A fire spread second-level prediction method and device based on multi-source data

Abstract

Embodiments of the present application provide a kind of fire field fire spread second-level prediction method and device based on multi-source data, by image acquisition equipment and wind field sensor real-time acquisition fire field visual data and wind speed and direction data, constitute multi-source information foundation;Then based on visual data analysis out combustible type, distribution range and current combustion area boundary, to extract directly related with fire spread combustible feature and fire state feature;Subsequently with unified geographic coordinate system as benchmark, the above-mentioned feature with real-time wind field data is spatially aligned, normalized and fused into standardization feature matrix;Finally the feature matrix is input into the lightweight fire spread prediction model pre-trained, and the direction, rate and burning boundary expansion range of future fire spread are output by model inference once.Not dependent on cloud transmission, can be completed from data acquisition to prediction result generation whole link operation with extremely low delay in edge end.

Description

Technical Field

[0001] This invention relates to the field of fire emergency monitoring and intelligent prediction technology, and in particular to a method and device for predicting the spread of fire in a fire scene in seconds based on multi-source data. Background Technology

[0002] Real-time prediction of fire spread is crucial for emergency response command, personnel evacuation, and firefighting force deployment. Currently, numerical simulation methods based on physical models are one of the mainstream technical approaches in the industry for achieving high prediction accuracy. For example, Fire Dynamics Simulators (FDS) can simulate the fire development process with relatively high precision by constructing a three-dimensional mesh and solving partial differential equations related to combustion, heat transfer, and fluid dynamics. While these methods can theoretically provide high-precision prediction results, their core algorithms result in enormous computational and memory requirements, with a complete simulation typically taking several minutes or even tens of minutes to complete. Due to this excessive computational delay, such solutions struggle to meet the stringent real-time requirements of emergency response at fire scenes.

[0003] Besides numerical simulation, existing technologies include empirical formula prediction and intelligent prediction based on vision or big data. Empirical formula methods rely only on a few macroscopic parameters such as wind speed and combustible material type for rough estimation, resulting in extremely low prediction accuracy and an inability to adapt to the non-uniform characteristics of complex fire scenes. Vision-based monitoring schemes are limited to open flame identification and fire alarms, failing to deeply analyze key information such as combustible material properties and spatial distribution contained in images, thus unable to output quantitative fire spread trends. A few prediction schemes that attempt to integrate historical fire big data typically use coarse-grained scene matching (such as "forest fire" or "building fire"), failing to deeply integrate the refined combustible material characteristics analyzed in real time with real-time wind field data. This leads to poor adaptability between historical models and the current scene, and limited confidence in the prediction results.

[0004] In summary, existing technologies suffer from the following inherent limitations: First, physical numerical simulation schemes, such as FDS, have excessively high computational complexity, with single predictions taking several minutes to tens of minutes, making them completely incapable of meeting the emergency real-time requirement of outputting fire spread results within seconds for the next 1-2 minutes. Second, existing technologies suffer from insufficient data fusion, often relying on a single data source (such as meteorological data or visual fire data), and generally lacking real-time analysis and fusion of refined combustible properties and their high-precision spatial distribution, resulting in significant deviations between prediction results and actual fire conditions. Third, existing big data prediction schemes exhibit low accuracy in matching historical models, failing to adaptively calibrate models based on the details of combustible distribution analyzed visually at the scene. This coarse-grained scene adaptation further exacerbates computational delays and prediction errors, failing to provide timely and effective decision support for on-site rescue. Summary of the Invention

[0005] This invention provides a method, device, and storage medium for predicting fire spread in seconds based on multi-source data, in order to solve the technical problems in existing fire prediction schemes, such as long calculation time that cannot meet the second-level real-time response, insufficient fusion of multi-source data leading to large deviations between prediction results and actual situation, and poor adaptability of historical models to field scenarios that further aggravates prediction errors.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting the spread of fire in a fire scene within seconds based on multi-source data, including: S1. Real-time acquisition of multi-source data from the fire scene, including visual data of the fire scene and real-time wind field data; S2. Extract combustible material features and fire status features based on the fire scene visual data. The combustible material features include the type of combustible material and the distribution range of combustible material. The fire status features include the current burning area boundary. S3. Using a unified geographic coordinate system as a reference, the characteristics of combustibles, the characteristics of fire status, and the real-time wind field data are spatiotemporally matched and normalized to generate a fused feature matrix. S4. Input the fused feature matrix into the pre-trained lightweight fire spread prediction model and output the fire spread prediction result, which includes the fire spread direction, spread rate and combustion boundary expansion range.

[0007] Preferably, S1 specifically includes: Visual data of the fire scene is collected by fixed surveillance cameras, inspection drones or individual law enforcement recorders, and the visual data of the fire scene includes fire scene video streams or static images. Real-time wind field data is collected by a portable weather station or a weather sensor mounted on a drone. The real-time wind field data includes real-time wind speed and wind direction data. The frame rate for acquiring the visual data of the fire scene shall not be less than 2fps and the resolution shall not be less than 1080P; the acquisition frequency for the real-time wind field data shall not be less than 1Hz.

[0008] Preferably, S2 specifically includes: S21. Perform preprocessing on the key frames of the fire scene visual data, including defogging, noise reduction, and contrast enhancement. S22. Input the preprocessed keyframes into the trained lightweight target detection model, and output the identification results of the combustible material type and the combustible material distribution range through the lightweight target detection model; S23. Input the preprocessed keyframes into the trained lightweight semantic segmentation model, and the lightweight semantic segmentation model outputs the segmentation results of the open flame core area, smoldering area and burned area to determine the boundary of the current burning area.

[0009] Preferably, S3 specifically includes: S31. Using a preset geographic coordinate system as a unified reference, convert the pixel coordinates of the combustible material distribution range and the current combustion area boundary into geographic spatial coordinates, and convert the real-time wind field data into wind field vector data within the same geographic space. S32. Construct a fusion feature matrix, which includes features in the following dimensions: the type of combustible material and its corresponding inherent combustion parameters, the spatial distribution coordinates of the combustible material, the wind field vector data, the coordinates of the current combustion zone boundary and center, and the historical fire spread feature parameters of the same scenario; wherein, the inherent combustion parameters include calorific value, standard spread rate, and ignition point; S33. After normalizing all feature dimensions in the fused feature matrix to the same numerical range, feature splicing is used for fusion to generate a standardized fused feature matrix.

[0010] Preferably, the lightweight fire spread prediction model is pre-trained based on a historical fire database using a combination of physical model constraints and data-driven methods, and then deployed on edge computing nodes after lightweight processing. The embedding methods for the physical model constraints include: Standardize and map key parameters in the classic physical laws of fire spread to multi-source data characteristics; A sparse mask is set in the self-attention mechanism of the lightweight fire spread prediction model. The sparse mask is based on the classical fire spread physics law. The sparse mask is used to shield invalid calculation areas and strengthen the focus on high-risk areas. Add a physical constraint loss term to the loss function of the lightweight fire spread prediction model during training; The lightweighting process includes: Model pruning is performed based on the contribution of each channel to the physical constraint loss; Quantize the 32-bit floating-point weight parameter into an 8-bit integer; The lightweight fire spread prediction model is trained by a teacher model that incorporates physical constraints, so that the lightweight fire spread prediction model fits the output of the teacher model, the attention weight distribution, and the physical prior features.

[0011] Preferably, in step S4, after the fused feature matrix is ​​input into the lightweight fire spread prediction model, the lightweight fire spread prediction model outputs fire spread prediction results for the next 1 minute and the next 2 minutes, and the fire spread prediction results also include high-risk spread areas. After outputting the fire spread prediction results, dynamic correction is also included: The actual combustion change data of the current video frame and the previous video frame are obtained, and the fire spread prediction result is dynamically corrected based on the actual combustion change data; wherein, the actual combustion change data includes the positional change of the boundary of the combustion area.

[0012] Preferably, the edge computing node is a locally deployed independent computing device, and the edge computing node is disconnected from the public network.

[0013] Secondly, embodiments of the present invention provide a fire spread prediction device based on multi-source data with a second-level prediction capability, comprising: The data acquisition module is used to collect multi-source data from the fire scene in real time, including visual data of the fire scene and real-time wind field data. The feature extraction module is used to extract combustible features and fire status features based on the fire scene visual data. The combustible features include the type of combustible and the distribution range of the combustible, and the fire status features include the current burning area boundary. The spatiotemporal fusion module is used to perform spatiotemporal matching and normalization feature fusion of the combustible material characteristics, the fire state characteristics and the real-time wind field data based on a unified geographic coordinate system, and generate a fusion feature matrix. The prediction and inference module is used to input the fused feature matrix into a pre-trained lightweight fire spread prediction model and output the fire spread prediction results, which include the fire spread direction, spread rate and combustion boundary expansion range.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the second-level prediction method for fire spread based on multi-source data as described in the first aspect of the present invention.

[0015] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the second-level prediction method for fire spread based on multi-source data as described in the first aspect of the present invention.

[0016] This invention provides a method and apparatus for second-level prediction of fire spread based on multi-source data. Addressing the problems of excessive computation time, insufficient data fusion, low prediction accuracy, and inability to meet the real-time emergency requirements of fire scenes in existing fire spread prediction technologies, this invention acquires real-time visual data of the fire scene and wind speed and direction data using image acquisition equipment and wind field sensors, forming a multi-source information foundation. Then, based on the visual data, the type and distribution range of combustibles, as well as the current boundary of the burning area, are analyzed to extract combustible characteristics and fire state characteristics directly related to fire spread. Subsequently, using a unified geographical coordinate system as a reference, the above features are spatially aligned and normalized with real-time wind field data and fused into a standardized feature matrix. Finally, this feature matrix is ​​input into a pre-trained lightweight fire spread prediction model, which infers and outputs the future direction, rate, and burning boundary expansion range of the fire in a single operation. It does not rely on cloud transmission and can complete the entire process from data acquisition to prediction result generation at the edge with extremely low latency. Compared with existing technologies, it has the following advantages: (1) By integrating the fine properties of combustibles, dynamic wind field and real-time fire status, the matching degree between the prediction results and the actual development of the fire is significantly improved, avoiding the deviation caused by a single data source.

[0017] (2) By adopting a lightweight model and performing compact feature expression under a unified spatiotemporal reference, the total time for a single prediction is compressed to within seconds, which completely solves the problem of computation delay of several minutes or even tens of minutes in traditional physical numerical simulation schemes, and truly meets the rigid requirements of fire emergency command for real-time performance.

[0018] (3) The entire prediction process can be completed independently on the edge computing node without connecting to the public network. It is suitable for various complex fire environments and provides reliable and timely quantitative decision support for rescuers to quickly formulate evacuation routes and deploy firefighting forces. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for predicting the spread of fire in a fire scene in seconds based on multi-source data, provided in an embodiment of the present invention. Figure 2 A schematic diagram of the lightweight fire spread prediction model architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a fire spread prediction device based on multi-source data provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a physical structure provided for an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Figure 1 This is a flowchart of a method for predicting the spread of fire in a fire scene in seconds based on multi-source data according to an embodiment of the present invention. (Refer to...) Figure 1 The method includes: S1. Real-time acquisition of multi-source data from the fire scene, including visual data and real-time wind field data. Visual data comprises images or videos reflecting the distribution of open flames, smoke, and combustibles, providing raw information on the appearance of combustibles and the morphology of the fire for subsequent analysis. Real-time wind field data provides instantaneous wind direction and speed data for the fire area, crucial environmental parameters influencing the direction and rate of fire spread. Both visual and real-time wind field data serve as fundamental inputs for fire spread prediction. In this embodiment, S1 involves simultaneously acquiring external visual conditions and meteorological environmental data from the fire scene. Addressing the shortcomings of existing technologies that rely on a single data source and incomplete input information, physical simulation schemes depend on offline meteorological parameters or rough estimates, while visual schemes completely ignore wind fields, resulting in predictions that lag significantly behind on-site changes. This step S1, by simultaneously acquiring both visual and wind field data, provides complete foundational information for subsequent predictions, solving the problem of insufficient prediction input dimensions. This ensures that subsequent analysis is supported by the real-world environment, improving the completeness of the prediction basis.

[0023] S2. Based on the visual data of the fire scene, extract combustible material features and fire status features. The combustible material features include the type of combustible material and the distribution range of combustible material. The fire status features include the current boundary of the burning area.

[0024] Among them, combustible material characteristics characterize the properties and spatial information of combustible substances in the fire scene; combustible material types distinguish the combustion characteristics of substances, referring to different materials present in the fire scene (such as wood, plastics, vegetation, chemical raw materials, etc.); combustible material distribution range refers to the spatial area occupied by various combustible materials; fire status characteristics reflect the current combustion status of the fire, and the current combustion area boundary defines the boundary between burned and unburned areas. This step S2 analyzes the combustion carrier and fire status information from visual data. Addressing the shortcomings of existing visual monitoring that only identifies open flames and does not analyze combustible material-related information, this step extracts combustible material and fire status characteristics to solve the problem of not being able to accurately quantify the combustion carrier and combustion range. This clarifies the basic conditions for fire spread and provides key combustion-related features for subsequent fusion calculations.

[0025] S3. Using a unified geographic coordinate system as a reference, the combustible material characteristics, fire status characteristics, and real-time wind field data are spatiotemporally matched and normalized to generate a fused feature matrix. The unified geographic coordinate system provides a unified spatial reference for all scene data, used to convert pixel coordinates into real geographic coordinates. Spatiotemporal matching aligns data from different sources at the same spatial location and time point. Normalized feature fusion transforms feature values ​​of different types and dimensions to the same numerical range and combines them into a multi-dimensional matrix. At a fire scene, the combustible material distribution obtained from visual analysis is in pixel coordinates, while the wind field data is in geographic vectors. Since their original reference systems are different, direct splicing will prevent the model from understanding the physical correspondence. In existing technologies, most solutions simply list data without strict spatial alignment and dimension normalization, making it difficult for the model to learn the correct spread patterns.

[0026] In this embodiment, S31 uses the CGCS2000 geographic coordinate system as a unified spatial reference. The affine projection formula for converting pixel coordinates to geographic spatial coordinates is as follows: (1) In formula (1): ( ) represents the geographic coordinates of the image reference point. u , v () represents pixel coordinates; () represents the pixel coordinates of the image reference point; Pixel-geographic scale factor; The attitude angle of the data acquisition equipment is determined through coordinate transformation. This ensures that every distribution point of the combustible material and every position on the combustion boundary can be precisely correlated with the wind vector at the corresponding spatial location, providing spatial consistency for subsequent data fusion.

[0027] S33 normalizes all feature dimensions in the fused feature matrix to the same numerical range using Min-Max normalization, with the following formula: (2) In equation (2): These are the original eigenvalues; The minimum / maximum value of the feature dimension; These are the normalized feature values. The normalization operation avoids slow convergence or bias caused by differences in feature scale during model training. After feature concatenation, a standardized fused feature matrix is ​​generated. This step takes less than 300ms.

[0028] S4. Input the fused feature matrix into the pre-trained lightweight fire spread prediction model and output the fire spread prediction result, which includes the fire spread direction, spread rate and combustion boundary expansion range.

[0029] The lightweight fire spread prediction model is a simplified and optimized fire prediction model that can be calculated quickly. The fire spread direction is the dominant direction of fire diffusion, the spread rate is the fire spread speed, and the combustion boundary expansion range is the spatial extent of the future burning area. In emergency command, decision-makers need to obtain the fire's trajectory for the next 1-2 minutes within seconds. However, traditional numerical models such as FDS (Fire Dynamics Simulator) require several minutes to tens of minutes for a single calculation, completely failing to meet real-time requirements. In this embodiment, S4 utilizes standardized data and a lightweight model to quickly output prediction results. Addressing the shortcomings of existing physical models—large computational load, long processing time, and inability to meet emergency real-time requirements—the lightweight model enables rapid inference, solving the problems of delayed prediction calculations and the inability to output results in a short period. It can quickly provide a quantitative fire spread situation, offering timely and effective decision support for emergency fire response.

[0030] Based on the above embodiments, as a preferred implementation, step S1 specifically includes: Visual data of the fire scene is collected through fixed surveillance cameras, patrol drones, or individual law enforcement recorders. This visual data includes video streams or still images of the fire scene. These devices are all general-purpose equipment at fire emergency sites and are easy to deploy quickly.

[0031] Real-time wind field data, including real-time wind speed and direction, is collected using portable weather stations or weather sensors mounted on drones. Drones, in particular, allow for flexible acquisition of wind field information from areas above fires or inaccessible regions.

[0032] The frame rate of the fire scene visual data acquisition is no less than 2fps and the resolution is no less than 1080P, ensuring the clarity and temporal continuity required for subsequent image analysis; the acquisition frequency of the real-time wind field data is no less than 1Hz, ensuring that the dynamic changes of the wind field can be captured.

[0033] It should be noted that all collected data is transmitted to the edge computing node in real time via 5G or industrial wireless LAN to avoid using public network links that may be interrupted. At the same time, a standardized historical fire database is pre-built and loaded into the storage unit of the edge computing node. This database contains fire spread models, spread rate parameters and combustion law data under different combustible material types, different wind field conditions and different scene categories (such as buildings, forests and chemical industrial parks), and pre-loads matching lightweight model weights according to the type of on-site scene.

[0034] Based on the above embodiments, as a preferred implementation, step S2 specifically includes: S21. Perform dehazing, denoising, and contrast enhancement preprocessing on the key frames of the fire scene visual data; wherein, a key frame refers to an image frame extracted from the video stream at preset time intervals (e.g., every 500ms), or directly using a captured still photograph, as the reference unit for subsequent analysis. The dehazing, denoising, and contrast enhancement preprocessing aims to overcome the adverse imaging conditions such as smoke obstruction and insufficient light at the fire scene; specifically, an adaptive histogram equalization combined with a guided filtering dehazing algorithm is adopted, wherein the adaptive histogram equalization is used to stretch the local contrast of the image, and the guided filtering dehazing is used to estimate and remove the influence of smoke. The two work together to improve image clarity, and the processing time per frame is controlled within 500ms.

[0035] S22. Input the preprocessed keyframes into the trained lightweight object detection model, and output the identification results of the combustible material type and the distribution range of the combustible material through the lightweight object detection model; wherein, the lightweight object detection model is specifically the YOLOv8n (You Only Look Once version 8 nano) model, which can also be replaced by lightweight models such as SSD-MobileNet or EfficientNet-Lite; the lightweight object detection model infers on the preprocessed keyframes and outputs the combustible material type (such as wood structure, plastic, vegetation, chemical raw materials, building insulation materials, etc.) and its distribution range (i.e., the area occupied and spatial position of each combustible material in the image), and can also match the pre-stored combustible material attribute library to output the inherent combustion parameters of various combustible materials (including calorific value, standard spread rate, ignition point, etc.).

[0036] S23. Input the preprocessed keyframes into the trained lightweight semantic segmentation model, and the lightweight semantic segmentation model outputs the segmentation results of the open flame core area, smoldering area and burned area to determine the boundary of the current burning area.

[0037] Specifically, the lightweight semantic segmentation model is the BiSeNetV2 (Bilateral Segmentation Network version 2) model, which performs pixel-level classification on the same preprocessed keyframe, dividing each pixel in the image into an open flame core area (intensely burning area), a smoldering area (area without open flame but with smoke or smoldering fire), or a burned area (overburned area), thereby accurately delineating the boundary of the current burning area. It is processed in parallel with combustible material detection, and the overall processing time is controlled within 800ms.

[0038] In this embodiment, the raw visual data is transformed into structured combustible features (type, distribution, combustion parameters) and fire state features (combustion boundary, location of each zone), which solves the technical defect in the prior art that it cannot provide fine combustible attributes for prediction due to the lack of in-depth analysis of image content, and provides a high-confidence input for subsequent multi-source fusion.

[0039] Based on the above embodiments, as a preferred implementation, step S3 specifically includes: S31. Using a preset geographic coordinate system as a unified reference, convert the pixel coordinates of the combustible material distribution range and the current combustion area boundary into geographic spatial coordinates, and convert the real-time wind field data into wind field vector data within the same geographic space.

[0040] Specifically, in this embodiment, the CGCS2000 geographic coordinate system is selected as a unified spatial reference. Pixel coordinates refer to the row and column positions of the combustible material distribution range and the combustion zone boundary in the original image, while geographic spatial coordinates refer to the longitude, latitude, or projected coordinates converted to the real world. Wind field vector data refers to a vector with spatial orientation formed by combining wind speed and wind direction. Initially, the combustible material and combustion boundary obtained from visual analysis are pixel coordinates, while the wind field data is a geographic vector, and the two cannot be directly correlated. Through coordinate transformation, every distribution point of the combustible material and every position on the combustion boundary can be accurately correlated with the corresponding spatial wind vector, providing spatial consistency for subsequent fusion.

[0041] S32. Construct a fusion feature matrix, which includes features in the following dimensions: the type of combustible material and its corresponding inherent combustion parameters, the spatial distribution coordinates of the combustible material, the wind field vector data, the boundary and center coordinates of the current combustion area, and the historical fire spread feature parameters of the same scenario; wherein, the inherent combustion parameters include calorific value, standard spread rate and ignition point.

[0042] Specifically, the fusion feature matrix is ​​a standardized data structure containing multi-dimensional features: in addition to the type of combustible material and its inherent combustion parameters (calorific value, standard spread rate, and ignition point, obtained by matching from a pre-stored attribute library), the spatial distribution coordinates of the combustible material, wind field vector data, and the coordinates of the current combustion zone boundary and center, it also introduces historical fire spread feature parameters from the same scenario. These parameters are extracted from the historical fire database pre-loaded on the edge computing nodes according to the current scenario type (such as buildings, forests, and chemical industrial parks) to provide statistical priors. The setting of these dimensions enables the model to simultaneously utilize real-time on-site information (combustible material, wind field, fire boundary) and historical statistical patterns (spread features from the same scenario), significantly improving the targeting and accuracy of predictions.

[0043] S33. After normalizing all feature dimensions in the fused feature matrix to the same numerical range, feature splicing is used for fusion to generate a standardized fused feature matrix.

[0044] Specifically, the normalization operation maps features with different dimensions and numerical ranges (for example, the calorific value of combustibles may be as high as tens of megajoules per kilogram, while the wind field vector value is between 0 and 30 meters per second) to the same numerical range (such as [0,1]), avoiding slow convergence or bias due to differences in feature scale during model training; feature concatenation refers to concatenating all normalized feature vectors end to end in sequence to form a long total feature vector, i.e., a fused feature matrix. This step takes less than 300ms to process.

[0045] The embodiments of the present invention achieve accurate alignment and compact representation of multi-source heterogeneous data within the same spatiotemporal framework, solving the technical defects in the prior art that make it difficult for subsequent models to learn effectively or have large prediction deviations due to missing data dimensions, insufficient coordinate system, and chaotic feature dimensions. It provides high-quality standardized input for the second-level inference of lightweight models.

[0046] Based on the above embodiments, as a preferred implementation method, such as Figure 2 As shown, the lightweight fire spread prediction model is based on a historical fire database and pre-trained using a combination of physical model constraints and data-driven methods. After lightweight processing, it is deployed on edge computing nodes. The historical fire database refers to a standardized dataset pre-stored on the edge computing nodes, containing fire spread models, spread rate parameters, and combustion law data under different combustible material types, wind field conditions, and scene categories (buildings, forests, chemical industrial parks). Physical model constraints refer to incorporating the mechanisms of classical fire spread physics (such as the Rothermel model) into the model training to avoid purely data-driven predictions that violate common sense; data-driven methods utilize real fire samples to fit statistical patterns.

[0047] The embedding methods for the physical model constraints include: This involves standardizing and mapping key parameters from the classical physical laws of fire spread to multi-source data characteristics. Specifically, the core formula for the spread rate in the Rothermel model is: ROS = R 0× K w × K s × K m ,in, ROS ( Rate of Spread ( ) represents the rate of fire spread. R 0 represents the baseline spread rate of combustible materials. K wThis is the wind field correction factor. K s This is the slope correction factor (which can be simplified in building / flat terrain scenarios). K m This is the humidity damping coefficient. These key parameters are normalized and mapped to features from multi-source data (combustible material properties, real-time wind field, scene terrain), for example, R 0 corresponds to the calorific value and ignition point of the combustible material. K w Corresponding to real-time wind speed and direction, K s and K m The corresponding scene terrain and environmental humidity are normalized to the [0,1] interval and then embedded into the model input layer, which is consistent with the input feature dimension of the Transformer model, so as to achieve feature adaptation without additional computational overhead and lay the foundation for subsequent fusion.

[0048] A sparse mask is set in the self-attention mechanism of the lightweight fire spread prediction model. The sparse mask is based on the classic fire spread physics law and is used to shield invalid calculation areas and strengthen the focus on high-risk areas. In terms of structural embedding and fusion, the Rothermel physics law is deeply embedded in the Transformer model architecture. A dual-branch input encoding layer is designed to achieve parallel processing. The data-driven branch converts multi-source features into a 64-dimensional lightweight token sequence through a lightweight linear projection layer, controlling the number of parameters from the source. The physical prior branch is a parameterless calculation module that calculates the fire physical baseline spread rate, spread direction baseline value, and high-risk area weight in real time based on the mapped physical parameters, outputs the physical prior feature sequence, and completes position encoding to ensure spatiotemporal matching. The two features are concatenated and input into the Encoder layer. At the same time, the self-attention mechanism is improved by setting a sparse mask based on the Rothermel law. Token pairs in upwind and non-combustible areas are shielded from invalid calculations, and high-risk areas are strengthened. A physical anchoring module is added to the decoding layer to force the spread rate and direction output by the model to be within the reasonable range of the Rothermel model, avoiding violation of physical common sense.

[0049] A physical constraint loss term is added to the loss function during the training of the lightweight fire spread prediction model. Regarding loss function fusion, a multi-objective total loss function is constructed. L total = L data + λ × L physics + L sparse The Rothermel physical constraint depth is incorporated into the training process. Ldata The data fitting loss (including IoU loss and mean squared error loss) is used to fit real fire samples; L physics The physical constraint loss includes baseline error loss, physical consistency penalty loss (e.g., the downwind spread rate must be greater than the upwind rate), and historical model adaptation loss. L sparse This provides a basis for L1 regularization loss and lightweight design. The physical loss weights are adjusted during the pre-training phase. λ Set it to 0.6, and set it to 0.25 in the fine-tuning stage to balance the physical laws and the accuracy of data fitting.

[0050] The physical constraint loss formula is: (3) In the above formula (3), This refers to the rate of fire spread predicted by the model. The baseline spread rate is calculated based on the Rothermel physical model. v 逆风 The rate of spread in the upwind direction, v 顺风 The propagation rate is given by the downwind direction; the combustion boundary IoU loss formula is: (4) In the above formula (4), The area (or set of pixels) enclosed by the boundary of the combustion zone predicted by the model. The area enclosed by the actual combustion zone boundary (Ground Truth, i.e., the actual observed combustion zone).

[0051] In the pre-training phase, the physical loss weight λ is set to 0.6, and in the fine-tuning phase it is set to 0.25, in order to balance the physical laws and the data fitting accuracy.

[0052] The lightweighting process includes: Model pruning is performed based on the contribution of each channel to the physical constraint loss.

[0053] The lightweight fire spread prediction model is trained by a teacher model that incorporates physical constraints, so that the lightweight fire spread prediction model fits the output of the teacher model, the attention weight distribution, and the physical prior features.

[0054] Specifically, in this embodiment of the invention, the number of parameters is compressed to less than 50M through a coordinated operation of "pruning + quantification + distillation": ①Physical contribution-oriented pruning: Calculate the contribution of each channel to the physical constraint loss, retain the core channels, prune redundant channels, and compress the number of parameters from the initial 80M to less than 60M without reducing the accuracy by more than 1%.

[0055] ②INT8 parameter quantization: The 32-bit floating-point weights and activation values ​​of the model are quantized into 8-bit integers, and the Rothermel physics calculation parameters are fixed-point quantized, with the accuracy loss controlled within 0.5%, further compressing the model size.

[0056] ③ Physics-aware knowledge distillation: Using a 200M parameter Transformer model that incorporates complete physical constraints as the teacher model, the lightweight student model not only fits the output of the teacher model but also its attention weight distribution and prior physical features, ensuring that the ability to meet physical constraints is not lost after lightweighting. Ultimately, the number of model parameters is stably controlled within 50M. After pre-training, it is deployed on edge computing nodes, supporting purely local edge inference without relying on cloud computing power.

[0057] Through the deep integration of the above physical laws and systematic lightweight processing, the model achieves three objectives at the edge: single inference time of less than 1000 milliseconds, prediction results that conform to physical common sense, and parameter quantity of less than 50M. It solves the core technical problems of excessive physical numerical simulation computation, lack of physical consistency of pure data-driven models, and inability to deploy large models at the edge in existing technologies.

[0058] Based on the above embodiments, as a preferred implementation, in step S4, after inputting the fused feature matrix into the lightweight fire spread prediction model, the lightweight fire spread prediction model outputs fire spread prediction results for the next 1 minute and the next 2 minutes. The fire spread prediction results also include high-risk spread areas. The next 1 minute and the next 2 minutes correspond to the most urgently needed recent fire trends in the emergency response. High-risk spread areas refer to areas predicted by the model as having a high probability of fire expansion, rapid spread, or posing a direct threat to critical facilities / personnel. These areas are generated by strengthening focus on high-risk areas through sparse masking. In specific implementation, after the fused feature matrix is ​​input into the model, the lightweight fire spread prediction model (which has undergone the aforementioned physical constraint embedding and collaborative lightweighting processing) outputs prediction results for both the next 1 minute and the next 2 minutes in a single forward inference, including the dominant fire spread direction, spread rate, combustion boundary expansion range, and high-risk spread areas. Unlike traditional numerical simulations that require multiple iterations or step-by-step extrapolation, this model uses a direct mapping method, significantly reducing latency.

[0059] After outputting the fire spread prediction results, dynamic correction is also included: The actual combustion change data of the current video frame and the previous video frame are obtained, and the fire spread prediction result is dynamically corrected based on the actual combustion change data; wherein, the actual combustion change data includes the positional change of the boundary of the combustion area.

[0060] Formula for actual change in combustion boundary: (5) In equation (5): The geospatial coordinates of the boundary of the burning area in the current video frame (which can be represented as the position vector of the boundary point or the regional features enclosed by the boundary). The geographic coordinates of the combustion boundary in the previous frame. This represents the change in geographic coordinates of the combustion boundary.

[0061] Prediction rate correction formula: (6) In equation (6): This is the dynamically corrected fire spread rate; The actual observed change in combustion boundary / the observed change in combustion boundary; The original predicted spread rate is output by the lightweight fire spread prediction model.

[0062] Specifically, dynamic correction refers to using the latest observed changes in actual fire intensity to calibrate the model's predictive output, forming a closed-loop feedback. This involves visually analyzing the positional changes of the fire boundary in two consecutive frames of images (e.g., the distance the fire front advances in a certain direction), comparing this actual change with the model's predicted change over the same time period, calculating the error, and then fine-tuning the current model's prediction. For example, if the actual spread rate is slightly lower than the model's prediction in the first second, the output for the following 1-2 minutes will correspondingly reduce the spread rate. This correction does not retrain the model; it performs lightweight calibration based solely on the most recent observations, keeping the correction time within 200 milliseconds. This maintains second-level response capability while ensuring the prediction results can promptly respond to sudden changes in fire rate or wind direction fluctuations.

[0063] Based on the above embodiments, as a preferred implementation, the edge computing node is a locally deployed independent computing device, and the edge computing node is disconnected from the public network.

[0064] In this embodiment, an edge computing node refers to a hardware device with computing capabilities deployed close to the data source (fire scene acquisition terminal), such as an industrial computer, an embedded AI computing box, or a ruggedized server; local deployment means that the node is located in the same field or a nearby local area network as the fire scene acquisition terminal, rather than a remote cloud data center; disconnection from the public network means that the node does not access the Internet and does not rely on any external network services or cloud computing power.

[0065] Edge computing nodes communicate directly with field data acquisition terminals (cameras, drones, weather stations) via 5G or industrial wireless LANs, and simultaneously connect to command and display terminals, forming a closed loop of "field data acquisition terminal—edge computing node—command and display terminal," with the entire process bypassing the public network. In actual deployment, the node is pre-loaded with a standardized historical fire database, pre-trained lightweight model weights, and various parameter mapping tables; all data is stored and processed locally. Because it is disconnected from the public network, the node naturally avoids the risks of transmission delays and communication interruptions caused by unstable network signals, insufficient bandwidth, or cloud service congestion. Furthermore, in fire environments such as forests, mountains, and underground structures where there is no public network coverage or the public network is damaged, the node can still operate independently, fully executing steps S1 to S4 and dynamic correction steps.

[0066] Secondly, embodiments of the present invention provide a fire spread prediction device based on multi-source data with a second-level prediction capability. Based on the multi-source data-based fire spread prediction method described in the above embodiments, the device 300 includes: The data acquisition module 310 is used to acquire multi-source data of the fire scene in real time, including visual data of the fire scene and real-time wind field data.

[0067] The feature extraction module 320 is used to extract combustible features and fire status features based on the fire scene visual data. The combustible features include the type of combustible and the distribution range of the combustible, and the fire status features include the current burning area boundary.

[0068] The spatiotemporal fusion module 330 is used to perform spatiotemporal matching and normalization feature fusion of the combustible material characteristics, the fire state characteristics and the real-time wind field data based on a unified geographic coordinate system, and generate a fusion feature matrix.

[0069] The prediction inference module 340 is used to input the fused feature matrix into the pre-trained lightweight fire spread prediction model and output the fire spread prediction result, which includes the fire spread direction, spread rate and combustion boundary expansion range.

[0070] Based on the same concept, this invention also provides a schematic diagram of a physical structure, such as... Figure 4As shown, the server may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions stored in the memory 430 to execute the steps of the second-level fire spread prediction method based on multi-source data as described in the above embodiments.

[0071] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0072] Based on the same concept, embodiments of the present invention also provide a non-transitory computer-readable storage medium storing a computer program containing at least one piece of code that can be executed by a master control device to control the master control device to implement the steps of the second-level prediction method for fire spread based on multi-source data as described in the above embodiments.

[0073] Based on the same technical concept, this application also provides a computer program, which, when executed by a main control device, is used to implement the above-described method embodiments.

[0074] The program may be stored, in whole or in part, on a storage medium packaged with the processor, or in part or in whole on a memory not packaged with the processor.

[0075] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the spread of fire in a fire scene within seconds based on multi-source data, characterized in that, include: S1. Real-time acquisition of multi-source data from the fire scene, including visual data of the fire scene and real-time wind field data; S2. Extract combustible material features and fire status features based on the fire scene visual data. The combustible material features include the type of combustible material and the distribution range of combustible material. The fire status features include the current burning area boundary. S3. Using a unified geographic coordinate system as a reference, the characteristics of combustibles, the characteristics of fire status, and the real-time wind field data are spatiotemporally matched and normalized to generate a fused feature matrix. S4. Input the fused feature matrix into the pre-trained lightweight fire spread prediction model and output the fire spread prediction result, which includes the fire spread direction, spread rate and combustion boundary expansion range.

2. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 1, characterized in that, S1 specifically includes: Visual data of the fire scene is collected by fixed surveillance cameras, inspection drones or individual law enforcement recorders, and the visual data of the fire scene includes fire scene video streams or static images. Real-time wind field data is collected by a portable weather station or a weather sensor mounted on a drone. The real-time wind field data includes real-time wind speed and wind direction data. The frame rate for acquiring the visual data of the fire scene shall not be less than 2fps and the resolution shall not be less than 1080P; the acquisition frequency for the real-time wind field data shall not be less than 1Hz.

3. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 1, characterized in that, S2 specifically includes: S21. Perform preprocessing on the key frames of the fire scene visual data, including defogging, noise reduction, and contrast enhancement. S22. Input the preprocessed keyframes into the trained lightweight target detection model, and output the identification results of the combustible material type and the combustible material distribution range through the lightweight target detection model; S23. Input the preprocessed keyframes into the trained lightweight semantic segmentation model, and the lightweight semantic segmentation model outputs the segmentation results of the open flame core area, smoldering area and burned area to determine the boundary of the current burning area.

4. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 1, characterized in that, S3 specifically includes: S31. Using a preset geographic coordinate system as a unified reference, convert the pixel coordinates of the combustible material distribution range and the current combustion area boundary into geographic spatial coordinates, and convert the real-time wind field data into wind field vector data within the same geographic space. S32. Construct a fusion feature matrix, which includes features in the following dimensions: the type of combustible material and its corresponding inherent combustion parameters, the spatial distribution coordinates of the combustible material, the wind field vector data, the coordinates of the current combustion zone boundary and center, and the historical fire spread feature parameters of the same scenario; wherein, the inherent combustion parameters include calorific value, standard spread rate, and ignition point; S33. After normalizing all feature dimensions in the fused feature matrix to the same numerical range, feature splicing is used for fusion to generate a standardized fused feature matrix.

5. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 1, characterized in that, The lightweight fire spread prediction model is based on a historical fire database, pre-trained using a combination of physical model constraints and data-driven methods, and then deployed on edge computing nodes after lightweight processing. The embedding methods for the physical model constraints include: Standardize and map key parameters in the classic physical laws of fire spread to multi-source data characteristics; A sparse mask is set in the self-attention mechanism of the lightweight fire spread prediction model. The sparse mask is based on the classical fire spread physics law. The sparse mask is used to shield invalid calculation areas and strengthen the focus on high-risk areas. Add a physical constraint loss term to the loss function of the lightweight fire spread prediction model during training; The lightweighting process includes: Model pruning is performed based on the contribution of each channel to the physical constraint loss; Quantize the 32-bit floating-point weight parameter into an 8-bit integer; The lightweight fire spread prediction model is trained by a teacher model that incorporates physical constraints, so that the lightweight fire spread prediction model fits the output of the teacher model, the attention weight distribution, and the physical prior features.

6. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 5, characterized in that, In step S4, after the fused feature matrix is ​​input into the lightweight fire spread prediction model, the lightweight fire spread prediction model outputs the fire spread prediction results for the next 1 minute and the next 2 minutes. The fire spread prediction results also include high-risk spread areas. After outputting the fire spread prediction results, dynamic correction is also included: The actual combustion change data of the current video frame and the previous video frame are obtained, and the fire spread prediction result is dynamically corrected based on the actual combustion change data; wherein, the actual combustion change data includes the positional change of the boundary of the combustion area.

7. The method for predicting fire spread in a fire scene within seconds based on multi-source data according to claim 5, characterized in that, The edge computing node is a locally deployed independent computing device, and the edge computing node is disconnected from the public network.

8. A fire spread prediction device based on multi-source data with second-level accuracy, characterized in that, include: The data acquisition module is used to collect multi-source data from the fire scene in real time, including visual data of the fire scene and real-time wind field data. The feature extraction module is used to extract combustible features and fire status features based on the fire scene visual data. The combustible features include the type of combustible and the distribution range of the combustible, and the fire status features include the current burning area boundary. The spatiotemporal fusion module is used to perform spatiotemporal matching and normalization feature fusion of the combustible material characteristics, the fire state characteristics and the real-time wind field data based on a unified geographic coordinate system, and generate a fusion feature matrix. The prediction and inference module is used to input the fused feature matrix into a pre-trained lightweight fire spread prediction model and output the fire spread prediction results, which include the fire spread direction, spread rate and combustion boundary expansion range.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting the spread of fire in a fire scene in seconds based on multi-source data as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting the spread of fire in a fire scene in seconds based on multi-source data as described in any one of claims 1 to 7.

Images