A fire parameter cloud map generation and large model explanation method based on reduced-order sample expansion

Abstract

The application discloses a fire parameter cloud map generation and large model explanation method based on reduced-order sample expansion, and belongs to the technical field of fire simulation and artificial intelligence. In view of the defects that the FDS simulation algorithm cost is high, the data-driven model has poor generalization, and the large model explanation is easy to present "illusion", the application first maps the high-dimensional cloud map to the low-dimensional coefficient space through the proper orthogonal decomposition, then generates the extended working condition in the physical feasible region, and after multidimensional physical screening, the prediction model is trained with the high weight strategy of the basic sample; the structured features such as the risk area and the propagation speed are extracted synchronously, and after the consistency verification closed loop, the field adaptation large model is driven to generate professional explanation. The cloud map generation time of the application is compressed from several hours to seconds, the explanation content and the measured data comprehensive consistency score are stably above 0.8, the algorithm power consumption is significantly reduced, the fire situation interpretation credibility is improved, and efficient support is provided for emergency decision-making.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of fire simulation and artificial intelligence, and in particular to a method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion. Background Technology

[0002] The development of a fire is influenced by a combination of factors, including the location of the fire source, heat release rate, fuel type, spatial geometry, obstacles, openings, ventilation conditions, and environmental boundaries. Fire parameters such as temperature, smoke concentration, hazardous gas concentration, visibility, and heat flux exhibit strong nonlinearity, strong coupling, and rapid evolution characteristics in both space and time. Accurately obtaining these fire parameter fields is crucial for fire risk assessment, personnel evacuation, safety design, emergency response, and fire training.

[0003] Currently, the fire dynamics simulation software FDS can output high-precision fire parameter fields based on fire mechanisms and has been applied in fields such as building fire protection design, tunnel smoke control, underground space fire assessment, warehouse safety analysis, and emergency simulation in industrial scenarios. However, when it is necessary to cover different combinations of fire sources, ventilation, spatial layout, and environmental boundaries, a large number of numerical simulations need to be executed repeatedly; each case contains multiple time steps and multiple parameter channels, resulting in long data acquisition cycles and high computational costs, making it difficult to meet the needs of rapid assessment and real-time simulation.

[0004] Existing data-driven fire prediction methods typically learn high-dimensional parameter fields directly, or build prediction models for fixed scenarios, single parameters, and finite operating conditions. These methods still have some areas for improvement in engineering applications.

[0005] (1) Existing fire dynamics simulation usually requires a large amount of grid calculation and time step iteration. The time cost and computing power cost of repeated simulation under multiple conditions are high, and it is difficult to quickly obtain fire parameter field data covering different fire sources, ventilation and space conditions.

[0006] (2) After an actual fire occurs, the situation on site changes rapidly. If we rely on real-time high-precision simulation to calculate on the spot, it is often difficult to predict the development trend of the fire and the changes in the dangerous area in a timely manner, and it is difficult to meet the needs of rapid auxiliary decision-making.

[0007] (3) Fire parameter cloud map contains a variety of professional parameters. Its risk interpretation usually relies on the experience of professionals. Ordinary people find it difficult to quickly understand the risk level, development trend and key points of handling. There is still room for improvement in interpretation efficiency and standardization.

[0008] Therefore, it is necessary to propose a method that integrates fire numerical simulation, reduced-order modeling, sample expansion, and intelligent interpretation to reduce the cost of multi-condition simulation and analysis, improve the efficiency of fire parameter cloud map generation, and provide numerical basis for fire situation interpretation. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems in the background technology and provide a method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion. Addressing the pain points of high computational cost and difficulty in real-time extrapolation in fire simulation, it constructs a full-chain technical path of "mechanism simulation - reduced-order representation - sample expansion - intelligent interpretation". First, it utilizes intrinsic orthogonal decomposition (POD) to map high-dimensional fire cloud maps to a low-dimensional coefficient space, transforming complex flow field prediction into a low-dimensional coefficient time-series evolution problem, significantly reducing computational complexity. Second, it introduces a physical constraint mechanism to automatically generate and select physically reasonable extended working conditions outside the basic simulation samples, enabling the training of a prediction model with strong generalization ability using only a small amount of high-precision FDS data. Finally, it extracts structured features such as risk areas and propagation speed to constrain the large model generation process and introduces a consistency verification closed loop, solving the problem of frequent "illusions" in general large models in vertical domains, and achieving efficient, reliable, and intelligent interpretation of fire situations.

[0010] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion, characterized by comprising the following steps:

[0011] S1: Determine the target fire scenario, establish an FDS fire simulation model, set fire source parameters, space parameters, ventilation parameters and environmental parameters, and execute multiple basic simulation conditions.

[0012] S2: Preprocess and reduce the order of the basic fire parameter field data to obtain the low-dimensional characterization data corresponding to the basic fire parameter field data.

[0013] S3: Based on the simulated working conditions and low-dimensional characterization data, construct a fire parameter field prediction model, and use the basic fire parameter field data to train, verify and test the fire parameter field prediction model to obtain an initial fire parameter field prediction model.

[0014] S4: Generate extended working conditions within the physically feasible range, generate extended fire parameter fields using the initial fire parameter field prediction model, and perform physical screening on the extended fire parameter fields.

[0015] S5: The basic samples are merged with the expanded samples that have passed physical screening to form an enhanced dataset, and the final fire parameter field prediction model is trained.

[0016] S6: Use the final fire parameter field prediction model to generate fire parameter field data and fire parameter time series cloud map corresponding to the target working condition, and extract structured fire situation features from the fire parameter field data.

[0017] S7: Construct a large text model training sample based on fire parameter values, structured fire situation characteristics, fire protection rules and professional interpretation labels, and use the large text model training sample to perform domain adaptation on the large text model to obtain a domain-adapted large text model.

[0018] S8: Input the fire parameter values ​​corresponding to the working condition to be explained into the domain-adapted text model to generate a professional explanation of the fire situation and perform consistency verification.

[0019] S9: Outputs a time-series cloud map of fire parameters and a professional interpretation of the fire situation that has passed the consistency check.

[0020] Furthermore, step S1 specifically includes the following steps:

[0021] S1.1: Establish an FDS fire simulation model based on the target fire scenario, set relevant simulation parameters of the fire scenario in the FDS fire simulation model, and set the measuring points and data slices. The relevant simulation parameters include geometric structure, material properties, fire source parameters, opening location, obstacle distribution, ventilation conditions and environmental conditions.

[0022] S1.2: Each basic simulation condition is represented as a condition parameter vector consisting of fire source parameters, space parameters, ventilation parameters, and environmental parameters. The basic simulation condition parameter vector is represented as follows:

[0023]

[0024] in, This represents the parameter vector corresponding to the i-th basic simulation condition; This represents the fire source parameter vector corresponding to the i-th basic simulation condition; This represents the spatial parameter vector corresponding to the i-th basic simulation condition; This represents the ventilation parameter vector corresponding to the i-th basic simulation condition; This represents the environmental parameter vector corresponding to the i-th basic simulation condition;

[0025] S1.3: Execute multiple FDS basic simulation conditions according to the aforementioned working condition parameter vector to obtain basic fire parameter field data corresponding to each basic simulation condition.

[0026] Furthermore, step S2 specifically includes the following steps:

[0027] S2.1: Perform time step alignment, unified two-dimensional coordinate mapping, and fixed grid resampling on the basic fire parameter field data to obtain fire parameter field data under a unified spatiotemporal scale; wherein, the basic fire parameter field data includes temperature field, smoke concentration field, harmful gas concentration field, and visibility field.

[0028] S2.2: Establish a geometric mask M(x) based on the solid regions, boundary regions, and uncomputable regions in the target fire scene, and use the geometric mask to mark invalid mesh regions:

[0029]

[0030] Where M(x)=1 indicates that spatial location x is an effective flow field grid, and M(x)=0 indicates that spatial location x is an invalid grid; after deleting non-converged samples, missing samples, and abnormal samples, each fire parameter channel is normalized, as shown in the following formula:

[0031]

[0032] in, This represents the fire parameters of type c at time t and spatial location x for the i-th working condition. and Let represent the mean and standard deviation of the parameter channels for fire type c, respectively. This represents the normalized fire parameter field; the normalization formula is used to unify the dimensions of different parameter channels and ensure that the model learns the fire parameter distribution only on the effective grid.

[0033] S2.3: Expand the normalized multi-channel fire parameter field into a fire parameter field snapshot vector, and calculate the average fire parameter field vector. And construct the mean-free snapshot matrix X:

[0034]

[0035] Singular value decomposition is performed on the snapshot matrix X. The singular value decomposition model is as follows:

[0036]

[0037] in, This represents a snapshot matrix composed of multiple snapshots of fire parameter fields. Let Σ denote the left singular vector matrix, and let Σ denote the singular value diagonal matrix. The transpose of the right singular vector matrix;

[0038] S2.4: Determine the number of retained POD modes based on the singular value cumulative energy contribution rate. The cumulative energy contribution rate calculation model is as follows:

[0039]

[0040] in, This represents the k-th singular value. This represents the total number of optional modes in the snapshot matrix. Indicates the cumulative energy threshold. This indicates the number of modes retained. The first r left singular vectors from the left singular vector matrix are selected to form the reduced-order basis matrix of POD:

[0041]

[0042] in, This represents the reduced-order basis matrix of POD composed of the first r principal spatial modes. This represents the first to the rth columns of the left singular vector matrix U;

[0043] The high-dimensional fire parameter field corresponding to the i-th basic simulation condition at time t is mapped to the low-dimensional space of POD, and the formula for calculating its low-dimensional coefficients is as follows:

[0044]

[0045] Where vec() represents vectorization operation, Represents the average field of the training samples. This represents the low-dimensional representation data of the i-th working condition at time t; the low-dimensional coefficients serve as the prediction objects for the subsequent time-series prediction network.

[0046] Furthermore, step S3 specifically includes the following steps:

[0047] S3.1: Normalize and encode the fire source parameters, spatial parameters, ventilation parameters, and environmental parameters corresponding to each basic simulation condition to obtain the condition encoding vector, and input the condition encoding vector into the multilayer perceptron to obtain the condition embedding vector:

[0048]

[0049] Where norm() represents the normalization encoding operation. This represents the mapping function of a multilayer perceptron. This represents the working condition embedding vector for the i-th working condition;

[0050] S3.2: Using the low-dimensional representation data obtained in step S2 as the model prediction object, the low-dimensional representation data of q consecutive historical time steps and the operating condition embedding vector are jointly input into the LSTM time series prediction network to predict the low-dimensional representation data of the next time step. The time series prediction model is as follows:

[0051]

[0052] in, This indicates that the i-th basic simulation condition is in the next time step. Predicting low-dimensional representation data, This represents the LSTM time series prediction function. This indicates that the i-th basic simulation condition is in continuous A low-dimensional representation data sequence of historical time steps, This represents the condition embedding vector corresponding to the i-th basic simulation condition. Indicates the length of the input time window. Indicates the time interval between adjacent time steps;

[0053] S3.3: Using the main spatial modes obtained in step S2, the predicted low-dimensional characterization data is linearly reconstructed to obtain a linearly reconstructed fire parameter field, and a nonlinear residual correction term is generated using a transposed convolutional neural network. The linear parameter field is calculated as follows.

[0054]

[0055] The predicted low-dimensional coefficients and the embedded vectors of the operating conditions are input into the nonlinear residual recovery network to obtain the nonlinear residual correction term:

[0056]

[0057] And obtain the predicted fire parameter field:

[0058]

[0059] in, This indicates that the fire parameter field is linearly reconstructed using the POD method. The parameter is Transposed convolutional neural network, This represents the nonlinear residual correction term. This represents the final predicted fire parameter field;

[0060] S3.4: Train the fire parameter field prediction model using the basic fire parameter field data in the training set, adjust the model parameters using the validation set and determine the optimal model, and evaluate the prediction accuracy of the optimal model using the test set; the following joint training loss is used during the training process:

[0061]

[0062] in, This represents the prediction loss based on low-dimensional coefficients. Indicates the fire parameter field reconstruction loss. Represents physical constraint loss. , and These represent the weight coefficients of the corresponding loss terms; the physical constraint loss includes geometric mask constraints, parameter value range constraints, temporal continuity constraints, and fire source neighborhood response rationality constraints; when the test results reach the preset accuracy threshold, the optimal model is determined as the initial fire parameter field prediction model.

[0063] Furthermore, step S4 specifically includes the following steps:

[0064] S4.1: Generate extended operating conditions within the physically feasible range of the basic simulation operating conditions. The generation model for the extended operating conditions is as follows:

[0065]

[0066] in, Represents the extended operating condition parameter vector. This represents the parameter vector for the i-th basic simulation condition. The parameter perturbation vector represents the parameters that satisfy the physical feasible range; the physical feasible range refers to the extended working condition that satisfies the following conditions: the fire source is located within the effective fluid area, the opening is located within the scene boundary, the obstacles do not overlap in a non-preset manner, the ventilation conditions do not conflict with each other, and each parameter does not exceed the preset upper and lower limits.

[0067] S4.2: Perform constraint judgment on the extended working condition to ensure that the extended working condition meets the following requirements: the fire source is not located in a solid area, the opening does not exceed the scene boundary, there is no non-preset overlap between obstacles, and there is no conflict between ventilation conditions.

[0068] S4.3: Input the extended working conditions determined by the constraints into the initial fire parameter field prediction model to generate corresponding extended fire parameter field data. Then, perform physical screening on the extended fire parameter field data based on the parameter value range, geometric mask, continuity of adjacent time steps, and rationality of the response in the fire source neighborhood. Determine the extended fire parameter field data that meets all screening conditions as valid extended fire parameter field samples.

[0069] Furthermore, step S5 specifically includes the following steps:

[0070] S5.1: Merge the base samples with the expanded samples that have passed the physical screening to form an augmented dataset;

[0071] S5.2: Set training weights for the base samples and expanded samples in the augmented dataset, and train the model according to the weighted training objective. The trained model is as follows:

[0072]

[0073] in, Represents the basic sample set, This represents the expanded set of samples that have passed physical screening. Represents the training weights of the basic samples. This represents the training weights for expanded samples, and + =1;

[0074] S5.3: Using the model parameters of the initial fire parameter field prediction model as initialization parameters, the fire parameter field prediction model is retrained or fine-tuned using the augmented dataset;

[0075] S5.4: Test the retrained or fine-tuned fire parameter field prediction model using independent test conditions. When the test results reach the preset accuracy threshold, the final fire parameter field prediction model is obtained.

[0076] Furthermore, step S6 specifically includes the following steps:

[0077] S6.1: Obtain the fire parameter field data corresponding to the target working condition in the initial time period, perform the same preprocessing as in step S2 on the fire parameter field data, and project it onto the POD reduced-order basis matrix to obtain the initial low-dimensional coefficient sequence corresponding to the target working condition; input the working condition parameters, the initial low-dimensional coefficient sequence and the target time series corresponding to the target working condition into the final fire parameter field prediction model, and recursively generate the predicted fire parameter field data corresponding to each target time step;

[0078]

[0079] in, These are the prediction parameters after inverse normalization. These are the normalized fire parameter values ​​predicted by the model. and These represent the mean and standard deviation of the parameter channels for fire type c, respectively.

[0080] S6.2: Generate fire parameter time series cloud maps according to a unified spatial scale, unified time interval, and unified color code range;

[0081] S6.3: Extract risk areas from the fire parameter field data based on fire parameter thresholds. For fire parameters with higher values ​​indicating higher risk, the risk area calculation model is as follows:

[0082]

[0083] For visibility parameters, the risk area is determined according to the following formula:

[0084]

[0085] in, This represents the risk area corresponding to the fire parameter Y at time t. Let x represent the risk area corresponding to the visibility at time t, and let x represent the spatial location in the two-dimensional grid. This represents the fire parameter value at spatial location x at time t. This represents the risk threshold corresponding to fire parameter Y. This indicates the visibility risk threshold;

[0086] S6.4: Perform connectivity analysis on the risk area to extract the risk area area, risk area center, area boundary, parameter peak value, parameter average value, gradient direction, propagation direction, spread velocity, and multi-parameter overlap as structured fire situation features, wherein the risk area area... Risk Area Center Expansion speed and multi-parameter overlap satisfy:

[0087]

[0088]

[0089]

[0090]

[0091] Furthermore, step S7 specifically includes the following steps:

[0092] S7.1: Organize fire parameter values, structured fire situation characteristics, and fire protection rules into structured text input;

[0093] S7.2: Generate professional interpretation labels based on fire parameter thresholds, risk area location, risk development trend, and fire protection rules. The professional interpretation labels include risk area, risk level, judgment basis, development trend, and handling suggestions.

[0094] S7.3: Combine the structured text input and professional explanation labels into a large text model training sample;

[0095] S7.4: Using the training samples of the aforementioned large text model, supervised fine-tuning or efficient parameter fine-tuning of the large text model is performed to obtain a domain-adapted large text model. Its training objective function is as follows:

[0096]

[0097] in, This represents the training loss of a large text model. This indicates the number of lexical units in the target explanatory text. Let Q represent the k-th target word, and let Q represent the input prompt consisting of fire parameter values, structured fire situation features, and fire protection rules. The parameter is The text model outputs the conditional probability of the target word.

[0098] Furthermore, step S8 specifically includes the following steps:

[0099] S8.1: Obtain the fire parameter values ​​corresponding to the working condition to be explained, and serialize the fire parameter values ​​into text input;

[0100] S8.2: Adapt the text input domain to the large text model to generate a professional interpretation of the fire situation, which includes risk areas, risk levels, judgment criteria, development trends, and disposal recommendations;

[0101] S8.3: The risk areas, risk levels, judgment criteria, development trends and handling suggestions in the professional interpretation of the fire situation are compared with the corresponding fire parameter values, structured fire situation characteristics and fire protection rules to obtain consistency verification results;

[0102] S8.4: Calculate the comprehensive consistency score based on the consistency scores of the risk area, development direction, risk level, and handling rules. The calculation model for the comprehensive consistency score is as follows:

[0103]

[0104] Where S represents the overall consistency score, Indicates the consistency score of risk areas. The score indicates the consistency of development direction. Indicates the consistency score of risk level. Indicates the consistency score of the handling rules. , , , These represent the corresponding weights.

[0105] Furthermore, step S9 specifically includes the following steps:

[0106] S9.1: Determine whether the overall consistency score has reached the preset consistency threshold;

[0107] S9.2: When the overall consistency score reaches the preset consistency threshold, output the fire parameter time series cloud map and the professional interpretation of the fire situation; the overall consistency threshold is preferably set to 0.8;

[0108] S9.3: When the overall consistency score does not reach the preset consistency threshold, the inconsistent fields, corresponding fire parameter values, structured fire situation characteristics and fire protection rules are re-entered into the domain-adapted text large model for regeneration.

[0109] S9.4: When the regenerated professional interpretation of the fire situation still fails to reach the preset consistency threshold, the preset rule model is invoked to generate a fallback interpretation result, and the fire parameter time series cloud map and the fallback interpretation result are output.

[0110] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0111] This invention transforms high-dimensional fire parameter fields into low-dimensional coefficient time-series prediction problems through POD (Programmable Analytical Design) modeling. It combines extended operational condition generation within physically feasible limits with a multi-dimensional physical screening mechanism to construct an enhanced dataset based on limited basic simulation samples. Employing a weighted training strategy with high weights on basic samples and auxiliary extended samples, it significantly reduces the number of repeated FDS (Fire Data Set) simulations while ensuring prediction accuracy. This effectively reduces the time and computational costs of acquiring multi-operational fire parameter fields and greatly improves the efficiency of generating fire parameter time-series cloud maps.

[0112] The final fire parameter field prediction model trained by this invention can rapidly generate a full-time, multi-channel fire parameter field and corresponding time-series cloud map based on target operating condition parameters, and simultaneously extract structured situational features such as risk area, propagation direction, expansion speed, and multi-parameter overlap. It can output fire development trends and hazardous area evolution results without performing real-time high-precision simulations, providing efficient data support for fire risk assessment and emergency situation analysis, and meeting the application needs of rapid simulation and decision support.

[0113] This invention integrates structured fire situation characteristics, fire protection rules, and a domain-adaptive text model to output standardized professional interpretations that include risk levels, judgment criteria, development trends, and handling suggestions, lowering the professional interpretation threshold of fire parameter cloud maps. Simultaneously, it introduces a multi-dimensional consistency verification mechanism, coupled with secondary generation and rule model fallback strategies, to ensure consistency between the interpretation results and parameter data and fire protection rules, effectively improving the reliability and engineering applicability of intelligent interpretation results. Attached Figure Description

[0114] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0115] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0116] Figure 2 This is a structural diagram of the parameterized reduced-order fire parameter field prediction model in this invention.

[0117] Figure 3 This is a flowchart of the Chinese text large model interpretation, generation, and consistency verification process in this invention. Detailed Implementation

[0118] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0119] This embodiment focuses on a two-dimensional slice fire parameter field in building rooms, underground spaces, tunnels, or storage spaces. System inputs include the target scene's geometric model, material properties, fire source heat release rate, fire source location, ventilation openings, wind speed, ambient temperature, simulation time step, and a fire safety rule base. System outputs include time-series cloud maps of temperature, smoke concentration, hazardous gas concentration, visibility, or heat flux under the target conditions, along with corresponding risk areas, risk levels, development trends, judgment criteria, and response recommendations.

[0120] To improve operability, this embodiment divides the entire process into ten executable steps: basic simulation, data standardization, reduced-order modeling, initial prediction, sample expansion, augmentation training, cloud map generation, structured feature extraction, large text model interpretation, and consistency verification. Figure 1 As shown, the inputs, outputs, and verification metrics for each stage are as follows.

[0121] 1. Basic working condition construction and FDS simulation

[0122] First, an FDS fire simulation model is established based on the target fire scenario, setting the geometry, material properties, obstacles, opening locations, boundary conditions, and measurement point slices. Each basic working condition is represented by a working condition parameter vector Pi, which is obtained by concatenating fire source parameters, spatial parameters, ventilation parameters, and environmental parameters.

[0123]

[0124] in, The parameters representing the ignition source for the i-th operating condition include at least the heat release rate, ignition source location, and fuel type. Spatial parameters include at least spatial dimensions, obstacle locations, and material properties; Ventilation parameters should include at least the opening area, opening location, and air velocity. This refers to environmental parameters, including at least initial temperature, pressure, or environmental boundary conditions. According to... After performing FDS calculations, a multi-time-step, multi-channel basic fire parameter field is obtained. .

[0125] The operational verification conditions for this step are: the FDS calculation ends normally; the target 2D slice has valid output at each sampling time; and the parameter channels such as temperature, flue gas concentration, harmful gas concentration, visibility, or heat flux have a unified time index. Basic operating conditions that do not converge, are missing, or are significantly beyond the physical range will not be included in subsequent training.

[0126] 2. Unified preprocessing, masking, and normalization

[0127] The basic fire parameter field output by FDS is time-step aligned, uniformly mapped to two-dimensional coordinates, and resampled using a fixed mesh to obtain the parameter field on a uniform mesh. For walls, obstacles, out-of-boundary regions, and uncomputable regions, a geometric mask M(x) is established:

[0128]

[0129] Calculate the training set mean for each parameter channel c. and standard deviation And normalize the effective grid:

[0130]

[0131] In the formula, This represents the fire parameters of type c at time t and spatial location x for the i-th working condition. This represents the normalized parameter field. This formula ensures that the model learns the fire parameter distribution only within the effective grid, avoiding interference from solid areas or areas outside the scene on the prediction results.

[0132] 3. Reduced-order representation and POD coefficient extraction

[0133] The normalized multi-channel parameter field is expanded into vectors according to channels and spatial grids, and a snapshot matrix X is constructed based on different operating conditions and time steps. Singular value decomposition is then performed on the snapshot matrix.

[0134]

[0135] The number of retained POD modes r is determined based on the singular value cumulative energy contribution rate:

[0136]

[0137] in, This represents the k-th singular value. This represents the total number of selectable modes, and η is the cumulative energy threshold, preferably 0.95. This indicates the number of modes retained. The first r left singular vectors from the left singular vector matrix are selected to form the reduced-order basis matrix of POD:

[0138]

[0139] For different types of scenarios, such as building rooms, underground spaces, tunnels, or storage spaces, corresponding reduced-order basis matrices of the Point of View (POD) are constructed to ensure that the reduced-order basis matrices match the spatial structure, boundary conditions, and fire parameter distribution characteristics of the target scenario. This represents the reduced-order basis matrix of POD composed of the first r principal spatial modes. This represents the first to the rth columns of the left singular vector matrix U;

[0140] The high-dimensional fire parameter field corresponding to the i-th basic simulation condition at time t is mapped to the low-dimensional space of POD, and the formula for calculating its low-dimensional coefficients is as follows:

[0141]

[0142] Where vec() represents vectorization operation, This represents the average field of the training samples. The interpretability of this step lies in: This represents the most prominent spatial distribution pattern in the fire parameter cloud map. This indicates the weight of each pattern at the current time step; therefore, changes in the high-dimensional cloud map can be interpreted as the temporal evolution of a small number of major fire spatial patterns.

[0143] 4. Operating Condition Embedding and Low-Dimensional Time Series Prediction

[0144] The structure of the parameterized reduced-order fire parameter field prediction model in this embodiment is as follows: Figure 2 As shown, the working condition parameter vector After normalization encoding, the data is input into a multilayer perceptron to obtain the working condition embedding vector. :

[0145]

[0146] The low-dimensional coefficients of q consecutive historical time steps are input together with the working condition embedding vector into the LSTM time series prediction network to predict the low-dimensional coefficients of the next time step:

[0147]

[0148] Where q is the length of the input time window, and Δt is the interval between adjacent time steps. (Work condition embedding) This enables the model to distinguish between different fire sources, spaces, and ventilation conditions; LSTM is used to describe the dynamic laws of fire parameter field propagation, diffusion, and decay over time. During engineering implementation, q can be set according to the FDS output time interval and the fire development speed, for example, using 5 to 20 consecutive time steps.

[0149] 5. Linear reconstruction, nonlinear residual recovery, and training loss

[0150] Linear reconstruction using predicted low-dimensional coefficients and main spatial modes yields the POD reconstruction parameter field:

[0151]

[0152] The predicted low-dimensional coefficients and the operating condition embedding vector are input into the nonlinear residual recovery network to obtain the nonlinear residual correction term:

[0153]

[0154] And obtain the predicted fire parameter field:

[0155]

[0156] The model training loss consists of low-dimensional coefficient error, parameter field reconstruction error, and physical constraint error.

[0157]

[0158] in, This represents the prediction loss based on low-dimensional coefficients. Indicates the fire parameter field reconstruction loss. Represents physical constraint loss. , and These represent the weighting coefficients of the corresponding loss terms. When the test results reach a preset accuracy threshold, the optimal model is determined as the initial fire parameter field prediction model.

[0159] 6. Extended operating condition generation and physical screening

[0160] Within the physically feasible range of the basic operating condition, perturbations are made to the ignition source, ventilation, space, or environmental parameters to generate extended operating conditions:

[0161]

[0162] Where ΔP is the parameter perturbation that satisfies engineering constraints. Input the initial fire parameter field prediction model to generate an extended fire parameter field. Before the extended samples are included in the augmentation dataset, they need to be physically screened. The screening criteria include: the predicted parameters are within the feasible numerical range; the predicted values ​​at invalid grids are close to zero; the rate of change between adjacent time steps does not exceed the preset threshold γ; and the response direction of the high-temperature zone, smoke zone, or heat flux during fire source enhancement conforms to the fire development law.

[0163] The augmented dataset is formed by merging the selected expanded samples with the base samples. During training, different weights are assigned to the base samples and the expanded samples, preferably with the weights of the base samples greater than those of the expanded samples.

[0164]

[0165] This weighting method ensures that the model is based primarily on real FDS base samples, while utilizing expanded samples selected through physical screening to enhance the coverage of operating conditions, thereby reducing the number of repeated FDS simulations.

[0166] 7. Target Condition Cloud Map Generation and Structured Situation Feature Extraction

[0167] When the user inputs the target operating condition parameters and target time series Afterwards, Input the final fire parameter field prediction model to obtain the normalized prediction field, and then perform inverse normalization processing:

[0168]

[0169] Time-series cloud maps are generated using a unified spatial scale, unified time interval, and unified color scale range. Risk areas are extracted from the fire parameter field data based on fire parameter thresholds. For fire parameters with higher values, indicating higher risk, the risk area calculation model is as follows:

[0170]

[0171] For visibility parameters, the risk area is determined according to the following formula:

[0172]

[0173] After identifying the risk area, connectivity analysis is performed to extract the risk area area, risk area center, area boundary, peak parameter value, average parameter value, gradient direction, propagation direction, spread velocity, and multi-parameter overlap as structured fire situation characteristics. The risk area area is among these parameters. Risk Area Center Expansion speed and multi-parameter overlap satisfy:

[0174]

[0175]

[0176]

[0177]

[0178] Here, ΔA represents the area of ​​a single grid cell, and I is the indicator function. These structured features directly correspond to the risk location, impact range, propagation direction, expansion speed, and degree of multi-risk coupling in fire situation interpretation. Therefore, the interpretation results no longer rely solely on the free generation of a large text model but have clear numerical basis.

[0179] 8. Construction of large-scale text model training samples and generation of professional interpretations

[0180] Fire parameter values, structured situation characteristics, and fire protection rules are organized into a structured text input Q. For example, Q includes fields such as target condition number, time step, maximum temperature, risk area area, risk center coordinates, smoke propagation direction, areas with visibility below the threshold, and fire protection rule trigger items. Based on Q, professional interpretation labels y are constructed, which include risk area, risk level, judgment basis, development trend, and handling recommendations.

[0181] Domain adaptation is performed on a large text model using supervised fine-tuning or parameter-efficient fine-tuning methods. The training objective function is:

[0182]

[0183] in, Indicates the number of lexical units in the target explanatory text. Let represent the k-th target word. Since Q comes from the prediction cloud and structured features, the input of the large text model has a clear numerical source; since y contains the judgment criteria and handling suggestions, the model output can form an explanatory chain of "parameter value - rule threshold - risk judgment - suggested action".

[0184] 9. Consistency verification, regeneration, and fallback output

[0185] After the large text model generates an interpretation, consistency checks are performed on the risk areas, development directions, risk levels, and handling rules within the interpretation. Figure 3 As shown, the overall consistency score S is calculated:

[0186]

[0187] in, Indicates the consistency score of risk areas. The score indicates the consistency of development direction. Indicates the consistency score of risk level. Indicates the consistency score of the handling rules. to Weight and satisfy + + + =1. If S≥ If S < If the inconsistent fields, corresponding numerical features, and fire safety rules are re-entered into the domain-adapted text model for regeneration, the model will be regenerated; if the number of regenerations reaches the preset number... If the verification still fails, the rule model is invoked to generate a fallback explanation.

[0188] Rule-based models prioritize threshold triggers and templated expressions for fallback interpretations, such as: when the area of ​​the high-temperature risk zone... When the area exceeds the preset area threshold and the center of the area moves towards the exit, the output "The high temperature risk area is expanding and spreading to the exit side. The exit evacuation route and smoke exhaust conditions should be checked first" is output. When the overlap between the area with visibility below the threshold and the personnel evacuation route exceeds the preset threshold, the output "The evacuation route is significantly affected by smoke. Smoke exhaust should be started and evacuation guidance should be adjusted" is output.

[0189] In summary, this embodiment fully verifies the feasibility of implementing the entire technology chain of "mechanism simulation - reduced-order representation - sample expansion - intelligent interpretation". During implementation, by setting parameters such as the POD cumulative energy threshold of 0.95 and the comprehensive consistency threshold of 0.8, the high-dimensional fire parameter field is mapped to a low-dimensional coefficient sequence. Combining working condition embedding and LSTM time-series prediction, a highly generalizable prediction model can be trained using only a small amount of FDS basic simulation data. The expanded sample selection and weighted training strategy under physical constraints effectively balances prediction accuracy and working condition coverage. The introduction of structured situation feature extraction and multi-dimensional consistency verification mechanisms avoids the "illusion" problem of content generated by general large models from the numerical root. Experimental results show that this method can reduce the generation time of multi-working-condition fire parameter cloud maps from the traditional several hours to seconds, and the consistency score between the generated interpretation content and the measured parameters is consistently above 0.85. It is fully adaptable to the rapid emergency simulation needs of scenarios such as buildings, tunnels, and underground spaces, providing efficient and reliable technical support for fire risk assessment and response decisions.

[0190] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion, characterized in that, Includes the following steps: S1: Determine the target fire scenario, establish an FDS fire simulation model, set fire source parameters, space parameters, ventilation parameters and environmental parameters, and execute multiple basic simulation conditions; S2: Preprocess and reduce the order of the basic fire parameter field data to obtain the low-dimensional characterization data corresponding to the basic fire parameter field data; S3: Based on the simulated working conditions and low-dimensional characterization data, construct a fire parameter field prediction model, and use the basic fire parameter field data to train, verify and test the fire parameter field prediction model to obtain an initial fire parameter field prediction model. S4: Generate extended working conditions within the physically feasible range, generate extended fire parameter fields using the initial fire parameter field prediction model, and perform physical screening on the extended fire parameter fields; S5: The basic samples are merged with the expanded samples that have passed the physical screening to form an enhanced dataset, and the final fire parameter field prediction model is trained. S6: Use the final fire parameter field prediction model to generate fire parameter field data and fire parameter time series cloud map corresponding to the target working condition, and extract structured fire situation features from the fire parameter field data; S7: Construct a large text model training sample based on fire parameter values, structured fire situation characteristics, fire protection rules and professional interpretation labels, and use the large text model training sample to perform domain adaptation on the large text model to obtain a domain-adapted large text model. S8: Input the fire parameter values ​​corresponding to the working condition to be explained into the domain-adapted text model to generate a professional explanation of the fire situation and perform consistency verification. S9: Outputs a time-series cloud map of fire parameters and a professional interpretation of the fire situation that has passed the consistency check.

2. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1: Establish an FDS fire simulation model based on the target fire scenario, set relevant simulation parameters of the fire scenario in the FDS fire simulation model, and set the measuring points and data slices. The relevant simulation parameters include geometric structure, material properties, fire source parameters, opening location, obstacle distribution, ventilation conditions and environmental conditions. S1.2: Each basic simulation condition is represented as a condition parameter vector consisting of fire source parameters, space parameters, ventilation parameters, and environmental parameters. The basic simulation condition parameter vector is represented as follows: in, This represents the parameter vector corresponding to the i-th basic simulation condition; This represents the fire source parameter vector corresponding to the i-th basic simulation condition; This represents the spatial parameter vector corresponding to the i-th basic simulation condition; This represents the ventilation parameter vector corresponding to the i-th basic simulation condition; This represents the environmental parameter vector corresponding to the i-th basic simulation condition; S1.3: Execute multiple FDS basic simulation conditions according to the aforementioned working condition parameter vector to obtain basic fire parameter field data corresponding to each basic simulation condition.

3. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1: Perform time step alignment, unified two-dimensional coordinate mapping, and fixed grid resampling on the basic fire parameter field data to obtain fire parameter field data under a unified spatiotemporal scale; wherein, the basic fire parameter field data includes temperature field, smoke concentration field, harmful gas concentration field, and visibility field. S2.2: Establish a geometric mask M(x) based on the solid regions, boundary regions, and uncomputable regions in the target fire scene, and use the geometric mask to mark invalid mesh regions: Where M(x)=1 indicates that spatial location x is an effective flow field grid, and M(x)=0 indicates that spatial location x is an invalid grid; after deleting non-converged samples, missing samples, and abnormal samples, each fire parameter channel is normalized, as shown in the following formula: in, This represents the fire parameters of type c at time t and spatial location x for the i-th working condition. and Let represent the mean and standard deviation of the parameter channels for fire type c, respectively. This represents the normalized fire parameter field; the normalization formula is used to unify the dimensions of different parameter channels and ensure that the model learns the fire parameter distribution only on the effective grid. S2.3: Expand the normalized multi-channel fire parameter field into a fire parameter field snapshot vector, and calculate the average fire parameter field vector. And construct the mean-free snapshot matrix X: Singular value decomposition is performed on the snapshot matrix X. The singular value decomposition model is as follows: in, This represents a snapshot matrix composed of multiple snapshots of fire parameter fields. Let Σ denote the left singular vector matrix, and let Σ denote the singular value diagonal matrix. The transpose of the right singular vector matrix; S2.4: Determine the number of retained POD modes based on the singular value cumulative energy contribution rate. The cumulative energy contribution rate calculation model is as follows: in, This represents the k-th singular value. This represents the total number of optional modes in the snapshot matrix. Indicates the cumulative energy threshold. This indicates the number of modes retained. The first r left singular vectors from the left singular vector matrix are selected to form the reduced-order basis matrix of POD: in, This represents the reduced-order basis matrix of POD composed of the first r principal spatial modes. This represents the first to the rth columns of the left singular vector matrix U; The high-dimensional fire parameter field corresponding to the i-th basic simulation condition at time t is mapped to the low-dimensional space of POD, and the formula for calculating its low-dimensional coefficients is as follows: Where vec() represents vectorization operation, Represents the average field of the training samples. This represents the low-dimensional representation data of the i-th operating condition at time t; the low-dimensional coefficients serve as the prediction objects for the subsequent time-series prediction network.

4. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S3 specifically includes the following steps: S3.1: Normalize and encode the fire source parameters, spatial parameters, ventilation parameters, and environmental parameters corresponding to each basic simulation condition to obtain the condition encoding vector, and input the condition encoding vector into the multilayer perceptron to obtain the condition embedding vector: Where norm() represents the normalization encoding operation. This represents the mapping function of a multilayer perceptron. This represents the working condition embedding vector for the i-th working condition; S3.2: Using the low-dimensional representation data obtained in step S2 as the model prediction object, the low-dimensional representation data of q consecutive historical time steps and the operating condition embedding vector are jointly input into the LSTM time series prediction network to predict the low-dimensional representation data of the next time step. The time series prediction model is as follows: in, This indicates that the i-th basic simulation condition is in the next time step. Predicting low-dimensional representation data, This represents the LSTM time series prediction function. This indicates that the i-th basic simulation condition is in continuous A low-dimensional representation data sequence of historical time steps, This represents the condition embedding vector corresponding to the i-th basic simulation condition. Indicates the length of the input time window. Indicates the time interval between adjacent time steps; S3.3: Using the main spatial modes obtained in step S2, the predicted low-dimensional characterization data is linearly reconstructed to obtain a linearly reconstructed fire parameter field, and a nonlinear residual correction term is generated using a transposed convolutional neural network. The linear parameter field is calculated as follows. The predicted low-dimensional coefficients and the embedded vectors of the operating conditions are input into the nonlinear residual recovery network to obtain the nonlinear residual correction term: And obtain the predicted fire parameter field: in, This indicates that the fire parameter field is linearly reconstructed using the POD method. The parameter is Transposed convolutional neural network, This represents the nonlinear residual correction term. This represents the final predicted fire parameter field; S3.4: Train the fire parameter field prediction model using the basic fire parameter field data in the training set, adjust the model parameters using the validation set and determine the optimal model, and evaluate the prediction accuracy of the optimal model using the test set; the following joint training loss is used during the training process: in, This represents the prediction loss based on low-dimensional coefficients. Indicates the fire parameter field reconstruction loss. Represents physical constraint loss. , and These represent the weight coefficients of the corresponding loss terms; the physical constraint loss includes geometric mask constraints, parameter value range constraints, temporal continuity constraints, and fire source neighborhood response rationality constraints; when the test results reach the preset accuracy threshold, the optimal model is determined as the initial fire parameter field prediction model.

5. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1: Generate extended operating conditions within the physically feasible range of the basic simulation operating conditions. The generation model for the extended operating conditions is as follows: in, Represents the extended operating condition parameter vector. This represents the parameter vector for the i-th basic simulation condition. This represents the parameter perturbation vector that satisfies the physical feasible range; S4.2: Perform constraint judgment on the extended working condition to ensure that the extended working condition meets the following requirements: the fire source is not located in a solid area, the opening does not exceed the scene boundary, there is no non-preset overlap between obstacles, and there is no conflict between ventilation conditions. S4.3: Input the extended working conditions determined by the constraints into the initial fire parameter field prediction model to generate corresponding extended fire parameter field data. Then, perform physical screening on the extended fire parameter field data based on the parameter value range, geometric mask, continuity of adjacent time steps, and rationality of the response in the fire source neighborhood. Determine the extended fire parameter field data that meets all screening conditions as valid extended fire parameter field samples.

6. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S5 specifically includes the following steps: S5.1: Merge the base samples with the expanded samples that have passed the physical screening to form an augmented dataset; S5.2: Set training weights for the base samples and expanded samples in the augmented dataset, and train the model according to the weighted training objective. The trained model is as follows: in, Represents the basic sample set, This represents the expanded set of samples that have passed physical screening. Represents the training weights of the basic samples. This represents the training weights for expanded samples, and + =1; S5.3: Using the model parameters of the initial fire parameter field prediction model as initialization parameters, the fire parameter field prediction model is retrained or fine-tuned using the augmented dataset; S5.4: Test the retrained or fine-tuned fire parameter field prediction model using independent test conditions. When the test results reach the preset accuracy threshold, the final fire parameter field prediction model is obtained.

7. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S6 specifically includes the following steps: S6.1: Obtain the fire parameter field data corresponding to the target working condition in the initial time period, perform the same preprocessing as in step S2 on the fire parameter field data, and project it onto the POD reduced-order basis matrix to obtain the initial low-dimensional coefficient sequence corresponding to the target working condition; input the working condition parameters, the initial low-dimensional coefficient sequence and the target time series corresponding to the target working condition into the final fire parameter field prediction model, and recursively generate the predicted fire parameter field data corresponding to each target time step; in, These are the prediction parameters after inverse normalization. These are the normalized fire parameter values ​​predicted by the model. and These represent the mean and standard deviation of the parameter channels for fire type c, respectively. S6.2: Generate fire parameter time series cloud maps according to a unified spatial scale, unified time interval, and unified color code range; S6.3: Extract risk areas from the fire parameter field data based on fire parameter thresholds, wherein the risk area calculation model for fire parameters is as follows: For visibility parameters, the risk area is determined according to the following formula: in, This represents the risk area corresponding to the fire parameter Y at time t. Let x represent the risk area corresponding to the visibility at time t, and let x represent the spatial location in the two-dimensional grid. This represents the fire parameter value at spatial location x at time t. This represents the risk threshold corresponding to fire parameter Y. This indicates the visibility risk threshold; S6.4: Perform connectivity analysis on the risk area to extract the risk area area, risk area center, area boundary, parameter peak value, parameter average value, gradient direction, propagation direction, spread velocity, and multi-parameter overlap as structured fire situation features, wherein the risk area area... Risk Area Center Expansion speed and multi-parameter overlap satisfy:

8. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S7 specifically includes the following steps: S7.1: Organize fire parameter values, structured fire situation characteristics, and fire protection rules into structured text input; S7.2: Generate professional interpretation labels based on fire parameter thresholds, risk area location, risk development trend, and fire protection rules. The professional interpretation labels include risk area, risk level, judgment basis, development trend, and handling suggestions. S7.3: Combine the structured text input and professional explanation labels into a large text model training sample; S7.4: Using the training samples of the aforementioned large text model, supervised fine-tuning or efficient parameter fine-tuning of the large text model is performed to obtain a domain-adapted large text model. Its training objective function is as follows: in, This represents the training loss of a large text model. This indicates the number of lexical units in the target explanatory text. Let Q represent the k-th target word, and let Q represent the input prompt consisting of fire parameter values, structured fire situation features, and fire protection rules. The parameter is The text model outputs the conditional probability of the target word.

9. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S8 specifically includes the following steps: S8.1: Obtain the fire parameter values ​​corresponding to the working condition to be explained, and serialize the fire parameter values ​​into text input; S8.2: Adapt the text input domain to the large text model to generate a professional interpretation of the fire situation, which includes risk areas, risk levels, judgment criteria, development trends, and disposal recommendations; S8.3: The risk areas, risk levels, judgment criteria, development trends and handling suggestions in the professional interpretation of the fire situation are compared with the corresponding fire parameter values, structured fire situation characteristics and fire protection rules to obtain consistency verification results; S8.4: Calculate the comprehensive consistency score based on the consistency scores of the risk area, development direction, risk level, and handling rules. The calculation model for the comprehensive consistency score is as follows: Where S represents the overall consistency score, Indicates the consistency score of risk areas. The score indicates the consistency of development direction. Indicates the consistency score of risk level. Indicates the consistency score of the handling rules. , , , These represent the corresponding weights.

10. The method for generating fire parameter cloud maps and interpreting large models based on reduced-order sample expansion according to claim 1, characterized in that, Step S9 specifically includes the following steps: S9.1: Determine whether the overall consistency score has reached the preset consistency threshold; S9.2: When the overall consistency score reaches the preset consistency threshold, output the fire parameter time series cloud map and the professional interpretation of the fire situation; S9.3: When the overall consistency score does not reach the preset consistency threshold, the inconsistent fields, corresponding fire parameter values, structured fire situation characteristics and fire protection rules are re-entered into the domain-adapted text large model for regeneration. S9.4: When the regenerated professional interpretation of the fire situation still fails to reach the preset consistency threshold, the preset rule model is invoked to generate a fallback interpretation result, and the fire parameter time series cloud map and the fallback interpretation result are output.

Images