A train fire multi-modal simulation dataset construction method and simulation dataset

Abstract

The application provides a train fire multi-modal simulation data set construction method and simulation data set, comprising the following steps: constructing a train carriage geometric model by using FDS, and setting material parameters and environmental conditions in the train carriage geometric model; simulating a fire full life cycle combustion process based on a segmented fire physical model, and synchronously outputting multi-modal data and heat release rate time series data as true value labels in a simulation process with a unified time stamp; performing time and space alignment on the multi-modal data, and adding a unified reference time stamp index; changing parameters of the segmented fire physical model by parameterized scanning, generating multiple different independent fire evolution conditions, and the parameters comprise a fire growth coefficient, a fire decay coefficient and a steady state duration; and performing structured storage on the multi-modal data, and generating a train fire multi-modal simulation data set. The application has the beneficial effect of realizing millisecond-level accurate time and space alignment of multi-modal data, and providing high-quality training data for a deep learning model.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of fire safety and artificial intelligence, and in particular relates to a method for constructing a multimodal simulation dataset for train fires and a simulation dataset. Background Technology

[0002] Subway train carriages are typical long and narrow tube corridor structures, characterized by dense crowds, difficult evacuation, and significant impact of piston wind on fire spread. Their characteristics differ greatly from those of ordinary building fires, making the research of high-precision real-time fire early warning and prediction systems of significant engineering value.

[0003] Currently, deep learning-based fire detection and prediction technologies have shown great potential and are gradually becoming a research hotspot in subway fire prediction technology. However, large-scale, high-quality, and multimodal datasets are the key foundation for supporting the training of deep learning models. The rail transit field has long lacked structured open datasets, mainly exhibiting the following technical shortcomings: existing datasets are mostly geared towards general scenarios such as indoor residential fires and open wildfires, failing to reflect the unique fire spread characteristics of subway trains; data types are singular and fragmented, primarily consisting of visible light images, with sensor and visual data not accurately aligned in time and space, thus failing to support the training of multimodal time-series models; and most datasets only support instantaneous classification tasks of "whether there is a fire," without coupling with fire dynamics parameters such as heat release rate, failing to meet the need for quantitative extrapolation of energy evolution trends. Furthermore, real-world subway fire experiments are extremely costly, dangerous, and non-reproducible, making it difficult to collect large-scale, full-lifecycle data.

[0004] Therefore, there is an urgent need for a simulation dataset construction method that is specifically designed for subway train scenarios, multimodal and time-aligned, and covers the entire life cycle of a fire, in order to solve the above-mentioned problems in existing technologies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for constructing a multimodal simulation dataset for train fires and a simulation dataset. This method is particularly suitable for solving technical problems such as mismatched scenarios, limited data types, spatiotemporal asynchrony, and inability to support trend prediction in existing datasets, thus providing a high-quality training data foundation for intelligent early warning and trend prediction of subway train fires.

[0006] The technical solution adopted in this invention is as follows: Firstly, a method for constructing a multimodal simulation dataset for train fires is provided, comprising the following steps:

[0007] A full-size train carriage geometric model was constructed using a fire dynamics simulator. The thermal properties of the interior materials, the combustion characteristics of the interior materials, the initial environmental parameters, and the ventilation boundary conditions were set in the carriage geometric model.

[0008] The simulation of the entire life cycle of a fire is based on a segmented fire physical model. During the simulation, multimodal data and time-series data of heat release rate as truth labels are synchronously output with a unified timestamp. The multimodal data includes 2D image sequences, 1D spatial temperature sequences and 0D point smoke concentration sequences.

[0009] The multimodal data is spatiotemporally aligned and a unified reference time timestamp index is attached.

[0010] By adjusting the parameters of the segmented fire physical model through parametric scanning, a variety of different independent fire evolution conditions are generated. The parameters of the segmented fire physical model include the fire growth coefficient, the fire decay coefficient, and the steady-state duration.

[0011] The multimodal data and heat release rate time series data are stored in a structured manner. The data is organized using a directory structure that is hierarchical according to fire evolution conditions, data types, and timestamps. A cross-modal mapping relationship table is generated, and the association between different modal data at the same reference time is established, thereby generating a train fire multimodal simulation dataset.

[0012] Furthermore, for the aforementioned carriage geometric model, a structured hexahedral mesh with verified mesh independence is used to divide the computational domain, including the following steps:

[0013] The computational domain was divided using structured hexahedral meshes of various sizes.

[0014] Simulations were performed under the same fire source parameters and environmental conditions to compare the fire dynamics characteristic parameters under different mesh sizes.

[0015] When the grid size is reduced to a certain value, the fire dynamics characteristic parameters no longer change significantly, and this size is determined to be the optimal grid size.

[0016] Furthermore, the side length of the structured hexahedral mesh is 0.1m.

[0017] Furthermore, the segmented fire physical model is used to define the boundary conditions for the heat release rate of the fire source in the FDS simulation. Its specific expression is:

[0018] Growth stage: , This represents the fire growth coefficient. This is the moment when growth reaches a steady state;

[0019] Steady-state phase: ,in Peak heat release rate, This is the time when decay begins;

[0020] Decay phase: ,in This is the fire attenuation coefficient. This marks the moment the fire ended.

[0021] Furthermore, by parametrically scanning and changing the parameters of the segmented fire physical model, various independent fire evolution scenarios are generated, including the following steps:

[0022] The growth coefficients in the segmented fire physical model were changed respectively. Fire attenuation coefficient and steady-state duration Select multiple fire growth coefficients covering different fire growth rates. Multiple fire decay coefficients covering different fire decay rates And orthogonally combine multiple steady-state duration values ​​covering different steady-state combustion durations, and based on the fire growth coefficient The corresponding fire scale matches the corresponding peak heat release rate. It generates multiple independent fire evolution scenarios covering the entire life cycle from smoldering, growth, steady state to decay, among which... .

[0023] Furthermore, the fire growth coefficient and fire attenuation coefficient The range of values ​​is The range of values ​​for the steady-state duration is: .

[0024] Furthermore, the spatiotemporal alignment step specifically involves: using the integer second time corresponding to the preset sampling frequency as a benchmark, aligning the 2D image sequence, 1D spatial temperature sequence, and 0D point smoke concentration sequence using a timestamp nearest neighbor matching algorithm.

[0025] Furthermore, the frequency of the unified timestamp synchronization output is 1Hz.

[0026] Furthermore, the cross-modal mapping table supports extracting multimodal data sequences within a corresponding time period based on a custom time window length.

[0027] Secondly, a multimodal simulation dataset for predicting train fire trends is provided, wherein the multimodal simulation dataset is constructed by any of the methods described above.

[0028] The advantages and positive effects of this invention are as follows: By adopting the above technical solution, millisecond-level accurate spatiotemporal alignment of multimodal data is achieved, providing high-quality training data for multimodal time-series deep learning models; based on the classic classification of fire dynamics, multiple independent fire evolution conditions covering the entire life cycle from smoldering, growth, steady state to decay are generated through parametric scanning, systematically covering all typical fire scenarios that may occur on subway trains; adopting a hierarchical directory structure and cross-modal mapping relationship table, it supports user-defined time window lengths, enabling rapid extraction of multimodal data sequences within any time period, facilitating use for different prediction tasks. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a simulation dataset construction method according to an embodiment of the present invention. Detailed Implementation

[0030] The present disclosure will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments of the present disclosure. The technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative effort are within the scope of protection of the present disclosure.

[0031] like Figure 1 As shown, this invention provides a method for constructing a multimodal simulation dataset for train fires, comprising the following steps:

[0032] S10. Construct a full-size train carriage geometric model using a fire dynamics simulator, and set the thermal properties of the interior materials, the combustion characteristics parameters of the interior materials, the initial environmental parameters, and the ventilation boundary conditions in the carriage geometric model.

[0033] Specifically, a 1:1 scale geometric model of a real subway train car was constructed using the Fire Dynamics Simulator (FDS). The train car's geometric model included longitudinal seats on both sides, doors, and windows, with accurate material parameters and environmental conditions set to ensure the simulation results accurately reflected the fire combustion characteristics inside the subway train car. The thermophysical properties of the interior materials included density, specific heat, and thermal conductivity; combustion characteristic parameters included heat of combustion and smoke production rate; initial environmental parameters included initial temperature and atmospheric pressure; and ventilation boundary conditions included the location and velocity of air inlets and outlets. The interior materials included polyurethane foam for the seats, rubber flooring, fiberglass for the side walls, aluminum alloy, and stainless steel (AISI 304).

[0034] Preferably, the carriage dimensions are 5.7m (length) × 3.3m (width) × 3.8m (height), and the interior includes core structures such as longitudinal seats on both sides (0.6m long × 1.1m wide × 0.95m high), doors (0.8m wide × 2.2m high, located at both ends of the carriage), and windows (1.025m wide × 0.95m high, two on each side), completely replicating the airflow circulation channels and flammable material distribution within the carriage. Seat polyurethane foam: density 32kg / m³ 3 Specific heat 1.4 kJ / (kg·K), thermal conductivity 0.034 W / (m·K), heat of combustion 24000 kJ / kg, smoke yield 0.12 g / g; Rubber flooring cloth: density 1100 kg / m 3 Specific heat 1.5 kJ / (kg·K), thermal conductivity 0.25 W / (m·K), heat of combustion 30000 kJ / kg, smoke yield 0.10 g / g; sidewall fiberglass: density 1800 kg / m³ 3 Specific heat 1.17 kJ / (kg·K), thermal conductivity 0.35 W / (m·K), heat of combustion 25000 kJ / kg, smoke yield 0.12 g / g; Aluminum alloy and stainless steel: as non-combustible structural materials, only heat conduction parameters are set, not heat of combustion and smoke yield. Initial environmental parameters: initial temperature 20℃, atmospheric pressure 101325 Pa. Ventilation boundary conditions: mechanical ventilation is used, with the air supply vent located at the center of the top of the carriage, and a wind speed of 2 m / s; the exhaust vent is located above the two end doors, with a wind speed of 2 m / s. The fire source is placed at the geometric center of the carriage, with an effective combustion area of ​​1.0 m × 1.0 m.

[0035] S20. Simulate the entire life cycle combustion process of a fire based on a segmented fire physical model. During the simulation, multimodal data and heat release rate time series data as truth labels are synchronously output with a unified timestamp. The multimodal data includes 2D image sequences, 1D spatial temperature sequences and 0D point smoke concentration sequences.

[0036] Specifically, a segmented fire physics model that conforms to the actual combustion laws of fires is adopted, which can accurately simulate the entire life cycle of a fire from smoldering, growth, steady state to decay. During the FDS simulation, multimodal data is synchronously output with a unified timestamp through a virtual measurement device. Each type of multimodal data is accompanied by HRR (Heat Release Rate) time-series data and a timestamp, serving as the true prediction value. This includes 2D image sequences reflecting the visual morphology of flames and smoke, 1D distributed fiber optic spatial temperature sequences reflecting the spatial temperature distribution inside the carriage, and 0D point smoke concentration sequences (1D time-series smoke concentration data and 1D time-series wind speed data reflecting single-point smoke concentration and wind speed), providing comprehensive and rich training data for the deep learning model. The virtual measurement device includes a virtual camera, distributed fiber optic temperature sensors, and point sensors. The virtual camera is placed inside the carriage to capture the visual morphology of flames and smoke; two distributed fiber optic temperature sensors are symmetrically laid along the longitudinal direction of the carriage to collect the spatial temperature distribution along the length of the carriage; the point sensors are placed at the center of the carriage ceiling to collect smoke concentration and wind speed data.

[0037] Preferably, the total FDS simulation duration is set to 1200s, with the time step automatically controlled by CFL conditions and a maximum Coulomb number of 1.0. The turbulence model uses Large Eddy Simulation (LES), the combustion model uses a mixed fractional model, and the radiation model uses a gray-body radiation model. During the simulation, the following four types of data are synchronously collected via a virtual measurement device at a preset sampling frequency of 1Hz: 2D image sequence: Video frames captured by a virtual camera are exported using the FDS SmokeView module, with a resolution of 1280×720 pixels. The virtual camera is positioned at one end of the carriage, facing the fire source, to capture the visual morphology of the flames and smoke. 1D spatial temperature sequence: Two distributed fiber optic temperature sensors are symmetrically laid longitudinally on both sides of the inner wall of the carriage top, with a height Z=2.8m, a total fiber length of 5.7m, and a spatial resolution of 0.1m, totaling 58 spatial sampling points, to collect the spatial temperature distribution along the length of the carriage. 0D point-based time-series data: Point sensors are placed at the center of the carriage ceiling (X=1.5m, Y=2.85m, Z=3.15m) to collect time-series data on flue gas concentration (% / m) and wind speed (m / s). HRR time-series data: Heat release rate data from FDS simulation output are collected as ground truth labels for training the deep learning model.

[0038] S30. Perform spatiotemporal alignment on the multimodal data and attach a unified reference time timestamp index;

[0039] Specifically, because FDS uses a dynamically adaptive time step for solving the problem, the actual output timestamps of the image and sensor data are not strictly integer seconds, leading to a natural timestamp mismatch between different modalities. Matching the closest image and sensor data for each reference time point enables precise spatiotemporal alignment of multimodal data, ensuring the quality of the training data.

[0040] S40. By parametrically scanning, the parameters of the segmented fire physical model are changed to generate a variety of different independent fire evolution conditions. The parameters include the fire growth coefficient, the fire decay coefficient, and the steady-state duration.

[0041] S50. The multimodal data and heat release rate time series data are stored in a structured manner. The data is organized in a directory structure that is hierarchical according to fire evolution conditions, data type and timestamp. A cross-modal mapping relationship table is generated to establish the association between different modal data at the same reference time, thereby generating a train fire multimodal simulation dataset.

[0042] Preferably, all collected data is stored in a structured manner, using a three-level directory structure of "operating condition-data type-timestamp": First-level directory: divided by fire evolution operating conditions, with each operating condition corresponding to an independent folder; Second-level directory: divided by data type, with each operating condition folder containing multiple subfolders; Third-level: data from different times is distinguished by the timestamp in the filename. A cross-modal mapping table is generated in CSV format, with each row corresponding to a baseline time and containing the storage path information of all modal data files at that time. Users only need to specify the start time and length of the input time window to quickly extract the corresponding multimodal data sequence from the mapping table without traversing the entire file system.

[0043] The above methods are highly targeted to specific scenarios, realistically recreating the characteristics of fire spread in subway trains and solving the problem of poor applicability of general datasets; they also simultaneously collect multimodal heterogeneous data and accurately align it in time and space, providing a high-quality training foundation for time-series deep learning models; and they employ hierarchical storage and mapping table structures to achieve efficient organization and flexible access to large-scale data.

[0044] To address the issue of undefined mesh size selection in FDS simulations and to avoid distorted results due to overly coarse meshes or computational overload due to overly fine meshes, this embodiment provides an implementation method.

[0045] In one embodiment, for the geometric model of the carriage, a structured hexahedral mesh with verified mesh independence is used to divide the computational domain, including the following steps:

[0046] The computational domain was divided using structured hexahedral meshes of various sizes;

[0047] Simulations were conducted under the same fire source parameters and environmental conditions to compare the fire dynamics characteristic parameters under different mesh sizes (size of structured hexahedral mesh). The fire dynamics characteristic parameters include the heat release rate (HRR) curve, the temperature of the plume axis at the center of the fire source, and the temperature of the lateral entrainment zone near the fire source.

[0048] When the grid size is reduced to a certain value, the fire dynamics characteristic parameters no longer change significantly, and this size is determined to be the optimal grid size.

[0049] Preferably, the computational domain was divided using seven different structured hexahedral meshes with side lengths of 0.10m, 0.15m, 0.20m, 0.25m, 0.30m, 0.35m, and 0.40m. Simulations were conducted under the same fire source parameters and environmental conditions, comparing the heat release rate (HRR) curves, the temperature of the plume axis at the center of the fire source, and the temperature of the near-field lateral entrainment zone of the fire source under different mesh sizes. Ultimately, when the mesh size was reduced to 0.10m, the aforementioned fire dynamics characteristic parameters no longer changed significantly, and the temperature of the plume axis at the center of the fire source remained within the theoretical temperature range of the continuous flame zone, which is 750–900℃. Therefore, 0.10m was determined to be the optimal mesh size.

[0050] By adopting the above method, the simulation results are ensured to be independent of the mesh size, thus ensuring the accuracy and reliability of numerical calculations; while ensuring accuracy, the computational efficiency is maximized, reducing the time cost of large-scale working condition simulations.

[0051] To address the problem that existing fire models cannot accurately describe the evolution of the entire life cycle, this embodiment provides an implementation method.

[0052] In one embodiment, a segmented fire physical model is used to define the boundary conditions for the heat release rate of the fire source in the fire dynamics simulator simulation. Its specific expression is:

[0053] Growth stage: , This represents the fire growth coefficient. This is the moment when growth reaches a steady state;

[0054] Steady-state phase: ,in Peak heat release rate, This is the time when decay begins;

[0055] Decay phase: ,in This is the fire attenuation coefficient. This marks the moment the fire ended.

[0056] The above method conforms to the classical laws of fire dynamics and accurately simulates the entire process of smoldering, growth, steady state, and decay; it provides a standardized model basis for subsequent parametric scanning to generate multiple working conditions.

[0057] To address the issues of existing parameter selection lacking scientific basis and incomplete coverage of operating conditions, this embodiment provides an implementation method.

[0058] In one embodiment, generating multiple different independent fire evolution scenarios by changing the parameters of a segmented fire physical model through parametric scanning includes the following steps:

[0059] Change the growth coefficient in the segmented fire physical model respectively Fire attenuation coefficient and steady-state duration Select multiple fire growth coefficients covering different fire growth rates. Multiple fire decay coefficients covering different fire decay rates And orthogonally combine multiple steady-state duration values ​​covering different steady-state combustion durations, and based on the fire growth coefficient The corresponding fire scale matches the corresponding peak heat release rate. It generates multiple independent fire evolution scenarios covering the entire fire lifecycle from smoldering, growth, steady state to decay, with the steady state duration being... .

[0060] Preferably, five typical fire growth coefficients are selected. Value and fire attenuation coefficient Value, 0.0029 kW / s 2 (Slow fire, corresponding to smoldering state), 0.01127 kW / s 2 (Medium-speed fire), 0.04689 kW / s 2 (Classic rapid fire), 0.1042kW / s 2 (Typical subway fire), 0.1876 kW / s 2 (Ultra-fast fire); Five typical steady-state duration values ​​of 30s, 60s, 90s, 120s, and 180s were selected; the above parameters were orthogonally combined to generate a total of 125 independent fire evolution scenarios; the corresponding peak heat release rate was matched according to the fire growth coefficient. :when In the event of a rapid or above fire, for ;when (In the case of medium-speed and slow-speed fires) for Each operating condition lasts for 1200 seconds, with a sampling frequency of 1 Hz. Each group can output 1200 frames of strictly time-aligned multimodal data (2D image sequence, 1D spatial temperature sequence, 0D point-based time series data), generating a total of approximately 150,000 sets of synchronous multimodal samples.

[0061] Using the above method, parameters are selected based on fire dynamics classification, and the system covers all typical fire development processes; an orthogonal combination method is used to achieve the most comprehensive scenario coverage with the fewest number of working conditions.

[0062] In one embodiment, the fire growth factor and fire attenuation coefficient The range of values ​​is The range of values ​​for the steady-state duration is: .

[0063] In one embodiment, peak heat release rate The matching rule is: when the fire growth coefficient Peak heat release rate for When the fire growth coefficient Peak heat release rate for .

[0064] To address the issue that the actual output image and sensor data timestamps of FDS are not strictly integer seconds due to the use of dynamic adaptive time steps for solving the problem, this embodiment provides an implementation method.

[0065] In one embodiment, the spatiotemporal alignment step specifically involves: using the integer second time corresponding to a preset sampling frequency as a benchmark, a timestamp nearest neighbor matching algorithm is employed to align the 2D image sequence, the 1D spatial temperature sequence, and the 0D point smoke concentration sequence. All modal data streams are resampled to a unified time grid and aligned by timestamp, ensuring strict alignment of image frames, line temperature data, and point sampling on the time axis.

[0066] Preferably, a timestamp nearest neighbor matching algorithm is used to achieve accurate spatiotemporal alignment of multimodal data: A baseline time sequence is determined: from t=0s to t=1200s, a total of 1201 integer seconds. For each baseline time t, among all images output by the FDS, the frame with the smallest absolute difference between the actual output time and t is found. For each baseline time t, among all sensor records output by the FDS, the set of temperature data, flue gas concentration data, and wind speed data with the smallest absolute difference between the actual recording time and t is found. When two records have the same absolute difference from the baseline time, the earlier record is selected as the matching result. The images, temperature, flue gas, wind speed, and HRR data for the same baseline time are packaged, and this baseline time is used as a unified timestamp index. Since the underlying solution step size of FDS is typically less than 0.1s, the maximum time error generated by the above matching will not exceed 0.05s, which is negligible compared to a 1s sampling interval and can meet the data synchronization accuracy requirements for deep learning model training.

[0067] In one embodiment, the frequency of the unified timestamp synchronization output is 1Hz.

[0068] In one embodiment, the cross-modal mapping table supports extracting multimodal data sequences within a corresponding time period based on a custom time window length.

[0069] In one embodiment, to evaluate the effectiveness of the dataset constructed according to the present invention, the dataset is randomly divided into a training set, a validation set, and a test set in an 8:1:1 ratio, and a Transformer-based deep learning model is used for training and testing. The model input is set to the multimodal historical data of the previous 300 seconds, predicting the HRR evolution curve of the next 600 seconds. The model training uses mean squared error (MSE) as the loss function and uses the Adam optimizer (Batch Size=4, initial learning rate 1×10⁻⁶). -4 This was accomplished on a single NVIDIA RTX 5090 GPU.

[0070] Based on embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0071] An electronic device includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the train fire multimodal simulation dataset construction method provided in this disclosure.

[0072] Electronic devices are intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0073] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the train fire multimodal simulation dataset construction method provided in this disclosure.

[0074] The various embodiments of this disclosure can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0075] A computer program product includes a computer program / instructions, which are executed by a processor. This disclosure provides a method for constructing a multimodal simulation dataset for train fires.

[0076] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0077] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0078] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for constructing a multimodal simulation dataset for train fires, characterized in that, Includes the following steps: A full-size train carriage geometric model was constructed using a fire dynamics simulator. The thermal properties of the interior materials, the combustion characteristics of the interior materials, the initial environmental parameters, and the ventilation boundary conditions were set in the carriage geometric model. The simulation of the entire life cycle of a fire is based on a segmented fire physical model. During the simulation, multimodal data and time-series data of heat release rate as truth labels are synchronously output with a unified timestamp. The multimodal data includes 2D image sequences, 1D spatial temperature sequences and 0D point smoke concentration sequences. The multimodal data is spatiotemporally aligned and a unified reference time timestamp index is attached. By changing the parameters of the segmented fire physical model through parametric scanning, a variety of different independent fire evolution conditions can be generated. The parameters of the segmented fire physical model include the fire growth coefficient, the fire decay coefficient, and the steady-state duration. The multimodal data and heat release rate time series data are stored in a structured manner. The data is organized using a directory structure that is hierarchical according to fire evolution conditions, data types, and timestamps. A cross-modal mapping relationship table is generated, and the association between different modal data at the same reference time is established, thereby generating a train fire multimodal simulation dataset.

2. The simulation dataset construction method according to claim 1, characterized in that, For the aforementioned carriage geometry model, a structured hexahedral mesh with verified mesh independence is used to divide the computational domain, including the following steps: The computational domain was divided using structured hexahedral meshes of various sizes; Simulations were performed under the same fire source parameters and environmental conditions to compare the fire dynamics characteristic parameters under different mesh sizes. When the grid size is reduced to a certain value, the fire dynamics characteristic parameters no longer change significantly, and this size is determined to be the optimal grid size.

3. The simulation dataset construction method according to claim 2, characterized in that: The side length of the structured hexahedral mesh is 0.1m.

4. The simulation dataset construction method according to claim 1, characterized in that, The segmented fire physical model is used to define the boundary conditions for the heat release rate of the fire source simulated by the fire dynamics simulator. Its specific expression is: Growth stage: , This represents the fire growth coefficient. This is the moment when growth reaches a steady state; Steady-state phase: ,in This represents the peak heat release rate. This is the time when decay begins; Decay phase: ,in This is the fire attenuation coefficient. This marks the moment the fire ended.

5. The simulation dataset construction method according to claim 4, characterized in that, The process of generating various independent fire evolution scenarios by changing the parameters of the segmented fire physical model through parametric scanning includes the following steps: The growth coefficients in the segmented fire physical model were changed respectively. Fire attenuation coefficient and steady-state duration Select multiple fire growth coefficients covering different fire growth rates. Multiple fire decay coefficients covering different fire decay rates And orthogonally combine multiple steady-state duration values ​​covering different steady-state combustion durations, and based on the fire growth coefficient The corresponding fire scale matches the corresponding peak heat release rate. It generates multiple independent fire evolution scenarios covering the entire life cycle from smoldering, growth, steady state to decay, among which... .

6. The simulation dataset construction method according to claim 5, characterized in that: Fire growth coefficient and fire attenuation coefficient The range of values ​​is The range of values ​​for the steady-state duration is: .

7. The simulation dataset construction method according to claim 1, characterized in that, The spatiotemporal alignment step specifically involves aligning the 2D image sequence, 1D spatial temperature sequence, and 0D point smoke concentration sequence using the integer second time corresponding to the preset sampling frequency as a benchmark and employing a timestamp nearest neighbor matching algorithm.

8. The simulation dataset construction method according to claim 1, characterized in that: The frequency of the unified timestamp synchronization output is 1Hz.

9. The simulation dataset construction method according to claim 1, characterized in that: The cross-modal mapping table supports extracting multimodal data sequences within a corresponding time period based on a custom time window length.

10. A multimodal simulation dataset for predicting train fire trends, characterized in that, The dataset is constructed using the method described in any one of claims 1-9.

Images