A method for predicting the trend of indoor fire spread

Abstract

The application discloses an indoor fire spreading trend prediction method and relates to the technical field of fire prevention and extinguishing. A physical information guided hybrid neural network model is constructed, and complex fire dynamics partial differential equations are embedded in network training in the form of loss function residual terms. In the reasoning stage, the model gets rid of the heavy large matrix iterative solving process of traditional numerical simulation, realizes full-field physical field reasoning within seconds through pure forward propagation, and meets the harsh requirements of real-time spreading trend prediction of fire emergency sites.

Description

Technical Field

[0001] This invention relates to the field of fire prevention and extinguishing technology, and more specifically to a method for predicting the spread trend of indoor fires. Background Technology

[0002] Indoor fires are characterized by their suddenness, transient variability, and destructiveness. Quantitative and real-time prediction of the spatiotemporal evolution of the fire's physical field (such as temperature, smoke velocity, and toxic smoke concentration) during a fire is crucial for on-site fire rescue command, personnel evacuation route planning, and building structural safety assessment.

[0003] Traditional techniques are mainly divided into two categories: The first category is numerical simulation methods based on the principles of fire dynamics, such as fire simulation software that utilizes computational fluid dynamics. This method obtains fire flow field information by numerically solving a system of partial differential equations involving mass, momentum, energy, and component transport. Although the physical mechanism of this method is clear and it has high prediction accuracy when the boundary conditions are accurate, its computational load is extremely large, the spatial grid is densely divided, and the iterative solution usually takes several hours or even days, which is completely unable to meet the urgent need for real-time prediction and dynamic response at the fire scene.

[0004] The second category is based on purely data-driven deep learning prediction methods. For example, these methods utilize long short-term memory networks or convolutional neural networks to perform black-box training on time-series data from sensors deployed inside buildings, directly predicting future temperature or smoke concentration changes. This method is computationally fast during the inference phase and can achieve real-time output. However, purely data-driven methods ignore the inherent physical laws of fire fluid dynamics. When sensor data is sparse, noisy, or subject to local failures, the prediction results are prone to numerical oscillations or non-convergence that defy physical laws (e.g., irrational drops in local temperature, failure to satisfy the law of conservation of mass, etc.). Furthermore, purely data-driven models heavily rely on massive amounts of historical training samples; for building spatial layouts that have not been trained on, their generalization ability is extremely poor, and prediction distortion is highly likely.

[0005] In particular, existing fire monitoring systems are often limited by hardware costs and transmission bandwidth, resulting in low sampling frequencies for multi-source sensors. Under low-frequency sampling conditions, the lack of temporal characteristics means that purely data-driven models cannot capture the transient changes in the early stages of a fire; while traditional physical simulation methods accumulate errors due to the lack of high-frequency boundary condition corrections. Furthermore, existing technologies struggle to couple fire smoke field prediction with the thermal rise failure of the building's steel structure in real time, failing to simultaneously provide structural collapse risk assessments while predicting fire spread. Moreover, they lack a quantitative characterization of the uncertainty in prediction results, thus failing to provide confidence level references for fire commanders.

[0006] Therefore, it is necessary to propose a method for predicting the spread trend of indoor fires to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to solve the problem that existing fire prediction methods are unable to output a high-fidelity, refined indoor fire spread trend field that strictly conforms to the laws of physical conservation while meeting real-time requirements, especially when the prediction accuracy drops sharply under sparse sensor arrangement and low-frequency sampling conditions.

[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0009] A method for predicting the spread trend of indoor fires includes:

[0010] Acquire real-time environmental parameter sequences from indoor multi-source sensors during a fire;

[0011] Construct a computational grid model of the building space to be predicted;

[0012] The real-time environmental parameter sequence is input into a trained physical information-guided hybrid neural network model, which infers and outputs the spatiotemporal evolution sequence of the fire physical field on the computational grid model in future time periods.

[0013] The fire physical field includes a temperature field, a velocity field, and a smoke concentration field;

[0014] The loss function of the physical information-guided hybrid neural network model during training includes at least the data fitting error term and the residual term of the fire dynamics control equation, and the loss function dynamically assigns weights to the residual term according to the intensity of the fire in the grid area.

[0015] Based on the spatiotemporal evolution sequence, the indoor fire spread trend is generated and output.

[0016] Furthermore, the multi-source sensor includes a temperature sensor, a flue gas concentration sensor, and an infrared thermal imager.

[0017] Furthermore, the physical information-guided hybrid neural network model adopts an encoder-decoder structure, in which the encoder maps the sequence of historical environmental parameters into physical latent space features, and the decoder reconstructs the future fire physical field from the physical latent space features.

[0018] Furthermore, the sampling frequency of the real-time environmental parameter sequence is from 0.2Hz to 5Hz.

[0019] Furthermore, the sampling frequency is 0.25Hz; at this sampling frequency, the physical information guides the hybrid neural network model to compensate for the lack of temporal information through physical constraints, achieving a prediction similarity of over 90%.

[0020] Furthermore, the fire dynamics control equations include energy conservation equations, momentum conservation equations, and component transport equations.

[0021] Furthermore, the fire dynamics control equation also includes a structural heat transfer equation describing the temperature rise of the steel structure, and the output fire physical field spatiotemporal evolution sequence also includes the temperature distribution of structural components to synchronously output the structural collapse risk trend.

[0022] Furthermore, the training of the physical information-guided hybrid neural network model adopts a course learning strategy, which first pre-trains the physical prior knowledge on simulation data, and then fine-tunes it using real fire experiment data.

[0023] Furthermore, the physical information-guided hybrid neural network model is constructed as a temporal dual-stream structure: one stream is based on a physical simulator to quickly infer macroscopic trends, and the other stream is based on a data network to correct local details. The two streams are coupled and corrected to each other during the inference process.

[0024] Furthermore, the generated and output indoor fire spread trend also includes an uncertainty cloud map of the spread path, providing risk areas at different confidence levels.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. This invention constructs a physical information-guided hybrid neural network model, embedding complex fire dynamics partial differential equations into the network training as residual terms of a loss function. During the inference phase, this model eliminates the cumbersome large-matrix iterative solution process of traditional numerical simulations, achieving full-field physical field inference within seconds through pure forward propagation, thus meeting the stringent requirements of real-time spread trend prediction in fire emergency scenarios.

[0027] 2. This invention utilizes the strong physical constraints of the fire dynamics control equations. The model can automatically compensate for missing temporal features and boundary conditions through physical conservation laws under extremely sparse temporal conditions with sensor sampling frequencies as low as 0.25Hz. This avoids the prediction distortion that is prone to occur in pure data-driven models when sampling at low frequencies, ensuring that the prediction similarity is above 90%, and significantly reducing the dependence on hardware transmission bandwidth and sensor response frequency.

[0028] 3. The loss function of this invention can dynamically adjust the weights of the residual terms of the physical control equations according to the intensity of the fire in the computational grid area. In areas with large gradient changes, such as the fire source center and the intense combustion front, the weights are adaptively increased or adjusted to force the neural network to accurately fit complex nonlinear fluid behavior, effectively solving the common problem of traditional PINN models easily getting trapped in local optima or gradient vanishing when dealing with high Reynolds number and high buoyancy driven flow fields. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a schematic diagram of the input and output of the method of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figure 1 and Figure 2 This invention provides a method for predicting the spread trend of indoor fires, the specific implementation of which is as follows:

[0033] Example 1

[0034] This embodiment details a method for predicting the spread of indoor fires within a standard large multi-story office building. The office building has a single-floor area of ​​2,000 square meters and includes a central corridor, symmetrically distributed offices, meeting rooms, and electrical distribution rooms (both high-voltage and low-voltage).

[0035] Step 1: Obtain the real-time environmental parameter sequence from indoor multi-source sensors during the fire process.

[0036] In this embodiment, an Internet of Things (IoT) multi-source sensor network had been pre-deployed inside the office building to be monitored. The multi-source sensors include point temperature sensors evenly installed on the ceilings of each room and corridor, infrared photoelectric smoke concentration sensors, and infrared thermal imagers deployed at key evacuation route intersections and the entrances to power distribution rooms.

[0037] When a fire occurs, all sensor nodes are activated and continuously collect data. Temperature sensors collect the Celsius temperature of their local space in real time; smoke concentration sensors collect the light attenuation rate or particulate matter concentration per cubic meter in the air; and infrared thermal imagers output a two-dimensional surface temperature thermal map within their field of view. Heterogeneous data collected by different hardware are uniformly aggregated to the edge computing center via an internal wireless LAN or wired bus. Within the edge computing center, the data is aligned according to a unified timestamp. For the two-dimensional image output by the infrared thermal imager, the highest and average temperatures of characteristic areas are extracted and combined with data from point sensors to form a continuous time series, i.e., a real-time environmental parameter sequence. In this embodiment, the sampling frequency of the time-series data from the multi-source sensors is set to 2Hz.

[0038] Step 2: Construct a computational grid model of the building space to be predicted.

[0039] Before a fire occurs or during system initialization, the Building Information Model (BIM) data of the office building is acquired. Based on the precise spatial dimensions of geometric walls, doors, windows, beams, and other components in the BIM, a voxelized mesh generation algorithm is used to divide the continuous airspace within the entire building into a discrete, regular computational mesh model composed of three-dimensional Cartesian meshes. To balance computational accuracy and model size, this embodiment sets the spatial mesh step size to 0.25 meters, dividing the entire floor into millions of tiny mesh units. Each mesh unit has a unique coordinate index in three-dimensional space. The mesh model clearly marks which units belong to solid walls, such as concrete walls and fire doors, and which units belong to airspace. During the fire's spread, the sensor data acquired in step one is automatically mapped and loaded as boundary conditions onto its nearest neighboring mesh units based on their absolute three-dimensional spatial coordinates.

[0040] Step 3: Input the real-time environmental parameter sequence into the trained physical information-guided hybrid neural network model, and infer and output the spatiotemporal evolution sequence of the fire physical field on the computational grid model in the future time period.

[0041] The core of this embodiment lies in the physical information-guided hybrid neural network model. Structurally, this model employs an advanced encoder-decoder architecture, specifically constructed as a temporal two-stream structure.

[0042] The encoder is composed of a three-dimensional spatiotemporal convolutional layer and a long short-term memory network layer. It receives a sequence of historical environmental parameters from the past thirty seconds as input, uses spatiotemporal convolution to extract spatial proximity features, and utilizes the long short-term memory network to extract nonlinear dynamic features that evolve over time. This compresses and maps high-dimensional time-series sensor data into a low-dimensional physical latent space feature vector. This physical latent space feature highly condenses the overall thermodynamic state of the current fire scene, the approximate intensity of the fire source, and the tendency of mass flow direction.

[0043] The decoder consists of multiple layers of three-dimensional transposed convolutional layers. Receiving the aforementioned physical latent space features, it reconstructs the full-field fire physical field at each time node of the entire computational grid model through layer-by-layer upsampling and reconstruction. This fire physical field specifically includes three key physical quantity fields: a three-dimensional temperature field, a three-dimensional air velocity vector field, and a three-dimensional smoke concentration field.

[0044] To ensure that the physical fields output by the neural network do not violate fundamental laws of physics, the model employs a deep physics fusion strategy during the training phase. Specifically, the model's overall loss function consists of two weighted parts: the first part is the data fitting error term, used to calculate the mean square error between the physical quantities predicted by the network and the labeled data generated by real experiments or high-precision numerical simulation software; the second part is the residual term of the core fire dynamics control equations.

[0045] In this embodiment, the fire dynamics control equations specifically include energy conservation equations, momentum conservation equations, and component transport equations. During training, the temperature, velocity, and smoke concentration output by the network at each grid cell are directly substituted into these partial differential equations. Automatic differentiation techniques are used to calculate their partial derivatives with respect to time and space, thereby determining the difference between the left and right sides of the equations. This difference is the equation residual. If the network predicts values ​​that do not conform to physical laws, the residual term will become extremely large, resulting in a strong penalty term that forces the network weights to optimize in a direction that conforms to physical laws.

[0046] The loss function has a mechanism for dynamically assigning weights to the residual terms based on the intensity of the fire in the grid region. Throughout the fire, the temperature and velocity gradients are extremely high in the fire core and flame front, resulting in intense chemical reactions and turbulence, while the physical field is relatively flat in areas far from the fire source. Therefore, this embodiment introduces an adaptive spatial weight coefficient into the loss function. This coefficient is a nonlinearly increasing function of the local temperature spatiotemporal gradient and the rate of change of smoke concentration in the grid cell. When calculating the loss, if the temperature or temperature gradient of a certain grid region is extremely high, indicating intense fire, the system will automatically assign a higher weight to the residual term of the equation corresponding to that local grid, for example, increasing it to three times that of the normal region. In regions downstream of the fire with a flat physical field, the weight of the residual term is reduced, allowing the network to focus its attention on the critical flow field region with intense fire and the strongest nonlinearity.

[0047] In actual online prediction, the model operates in parallel with a temporal dual-stream structure: one stream is based on a lightweight physics simulator pre-simplified and embedded in GPU memory, which rapidly iterates to generate a macroscopic, coarse-grained overall trend of fire spread using an extremely coarse mesh; the other stream consists of the aforementioned deep data network, which corrects local details based on real-time received multi-source sensor parameters. These two streams are coupled and mutually corrected through a latent feature fusion layer at each time step of inference. The physics simulator provides macroscopic boundary convergence guidance for the data network, while the data network dynamically compensates for the cumulative drift caused by the simplification of physical processes in the physics simulator using measured data.

[0048] Step 4: Based on the spatiotemporal evolution sequence, generate and output the indoor fire spread trend.

[0049] After the decoder outputs the spatiotemporal evolution sequence for the next five minutes, the edge computing center performs post-processing on this continuous three-dimensional full-field mesh data. Specifically, it extracts spatial isosurfaces with temperatures exceeding 60 degrees Celsius and smoke concentration isosurfaces with visibility below two meters. The trajectory of these isosurfaces expanding outward over time is superimposed on the three-dimensional spatial map of the building to generate an intuitive trend of indoor fire spread.

[0050] In this embodiment, the generated and output indoor fire spread trend also includes an uncertainty cloud map of the spread path. Since single predictions may contain random errors, this embodiment uses Monte Carlo random sampling to sample the physical latent space features output by the encoder with minor perturbations during inference, repeatedly running the decoder ten times to obtain ten different future spatiotemporal evolution sequences. Statistical analysis is performed on these ten results to calculate the probability that each grid cell will be engulfed by flames or high temperatures at a future time. The probability values ​​are divided into different confidence levels, such as a 95% high-risk certainty zone, an 80% potential spread zone, and a 5% possible impact zone. Finally, these areas with different confidence levels are visualized on the system terminal interface as cloud maps of varying color shades, providing clear risk zone boundaries for fire commanders to plan the safest rescue and evacuation routes.

[0051] Example 2

[0052] This embodiment details how the present invention achieves high-precision prediction in extreme scenarios where the sampling frequency of multi-source sensors is extremely low. Examples include situations where electromagnetic interference, bandwidth limitations, or hardware entering a low-power sleep mode are encountered. The application of this embodiment is a large underground commercial warehousing center.

[0053] Step 1: Obtain the real-time environmental parameter sequence from indoor multi-source sensors during the fire process.

[0054] In the extremely harsh environment of the current underground storage center, the IoT system forcibly resets the sensor data reporting frequency to 0.25Hz, meaning data is transmitted only once every four seconds. The multi-source sensors also consist of low-power point thermometers, gas characteristic component concentration sensors, and smoke infrared detectors deployed throughout the warehouse. Due to the extremely low sampling frequency, in the rapid initial outbreak phase of a fire, there is often a significant information gap between adjacent sampling points. If a traditional time-series prediction network is used, the lack of intermediate process information injection makes the network prone to severe prediction lag or complete underreporting of transient jumps.

[0055] Step 2: Construct a computational grid model of the building space to be predicted.

[0056] A 3D digital floor plan of the underground storage center was retrieved to establish a large-space Cartesian mesh model. Since the large-space storage area lacks fixed partitions but contains densely packed shelves, the physical space containing the shelves was defined as a porous solid obstruction in the mesh model, and a corresponding flow resistance coefficient was assigned to it. The 0.25Hz sensor sequence acquired at low frequencies was then mounted onto the corresponding mesh nodes according to the spatial coordinates of the shelf or aisle it was located on.

[0057] Step 3: Input the real-time environmental parameter sequence into the trained physical information-guided hybrid neural network model, and infer and output the spatiotemporal evolution sequence of the fire physical field on the computational grid model in the future time period.

[0058] In order to maintain a prediction similarity of over 90% at an extremely low sampling frequency of 0.25Hz, the physical information-guided hybrid neural network model in this embodiment has a unique physical constraint compensation mechanism in training and design.

[0059] During the model training phase, a rigorous learning strategy was employed to address the low-frequency input characteristic of 0.25Hz. The specific training process was as follows: First, a large-scale numerical simulation fire database was constructed. This database was obtained by running thousands of high-precision CFD fire numerical simulation software programs with different fire source locations and fire growth rates. In the pre-training phase, these simulation data were downsampled to 0.25Hz and fed as input to the hybrid neural network model. At this point, the weight coefficient of the residual term in the fire dynamics control equation of the overall loss function was set extremely high, accounting for 80% of the total loss. The control equations include the Navier-Stokes equations describing the interaction of mass and momentum, the energy equation, and the component transport equation. The aim was to enable the network, when faced with two discrete input frames spanning four seconds, to perfectly "fill in" the gaps in the physical latent space and establish prior physical knowledge of the continuous fluid evolution within those four seconds, relying entirely on the mathematical and physical intrinsic connections between the terms in the equations.

[0060] After completing the pre-training of physical prior knowledge on the simulation data, the system enters the fine-tuning stage. At this stage, a small amount of low-frequency data from real fire experiments collected in a full-scale physical fire laboratory is introduced. During fine-tuning, the weight of the residual terms in the governing equations is gradually reduced, while the weight of the error terms in the real-world data fitting is increased. Through this learning strategy, the network not only firmly grasps the fundamental laws of fluid mechanics but also learns about the nonlinear perturbations caused by the complexity of materials in the real world.

[0061] When running inference at a sampling rate of 0.25Hz in this embodiment, because a powerful physical control equation constraint network is embedded within the model, when the historical input sequence advances four seconds from the current moment and the next measured data frame arrives, the decoder within the network does not mechanically perform linear interpolation between the two points. Instead, based on the laws of conservation of momentum and energy, it automatically calculates the entire process of air acceleration, plume rise velocity, and subsequent temperature diffusion caused by thermal buoyancy within those four seconds. The physical constraints perfectly compensate for the severe lack of temporal information.

[0062] To verify the prediction accuracy, the future field prediction results output by this method were compared with the full flow field benchmark data obtained using ultra-dense grid high-precision CFD software in the same scenario. The comparison metrics used were the full-field grid spatial correlation coefficient and the structural similarity index (SSIM). Experimental data show that, due to the strong pull of physical fluid laws, the predicted similarity of the full-field temperature and flue gas concentration of this model can still be stably maintained above 92% at an ultra-low sampling rate of 0.25Hz, completely avoiding the prediction failure and numerical divergence that inevitably occur in traditional data-driven models at low-frequency inputs.

[0063] Finally, step four is executed to generate and output the indoor fire spread trend based on the spatiotemporal evolution sequence.

[0064] The continuously reconstructed three-dimensional temperature, flow rate, and smoke concentration sequence of the entire field over the next few minutes is smoothly displayed on the large screen in the monitoring center as a dynamic evolution trend chart. Even though the sensors only update once every four seconds, the fire physical field spread trend on the system interface can be smoothly displayed in real-time panoramic view of heat convection and smoke layer settling within the scene in the form of high-definition animation at 24 frames per second.

[0065] Example 3

[0066] This embodiment details a method for predicting the spread of indoor fires in a high-rise modern steel-structure office building. Because the building's main structure is a steel frame load-bearing structure, the high temperatures during a fire can easily cause a sharp decrease in the yield strength of the steel, potentially leading to the collapse of the entire structure. Therefore, this embodiment focuses on demonstrating how to simultaneously couple flow field prediction with overall structural safety assessment.

[0067] Step 1: Obtain the real-time environmental parameter sequence from indoor multi-source sensors during the fire process.

[0068] In key fire-prevention areas of this high-rise steel structure building, in addition to point-type thermometers and smoke photoelectric sensors, high-temperature thermocouple sensors are densely attached to the exposed steel beams and columns, and an array-type infrared thermal imager is integrated into the circuitry. When a fire breaks out on a certain floor, the multi-source sensor network simultaneously captures the air temperature, smoke particle concentration, and local infrared radiation thermal images of the main load-bearing steel components at a sampling frequency of 1Hz.

[0069] Step 2: Construct a computational grid model of the building space to be predicted.

[0070] When constructing the computational mesh model, not only was the indoor air circulation area meshed in three dimensions using voxelization, but also all load-bearing steel columns, I-beams, profiled steel sheet composite floor slabs, and other solid structural components, whether hidden inside the walls or exposed, were meshed in three dimensions according to the building's structural construction drawings. The air mesh and the steel structure solid mesh achieved perfect geometric alignment at the interface to facilitate numerical transfer of conjugate heat between the fluid and solid layers.

[0071] Step 3: Input the real-time environmental parameter sequence into the trained physical information-guided hybrid neural network model, and infer and output the spatiotemporal evolution sequence of the fire physical field on the computational grid model in the future time period.

[0072] To enable simultaneous assessment of structural collapse risk, this embodiment significantly extends the fire dynamics control equations in the physical information-guided hybrid neural network model. In addition to including energy conservation equations, momentum conservation equations, and component transport equations to constrain the airflow field, a structural heat transfer equation specifically describing the temperature rise process of the solid steel structure is introduced. This structural heat transfer equation, based on Fourier's law of thermal conductivity, considers the heat flux density input to the steel structure surface by fire smoke through convective and radiative heat transfer, as well as the anisotropic diffusion of heat within the steel components along the isotropic thermal conductivity.

[0073] During model training, the residual terms in the loss function include the residuals of this structural heat transfer equation. The network must not only ensure that the predicted air temperature field conforms to fluid dynamics, but also that the predicted temperature distribution within the solid mesh of the steel structure fully conforms to the laws of solid thermal conduction and the thermal equilibrium conditions of the fluid-solid interface.

[0074] During online inference and prediction, because the network has thoroughly grasped the laws of fluid-structure interaction and heat conduction, its decoder outputs a spatiotemporal evolution sequence of the fire physical field with dual characteristics when outputting the spatiotemporal evolution sequence for future time periods. This sequence not only includes the three-dimensional temperature field, smoke velocity field, and smoke concentration field of the air domain within the next five minutes, but also the dynamic three-dimensional temperature distribution evolution sequence within the entire building's load-bearing framework system—all load-bearing steel beams and columns. The system can clearly capture the detailed process of how the internal temperature of a core main column gradually rises from room temperature to 500 degrees Celsius and then 600 degrees Celsius due to the intense heat of the fire.

[0075] Step 4: Based on the spatiotemporal evolution sequence, generate and output the indoor fire spread trend.

[0076] In this embodiment, the system incorporates a mathematical analytical function that describes the degradation of standard steel's mechanical properties with temperature, based on the temperature distribution of structural components within the output fire physical field spatiotemporal evolution sequence. As the temperature exceeds 400 degrees Celsius, the steel's yield strength decreases exponentially by a specific coefficient. When it is predicted that the average temperature of a critical load-bearing column will reach 600 degrees Celsius within the next three minutes, thus losing its load-bearing capacity, the system immediately activates the structural damage calculation module. This module not only outputs the indoor fire spread trend path of air and smoke but also simultaneously outputs the structural collapse risk trend in a 3D visualization scene. During output, the system directly highlights steel components with a very high probability of yielding failure, loss of support, and subsequent chain collapse of upper floors in a striking deep red flash on the screen, along with a countdown timer indicating the expected collapse. Simultaneously, in conjunction with the uncertainty cloud map of the output fire spread path, it clearly marks the extreme high-risk red line zone where the building interior is absolutely inaccessible due to the spread of collapse risk. This significantly improves the scientific nature of fire command in high-rise building fires and ensures the safety of firefighters.

[0077] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the content of the present invention's specification shall also be included within the scope of protection of the present invention.

Claims

1. A method for predicting the spread trend of indoor fires, characterized in that, include: Acquire real-time environmental parameter sequences from indoor multi-source sensors during a fire; Construct a computational grid model of the building space to be predicted; The real-time environmental parameter sequence is input into a trained physical information-guided hybrid neural network model, which infers and outputs the spatiotemporal evolution sequence of the fire physical field on the computational grid model in future time periods. The fire physical field includes a temperature field, a velocity field, and a smoke concentration field; The loss function of the physical information-guided hybrid neural network model during training includes at least the data fitting error term and the residual term of the fire dynamics control equation, and the loss function dynamically assigns weights to the residual term according to the intensity of the fire in the grid area. Based on the spatiotemporal evolution sequence, the indoor fire spread trend is generated and output.

2. The method for predicting the spread trend of indoor fires according to claim 1, characterized in that, The multi-source sensor includes a temperature sensor, a flue gas concentration sensor, and an infrared thermal imager.

3. The method for predicting the spread trend of indoor fires according to claim 1, characterized in that, The physical information-guided hybrid neural network model adopts an encoder-decoder structure, in which the encoder maps the sequence of historical environmental parameters into physical latent space features, and the decoder reconstructs the future fire physical field from the physical latent space features.

4. The method for predicting the spread trend of indoor fires according to claim 1, characterized in that, The sampling frequency of the real-time environmental parameter sequence is from 0.2 Hz to 5 Hz.

5. The method for predicting the spread trend of an indoor fire according to claim 4, characterized in that, The sampling frequency is 0.25Hz; at this sampling frequency, the physical information guides the hybrid neural network model to compensate for the lack of temporal information through physical constraints, achieving a prediction similarity of over 90%.

6. The method for predicting the spread trend of indoor fires according to claim 1, characterized in that, The fire dynamics control equations include the energy conservation equation, the momentum conservation equation, and the component transport equation.

7. The method for predicting the spread trend of an indoor fire according to claim 6, characterized in that, The fire dynamics control equations also include structural heat transfer equations describing the temperature rise of the steel structure, and the output fire physical field spatiotemporal evolution sequence also includes the temperature distribution of structural components to synchronously output the structural collapse risk trend.

8. The method for predicting the spread trend of indoor fires according to claim 1, characterized in that, The training of the physical information-guided hybrid neural network model adopts a course learning strategy, which first pre-trains the physical prior knowledge on simulation data, and then fine-tunes it using real fire experimental data.

9. The method for predicting the spread trend of an indoor fire according to claim 1, characterized in that, The physical information-guided hybrid neural network model is constructed as a temporal dual-stream structure: one stream is based on a physical simulator to quickly infer macroscopic trends, and the other stream is based on a data network to correct local details. The two streams are coupled and corrected to each other during the inference process.

10. The method for predicting the spread trend of an indoor fire according to claim 1, characterized in that, The generated and output indoor fire spread trend also includes an uncertainty cloud map of the spread path, providing risk areas at different confidence levels.

Images