of the validation and test sets. The minimum MAE of the validation set is used as the criterion for selecting the optimal model. At the same time, comparison charts of "real curves and generated curves" under several typical working conditions are output to intuitively verify the model's fitting effect on the characteristics of the main cable's temperature rise and cooldown stages.

[0132] In one example of the present invention, the generator E introduces a reconstruction term of the main cable temperature sequence on top of the adversarial loss, and the weighted combined loss function is: The discriminator E loss uses the standard binary cross-entropy form: in, denoted by z, representing the mathematical expectation in the inversion model; z is the noise vector; c is the conditional scalar. This represents the probability function of the discriminator E's output being true; This represents the main cable temperature sequence function generated by generator E; P represents the hyperparameters in the inversion model; y represents the true main cable temperature sequence sampled from the real data distribution; P represents the hyperparameters in the inversion model. data This represents the data distribution of the actual main cable temperature sequence, reflecting the overall characteristics of the temperature data obtained from COMSOL simulation or experimentation.

[0133] In one example of the invention, the external temperature for damage grading is a predicted value of the highest temperature on the main cable surface. The internal temperature is measured using the highest surface temperature of the outer steel wire of the main cable. Measured by the predicted time-temperature value inside the main cable; The safety threshold for main cable materials includes the first surface temperature threshold. Second surface temperature threshold and the temperature threshold of the outer steel wire Set the over-threshold time threshold. Cumulative time exceeding the threshold The cumulative duration during which the surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold. when When, it is determined to be undamaged; when ,and If it is, then it is judged as a minor injury; when ,and If it is, then it is judged as a minor injury; when ,and If the injury is moderate, it is considered a moderate injury; if the injury is severe, it is considered a moderate injury. ,and If so, it is determined to be a severe injury.

[0134] Specifically, when classifying damage, the damage classification criteria are determined by referring to both external and internal temperatures, with the external temperature being the predicted highest temperature on the main cable surface. The measurement is performed using the surface temperature of the outer steel wire of the main cable, measured by the predicted time-temperature value inside the main cable, and using the highest surface temperature of the outer steel wire of the main cable as the measurement. The comparison revealed that the safety threshold for the main cable material includes the first surface temperature threshold. Second surface temperature threshold and the temperature threshold of the outer steel wire ; In surface temperature grading, a first surface temperature threshold is preset. With the second surface temperature threshold This is used to characterize the degree of thermal effect on the surface of the main cable; in the internal temperature grading, the temperature threshold of the outer steel wire is preset. This is used to characterize the initial level of the thermal effect of the outer steel wire; an additional threshold time for exceeding the threshold is set. Cumulative time exceeding the threshold The cumulative duration during which the surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold. The temperature thresholds mentioned above are determined by the material library and / or specification library, and the over-threshold time threshold is determined by the material library and experimental calibration. When the predicted maximum surface temperature of the main cable is not higher than the first surface temperature threshold, that is... If so, then there is no loss; When the predicted maximum surface temperature of the main cable is higher than the first surface temperature threshold but not higher than the second surface temperature threshold, and the maximum surface temperature of the outer steel wire of the main cable is not higher than the outer steel wire temperature threshold, i.e. ,and If so, it is considered a minor injury; When the predicted maximum surface temperature of the main cable is higher than the second surface temperature threshold, and the maximum surface temperature of the outer steel wire of the main cable is not higher than the outer steel wire temperature threshold, i.e. ,and If so, it is considered a minor injury; When the highest surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold, and the cumulative time exceeding the threshold does not exceed the threshold time, i.e. ,and The injury was moderate. When the highest surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold, and the cumulative time exceeding the threshold is greater than the threshold time, i.e. ,and This constitutes severe injury; Specifically, based on literature review, 95℃ is the temperature limit for changes in the physicochemical properties of the outermost wrapping tape or topcoat of the main cable; the temperature limit for irreversible damage to the aerogel felt is set at 600℃ based on literature review; and the temperature limit for no damage to the high-strength steel wire of the main cable is set at 300℃ based on literature review and experimental research, thus establishing the first surface temperature threshold. 95℃, second surface temperature threshold The temperature threshold is 600℃ and the outer steel wire temperature. 300℃; Therefore, if the predicted maximum surface temperature of the main cable is not higher than 95℃, it is considered undamaged. When the predicted maximum surface temperature of the main cable is higher than 95℃ but not higher than 600℃, and the maximum surface temperature of the outer steel wire of the main cable is not higher than 300℃, it is considered minor damage. When the predicted maximum surface temperature of the main cable is higher than 600℃ and the maximum surface temperature of the outer steel wire of the main cable is not higher than 300℃, it is considered minor damage. When the highest surface temperature of the outer steel wire of the main cable exceeds 300℃ and the cumulative time exceeding the threshold does not exceed the threshold time threshold, it is considered moderate damage. When the highest surface temperature of the outer steel wire of the main cable exceeds 300℃ and the cumulative time exceeding the threshold exceeds the threshold time, it is considered severe damage. The threshold time for exceeding the threshold is determined by a steel wire material library and / or by experimental calibration.

[0135] According to a second aspect of the present invention, a processing method for an intelligent analysis system for vehicle fires on cable bridges based on a large model, as described above, includes the following steps: S10: Users input their needs in natural language through the human-computer interaction and parameter extraction module's human-computer interaction interface; S20: The large language model scheduling and management module understands user intent and extracts and verifies standardized working condition parameter sets through multi-turn dialogue; S30: The large language model scheduling and management module schedules the parameter set to the pre-disaster defense module, emergency assessment module, or post-disaster assessment module according to the intent. S40: The invoked pre-disaster preparedness module, emergency assessment module, or post-disaster assessment module performs calculations and analyses, and generates corresponding structured professional reports; S50: The processing system outputs the final structured report to the user.

[0136] This approach enables intelligent and integrated management throughout the entire lifecycle, fundamentally changing the traditional phased and fragmented passive response model. This application organically integrates three previously independent technical aspects—"pre-disaster preparedness," "disaster emergency response," and "post-disaster assessment"—through an intelligent scheduling hub incorporating a large language model, constructing a complete business loop. The system can automatically identify scenario intents based on user needs and invoke corresponding modules, achieving seamless operation throughout the entire process from proactive prevention and rapid response to precise repair, significantly improving the systematic, proactive, and forward-looking nature of bridge cable fire safety management.

[0137] This processing method significantly improves the intelligence level and decision-making efficiency of each stage. In each core module, this application deeply integrates advanced deep learning models trained with physical constraints, replacing traditional methods that rely on empirical formulas or time-consuming high-fidelity numerical simulations. This reduces core computational tasks such as generating temperature protection curves, performing real-time fire simulations, and retrieving internal temperature histories from hours or even days to minutes or even seconds, achieving true "rapid analysis and real-time response." This provides unprecedentedly efficient and intelligent decision support for engineering design, emergency command, and post-disaster assessment.

[0138] This processing method constructs an "evolvable and scalable" intelligent system kernel, ensuring the continuous technological leadership and long-term application value. This application enables the system to self-optimize and grow through iterative learning and module upgrade management modules. On the one hand, the system can continuously fine-tune its internal predictive models and knowledge base based on user feedback, becoming more accurate with use; on the other hand, based on a loosely coupled modular architecture, new algorithms and functions can be easily integrated into the system like "building blocks," achieving smooth upgrades without requiring a complete overhaul. This effectively solves the pain point of traditional software systems easily becoming rigid and obsolete with technological development, greatly extending the technical lifecycle of this system.

[0139] This processing method provides intuitive and accurate visualization and report generation capabilities, significantly lowering the professional threshold and improving communication efficiency. The system not only offers dynamic, three-dimensional visualization of flame and temperature field risks in the emergency module, enabling non-professionals to intuitively understand the risk situation; more importantly, all three core modules can automatically generate well-structured, richly illustrated, and clearly concluded professional reports. This completely changes the inefficient model that previously relied on experts manually interpreting data and writing reports, freeing engineers from tedious paperwork so they can focus on higher-value decisions, while also greatly facilitating technical communication and collaboration between different departments and units.

[0140] This approach, through a localized and modular system architecture, ensures high security, stability, and maintainability while pursuing high performance. All core models and data are deployed locally, eliminating the risk of sensitive engineering data leakage. The high cohesion and low coupling of each functional module not only facilitates expansion but also ensures that the maintenance, updates, and troubleshooting of individual modules do not affect the overall system operation, significantly improving system stability and maintainability and reducing long-term maintenance costs.

[0141] By introducing multimodal information recognition capabilities, this processing method enables the system to directly and objectively extract key parameters (such as flame size, smoke color, and structural burn marks) from on-site images. This effectively compensates for the omissions, subjectivity, or misjudgments that may occur when relying solely on manual descriptions, further improving the accuracy and reliability of the input data and laying a more solid data foundation for subsequent intelligent prediction and evaluation.

[0142] This approach not only addresses the specific pain points of existing technologies, such as slow computation, lack of intuitiveness, fragmented processes, and high barriers to entry, but also constructs an intelligent ecosystem capable of self-learning and continuously adapting to future technological developments by introducing an "evolvable" design concept. This establishes a new and sustainable technological paradigm for the safe operation and maintenance of bridge cables and similar major engineering structures throughout their entire lifecycle, possessing significant engineering application value and long-term economic benefits.

[0143] The foregoing description, with reference to preferred embodiments, details an exemplary implementation of the intelligent assessment system and method for vehicle fires on cable bridges based on a large model proposed in this invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of this invention, and various combinations can be made to the various technical features and structures proposed in this invention, without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A smart fire detection system for cable bridge vehicles based on a large model, characterized in that, include: The system includes modules for human-computer interaction and parameter extraction, large language model scheduling and management, pre-disaster preparedness, emergency assessment, and post-disaster assessment. The human-computer interaction and parameter extraction module is configured to extract and verify user needs and operating condition parameters through natural language interaction and multimodal information recognition. The pre-disaster preparedness module is configured to generate conservative temperature preparedness curves and preparedness reports based on multi-condition sampling and temperature field prediction models. The emergency assessment module is configured to perform rapid temperature field and flame pattern prediction, three-dimensional visualization of risk mapping, and generate an emergency assessment report. The post-disaster assessment module is configured to perform internal temperature inversion of the main cable and prediction of the cable surface temperature field, damage classification, and generate a damage assessment report; and The large language model scheduling and management module is configured to understand user intent, verify parameters, and schedule the pre-disaster defense module, emergency assessment module, and post-disaster assessment module to perform corresponding tasks.

2. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 1, characterized in that, The human-computer interaction and parameter extraction module includes: The multimodal information fusion unit is configured to compare, supplement, and fuse the operating condition parameters described by the user in natural language with the visual parameters automatically identified from field images or videos to generate a more complete and accurate set of standardized operating condition parameters. The parameter conflict resolution unit is configured to initiate a follow-up questioning or confirmation process when there is a contradiction between the natural language description parameters and the visual recognition parameters, guiding the user to clarify and ensuring the reliability of the input parameters.

3. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 1, characterized in that, The disaster preparedness module includes: The first parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized first working condition parameter dataset. The first model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the first working condition parameter dataset, and generate a call instruction for the cable surface temperature field prediction model. The first deep learning unit is configured to load and obtain the corresponding cable surface temperature field prediction model based on the cable surface temperature field prediction model call command, perform temperature field spatiotemporal evolution simulation calculation, and output the cable surface temperature field prediction result. The first report generation unit is configured to integrate the first working condition parameter dataset and the temperature field prediction results to generate a report containing parameter vectors, temperature protection curves for bridge cable vehicle fires, protection indicators, and protection recommendations.

4. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 1, characterized in that, The first deep learning unit includes: The first fire simulation dataset sub-unit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The first data preprocessing and standardization subunit is configured to map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the operating parameters and temperature values ​​to form an operating parameter set, which is then divided into a training set, a validation set, and a test set. The first adversarial network model subunit is configured to generate a first conditional generative adversarial network model consisting of a generator A and a discriminator A. The generator A is configured to generate a predicted temperature curve value based on a noise vector and normalized working condition parameters. The discriminator A is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The first definition is a weighted loss function subunit, configured to use a weighted combined loss function that includes adversarial loss and physical constraints to train the first conditional generative adversarial network model through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. The first model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between discriminator A and generator A to train the first conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning efficiency is 2 × 10⁻⁶. -4 The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The first model validation and deployment subunit is configured to evaluate the performance of the first conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The mean absolute error (MAE) is used to evaluate the first-condition generative adversarial network model, where the coefficient of determination R0 is... 2 >0.9, with a mean absolute error (MAE) <50℃. After successful verification, the weight file of the fully trained first-condition generative adversarial network model is deployed to a local professional model library for use by large language models.

5. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 4, characterized in that, The generator A includes a first input layer, three first fully connected hidden layers, and a first output layer. The first input layer is a concatenated vector of a noise vector and normalized working condition parameters, used for data input and feature initialization. The three first fully connected hidden layers are connected sequentially, with each first fully connected hidden layer followed by a batch normalization layer and a LeakyReLU activation function, used for feature extraction and nonlinear transformation at different levels, respectively. The first output layer includes a first fully connected layer and a Tanh activation function, used to output predicted temperature values ​​at multiple standard altitude points. The discriminator A is a two-stream network structure, including a first conditional stream, a first data stream, a first concatenation layer, two second fully connected hidden layers, and a second output layer. The first conditional stream includes a second fully connected layer and a LeakyReLU activation function, used to process the operating condition vector and extract conditional features. The first data stream includes a third fully connected layer and a LeakyReLU activation function, used to process the temperature curve vector and extract data features. The first concatenation layer concatenates the feature vectors output from the first conditional stream and the first data stream into a fused feature vector. The two second fully connected hidden layers are connected sequentially, with each second fully connected hidden layer followed by a LeakyReLU activation function, used for feature fusion and discriminant analysis, respectively. The second output layer includes a fourth fully connected layer and a Sigmoid activation function, used to output the discriminant probability value.

6. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 4, characterized in that, The weighted combination loss function L used by generator A total From the counter-loss L adv and physical constraint loss L smooth Weighted according to the following formula: Among them, the resistance loss L adv The calculation formula is: Among them, physical constraint loss L smooth The calculation formula is: In the formula, λ is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data (C) represents the data distribution under actual operating conditions; D(.) represents the output probability of discriminator A; G(.) represents the output of generator A; N represents the length of the temperature sequence; T i This represents the predicted temperature value at the i-th elevation point; The loss function used by discriminator A The expression is: In the formula, Discriminator A represents the probability of outputting "true" under given operating conditions.

7. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 1, characterized in that, The emergency assessment module includes: The second parameter extraction unit is configured to extract and confirm key working condition parameters of bridge cable vehicle fire scenarios through multiple rounds of human-computer interaction based on a large language model, and obtain a standardized second working condition parameter dataset. The second model scheduling and management unit is configured to call the local knowledge base to verify parameters based on the standardized second working condition parameter dataset, and generate a deep learning model calling instruction for predicting the temperature field and flame morphology of bridge cables. The second deep learning unit is configured to load and obtain the corresponding cable surface temperature field prediction model and flame morphology prediction model based on the deep learning model call instruction, train them, perform spatiotemporal evolution simulation calculations of temperature field and flame morphology, and output the prediction results of bridge cable temperature field and flame morphology. The second report generation unit is configured to map the predicted results of the bridge cable temperature field and flame pattern onto the bridge's three-dimensional model, and perform dynamic risk analysis and visualization; it integrates key operating parameters, bridge cable temperature field distribution map, flame pattern data, risk level assessment and emergency measure recommendations to generate a structured emergency assessment report.

8. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 7, characterized in that, The second deep learning unit includes: The second fire simulation dataset sub-unit is configured to generate flame morphology data under different standardized working condition parameter combinations through simulation software. The second data preprocessing and standardization subunit is configured to map the non-uniform spatial temperature data output by the simulation model to a preset regular grid through cubic spline interpolation, normalize the temperature data to form a set of operating parameters, and divide the set of operating parameters into a training set, a validation set and a test set. The second adversarial network model subunit is configured to construct a second conditional generative adversarial network model. The second conditional generative adversarial network model adopts a generator-discriminator dual-path system, including a generator B and a structure discriminator B. The generator B is configured to generate flame shape prediction values ​​based on noise vectors and normalized chemical condition parameters. The discriminator B is configured to distinguish between generated flame shapes and real flame shapes and output a discrimination probability value. The second definition is a weighted loss function subunit, configured so that generator B adopts a weighted combined loss function that includes adversarial loss and physical constraints. The second conditional generative adversarial network model is trained through the training set, wherein the adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the flame shape reasonably satisfies the empirical formula. The second model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It uses an alternating iterative update strategy of discriminator B and generator B to train the second conditional generative adversarial network model through the training set. During the training process, the training effect is monitored through the validation set and the training parameters are adjusted. The second model validation and deployment subunit is configured to evaluate the performance of the second conditional generative adversarial network model on a test set, using the coefficient of determination R. 2 The second-condition generative adversarial network model is evaluated using the mean absolute error (MAE) and the coefficient of determination (R²). 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained second-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

9. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 8, characterized in that, Weighted combination loss function Adversarial loss from generator B and physical constraints Together they constitute, specifically as follows: in, The adversarial loss of generator B, whose goal is to deceive discriminator B, is calculated using the following formula: in, This is the physical constraint loss, used to ensure the flame shape conforms to the empirical formula, and its calculation formula is as follows: In the formula, λ1 is the hyperparameter; E represents the mathematical expectation; P Z (Z) represents the probability distribution of random noise; P data The distribution of real data; D flame (·) represents the output probability of discriminator B; G flame (·) represents the output of generator B; H pred H represents the predicted flame height. empirical The results are calculated using empirical formulas based on the classic fire plume model; P represents the fire source power. The mass combustion rate.

10. The intelligent analysis system for vehicle fires on cable bridges based on a large model according to claim 7, characterized in that, The second deep learning unit also includes: The third fire simulation dataset sub-unit is configured to generate bridge height-temperature data under different standardized working condition parameter combinations through simulation software. The third data preprocessing and standardization subunit is configured to uniformly map the non-uniform height-temperature distribution in simulation data and experimental data to a standard height sequence through an interpolation algorithm, and to normalize the working parameters and temperature values ​​to form a working parameter set, which is then divided into a training set, a validation set, and a test set. The third adversarial network model subunit is configured to construct a third conditional generative adversarial network model, which consists of a generator C and a discriminator C. The generator C is configured to generate predicted temperature curve values ​​based on noise vectors and normalized working condition parameters. The discriminator C is configured to distinguish between the generated temperature curve and the real temperature curve and output a discrimination probability value. The third definition is a weighted loss function subunit, configured so that the generator C adopts a weighted combined loss function that includes adversarial loss and physical constraints. The third conditional generative adversarial network model is trained through the training set. The adversarial loss ensures that the generated distribution approximates the real distribution, and the physical constraints ensure that the generated curve is smooth and meets the corresponding physical laws. The third model training and optimization subunit is configured with a gradient descent-based optimization algorithm and its hyperparameter configuration strategy. It employs an alternating iterative update strategy between the discriminator C and the generator C to train the third conditional generative adversarial network model using the training set. During training, the training effect is monitored using a validation set to adjust the training parameters. The hyperparameter learning rate is 2×10⁻⁶. -4 The minimum number of training rounds is 500 to ensure prediction accuracy, and the maximum is 1000 to prevent overfitting. The third model validation and deployment subunit is configured to evaluate the performance of the third conditional generative adversarial network model on the test set, using the coefficient of determination R. 2 The third-condition generative adversarial network model is evaluated using the mean absolute error (MAE) and the coefficient of determination (R²). 2 >0.9, Mean Absolute Error (MAE) <5×10 -3 After successful verification, the fully trained third-condition generative adversarial network model weight file is deployed to a local professional model library for use by large language models.

11. The intelligent analysis system for vehicle fires on cable bridges based on a large model, as described in claim 7, is characterized in that... The second report generation unit includes: The spatial mapping subunit is configured to establish a spatial mapping relationship between the bridge's 3D model and the predicted data, and to render the temperature field on the surface of the bridge cables through a heat map. The flame model sub-unit is configured to generate a dynamic flame model based on flame morphology characteristics and uses a particle system to simulate the flame morphology change process. The spatial relationship calculation subunit is configured to calculate the spatial relationship between flames and bridge cables in real time, and to perform licking detection and risk area identification. The risk level sub-unit is configured to display the risk level through color coding; where red represents high risk, yellow represents medium risk, and green represents low risk.

12. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 1, characterized in that, The post-disaster assessment module includes: The third parameter extraction unit is configured to identify key working condition parameters and main cable protection type in bridge cable vehicle fire scenarios based on a large language model, and obtain a standardized third working condition parameter dataset including fire source power, fire source distance and bridge surface wind speed. The third model scheduling and management unit is configured to call the local knowledge base for parameter verification based on the standardized third working condition parameter dataset, and generate calling instructions for the cable surface temperature field prediction model and the main cable internal temperature inversion model. The third deep learning unit is configured to use the calling command based on the cable surface temperature field prediction model and the main cable internal temperature inversion model to load the bridge main cable internal temperature inversion model and the bridge cable surface temperature field prediction model and train them; according to the verified working condition parameters, the bridge main cable internal temperature inversion model is used to output the predicted value of the time-temperature inside the main cable, and the cable surface temperature field prediction model is used to output the predicted value of the highest temperature on the surface of the main cable. The third report generation unit is configured to compare the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the safety threshold of the main cable material to classify the damage; and to integrate the standardized third working condition parameter dataset, the predicted time-temperature inside the main cable, the predicted maximum temperature on the main cable surface, and the damage level to generate a report containing parameter vectors, the time-temperature curve inside the main cable, the maximum temperature on the main cable surface, the damage level, and repair recommendations.

13. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 12, characterized in that, The third deep learning unit includes: The first data simulation subunit is configured to build a bridge vehicle fire simulation model, and the simulation model calculates and outputs peak temperature-height data of the main cable surface under different working conditions. The first dataset is divided into sub-units, configured to construct a dataset from the peak temperature-height data output by the simulation model, the experimental data on the surface of the main cable, and the corresponding working parameters. After preprocessing, the dataset is divided into training set, validation set, and test set. The first inversion model training subunit is configured to construct a bridge main cable surface temperature prediction model based on a conditional generative adversarial network. This model has a generator D and a discriminator D. The generator D is configured to generate predicted temperature curves based on noise vectors and normalized working condition parameters. The discriminator D is configured to distinguish between the generated temperature curves and the actual temperature curves and output a discrimination probability value. The bridge main cable surface temperature prediction model using a loss function is trained on the training set. During the training process, the training effect is monitored and the training parameters are adjusted through the validation set. The performance of the bridge main cable surface temperature prediction model is evaluated on the test set to obtain the trained bridge main cable surface temperature prediction model. The first prediction value output subunit is configured to input the operating condition parameters into the trained bridge main cable surface temperature prediction model and output the predicted value of the highest surface temperature of the main cable.

14. The intelligent analysis system for vehicle fires on cable bridges based on a large model, as described in claim 12, is characterized in that... The third deep learning unit includes: The second data simulation subunit is configured to input operating parameters into the bridge vehicle fire simulation model, and output the time-temperature data of the main cable surface by the simulation model; input the time-temperature data of the main cable surface into the main cable heat transfer simulation model, and output the time-temperature data of the main cable interior. The second dataset is divided into sub-units and configured to construct a dataset from the output time-temperature data inside the main cable, the experimental data inside the main cable, and the corresponding operating parameters. After preprocessing, the dataset is divided into training set, validation set, and test set. The second inversion model training subunit is configured to construct a bridge main cable internal temperature inversion model based on LSTM conditional generative adversarial network. It adopts a weighted combination loss function that includes adversarial loss and physical constraints. The bridge main cable internal temperature inversion model is trained by the training set. During the training process, the training effect is monitored and the training parameters are adjusted by the validation set. The performance of the bridge main cable internal temperature inversion model is evaluated on the test set to obtain the trained bridge main cable internal temperature inversion model. The second prediction output subunit is configured to input operating parameters into the trained bridge main cable internal temperature inversion model and output the predicted time-temperature value inside the main cable.

15. The intelligent analysis system for vehicle fires on cable bridges based on a large model as described in claim 14, characterized in that, The heat transfer simulation model of the main cable is derived using a rectangular finite element model. The rectangular finite element model consists of the main cable steel wires and the air gaps between the steel wires. The internal steel wire dimensions and porosity are consistent with the experimental model. There is surface-to-surface thermal radiation between the steel wire gaps. The radiating surface is defined as a diffuse reflective surface and its emissivity is set. The radiation direction is set to be controlled by opacity. The opacity of the steel wires and the air is set to opaque and transparent, respectively. The air is set to be non-flowing, and the thermal convection of the air is ignored. The main cable structure is regarded as a porous medium of "steel wire-void". By fitting the equivalent thermal conductivity and correcting the equivalent specific heat capacity, a simplified equivalent model of the main cable structure of steel wire-void is established. The input is the time-temperature data of the main cable surface, and the output is the time-temperature data of the main cable interior.

16. The intelligent analysis system for vehicle fires on cable bridges based on a large model according to claim 12, characterized in that, The external temperature for damage grading is the predicted value of the highest surface temperature of the main cable. The internal temperature is measured using the highest surface temperature of the outer steel wire of the main cable. Measured by the predicted time-temperature value inside the main cable; The safety threshold for main cable materials includes the first surface temperature threshold. Second surface temperature threshold and the temperature threshold of the outer steel wire ; Set the over-threshold time threshold Cumulative time exceeding the threshold The cumulative duration during which the surface temperature of the outer steel wire of the main cable exceeds the outer steel wire temperature threshold. when When, it is determined to be undamaged; when ,and If it is, then it is judged as a minor injury; when ,and If it is, then it is judged as a minor injury; when ,and If the injury is moderate, it is considered a moderate injury; if the injury is severe, it is considered a moderate injury. ,and If so, it is determined to be a severe injury.

17. A judgment method for a large-model-based intelligent judgment system for vehicle fires on cable bridges, as described in any one of claims 1 to 16, characterized in that, Includes the following steps: S10: Users input their needs in natural language through the human-computer interaction and parameter extraction module's human-computer interaction interface; S20: The large language model scheduling and management module understands user intent and extracts and verifies standardized working condition parameter sets through multi-turn dialogue; S30: The large language model scheduling and management module schedules the parameter set to the pre-disaster defense module, emergency assessment module, or post-disaster assessment module according to the intent. S40: The invoked pre-disaster preparedness module, emergency assessment module, or post-disaster assessment module performs calculations and analyses, and generates corresponding structured professional reports; S50: The processing system outputs the final structured report to the user.

Images