A fire point tracing method and system for a fire site, and a storage medium

Abstract

The present application relates to a kind of fire starting point fire starting point tracing method and system, storage medium, it is based on space correlation and fire scene trace image intelligent analysis, belong to information tracing field. Including the following steps: data acquisition and database construction, and carry out standardization processing and label management;Fire trace feature extraction in three-dimensional space, including spread direction trace feature, fire starting site local feature and fire starting point local trace feature extraction;Mapping model is constructed from three-dimensional space to two-dimensional image, and cross-modal feature correlation model is established;Machine learning analysis, using deep learning algorithm, the training of the large amount of fire scene data and laboratory simulation data collected, compare field image and database image, calculate similarity and determine fire point position;Model optimization and dynamic learning, by dynamic incremental learning mechanism, continuously introduce new data, optimize model performance, improve accuracy and generalization ability.The present application greatly improves the scientificity and accuracy of fire investigation.

Description

TECHNICAL FIELD

[0001] The present application relates to a fire site fire point tracing method and system, storage medium based on spatial correlation and fire scene trace image intelligent analysis, belonging to the technical field of information tracing. BACKGROUND

[0002] As a sudden and destructive disaster, fire causes 800,000 to 900,000 fire-related cases every year, seriously threatening public safety. At the same time, the fire scene is complex and variable, and the investigation is difficult. The project focuses on the key technology of identifying the fire point of fire-related cases. On the basis of first building a "data collection" standard system, collecting fire scene data, building a database platform, and then developing an efficient multi-source data fusion algorithm, accurately extracting key information, based on big data and deep learning analysis of fire range, spread direction, fire site, and fire point trace characteristics, providing support for unknown fire scene fire point identification, establishing "data analysis" and "data application" standard specifications, helping to accurately identify the fire point of fire-related cases, and providing professional law enforcement for public security, fire, emergency and housing departments, ensuring the scientificity and accuracy of fire-related case handling, and maintaining judicial justice.

[0003] In traditional fire investigation, the determination of the fire point depends on the experience of investigators, and is inferred by observing two-dimensional features such as smoke marks and burning residues, which has strong subjectivity and low efficiency. The fire detection method disclosed in the prior art CN114998822A can identify fire, but lacks quantitative analysis of three-dimensional spatial features; the fire source positioning technology of CN119445802A relies on sensor data and is difficult to apply to complex scenes. Currently, three-dimensional scanning technology can restore the spatial structure of the fire scene with high precision, but how to combine three-dimensional data with two-dimensional trace feature images to realize intelligent tracing of fire sites and fire points is still a technical difficulty. SUMMARY

[0004] In view of the deficiencies in the prior art, the present application proposes a fire site fire point tracing method and system based on spatial correlation and fire scene trace image intelligent analysis, storage medium, which systematically collects three-dimensional scanning scene data and laboratory simulation data of fire-related cases, constructs a fire point database containing fire site, fire cause, fire site, fire point, time and geographic information, and develops a fire site fire point analysis large model using deep learning technology, breaking through the limitations of traditional methods and realizing efficient and accurate identification of fire sites and specific fire points.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] A fire site fire point tracing method based on spatial correlation and fire scene trace image intelligent analysis, comprising the following steps:

[0007] S 1. Data collection and database construction, collecting three-dimensional scanning data and image data of fire scenes, generating standard data sets by combining laboratory simulated fires, establishing a fire point database containing multi-dimensional and multi-modal information of fire sites, fire causes, fire locations, fire points, and time, and performing standardization processing and labeling management;

[0008] S2. Fire scene trace feature extraction in three-dimensional space, including spread direction trace feature extraction, fire location local feature extraction, and fire point local trace feature extraction;

[0009] S3. Mapping model from three-dimensional space to two-dimensional image, generating a two-dimensional image set by multi-view rendering projection of three-dimensional data, and extracting texture features, color changes, and edge contour features using a convolutional neural network; feature fusion and correlation modeling, deep fusion of three-dimensional space features and two-dimensional image features, and establishment of a cross-modal feature correlation model;

[0010] S4. Machine learning analysis, using deep learning algorithms to train a large amount of collected fire scene data and laboratory simulation data, constructing a fire location and fire point analysis model, inputting a large amount of labeled three-dimensional space feature data and two-dimensional image data to make the model learn the feature patterns and rules of fire location and fire point in different fire scenarios; intelligent comparison of images based on a twin neural network architecture combined with an attention mechanism, comparing on-site images with database images, calculating similarity and determining fire point location; S5. Model optimization and dynamic learning, continuously introducing new data through a dynamic incremental learning mechanism to optimize model performance, improve accuracy and generalization ability.

[0011] Further, in S1, the data sources include:

[0012] Three-dimensional scanning data of fire-related cases: using three-dimensional laser scanning equipment to obtain spatial data of real fire scenes, recording complete information of smoke, burning, and deformation;

[0013] Laboratory simulation data: building experimental models of different fire causes and situations in a controllable environment, and recording fire process and feature information at each stage using three-dimensional scanning technology;

[0014] Data preprocessing: using MeshLab, CloudCompare three-dimensional processing software to remove noise, reconstruct discrete point clouds, and standardize the collected data, and labeling the fire point and important feature positions with multi-level and multi-dimensional labels.

[0015] Further, in S2, the three-dimensional space fire scene trace feature extraction specifically includes:

[0016] Trace feature extraction of spreading direction: identify trace features of charring, smoking, and melting on the surface of combustible and non-combustible materials, and infer the fire spreading path;

[0017] Trace feature extraction of fire starting site: analyze the material properties and local deformation damage signs of the fire starting site, and establish a preliminary judgment of the fire starting location;

[0018] Trace feature extraction and comprehensive determination of fire starting point: refine the trace features of the fire starting point and determine the location of the fire starting point based on the comprehensive characteristics of the fire scene.

[0019] Further, in S2, the refined trace features of the fire starting point include ash accumulation and electrical fault traces.

[0020] Further, in S3, the mapping modeling of the two-dimensional image uses a combination of orthogonal projection and perspective projection to convert three-dimensional spatial features into multi-view two-dimensional images, preserving the spatial topological relationship, and using a convolutional neural network (CNN) to extract and enhance features from the mapped two-dimensional images, ensuring the integrity and robustness of the two-dimensional image in terms of spatial representation.

[0021] Further, in S4, the image intelligent comparison uses a twin network combined with an attention mechanism to perform high-dimensional feature matching and fine-grained similarity determination between the target image and the database image; the similarity determination uses a comprehensive determination method that combines cosine similarity and Euclidean distance weighting, thereby improving the accuracy and stability of the fire starting point location matching.

[0022] Further, in S5, the model optimization specifically includes:

[0023] Incremental learning: introducing newly collected fire case data to continuously update the deep neural network with small steps, enhancing the model's adaptability to new types of fire scenes;

[0024] Knowledge distillation: transferring the knowledge of large and complex models to lightweight sub-models to improve model deployment efficiency while maintaining determination accuracy, adapting to terminal device operation requirements;

[0025] Automatic optimization of hyperparameters: using Bayesian optimization algorithms to automatically optimize key hyperparameters such as learning rate, regularization coefficient, and loss function weight, improving the overall training efficiency and prediction performance of the model.

[0026] A fire starting site and fire starting point tracing system based on spatial correlation and intelligent analysis of fire scene trace images, which uses the above-mentioned fire starting site and fire starting point tracing method, comprising:

[0027] Data acquisition device: used to collect three-dimensional scanning data, image data, and laboratory simulation data from fire scenes;

[0028] Data storage unit: for constructing a fire point database containing multi-dimensional data such as fire site, fire cause, fire site, fire point, time information and geographic location;

[0029] Trace feature extraction module: for extracting fire-related spatial features from three-dimensional reconstruction data, including combustible and non-combustible spread direction trace features, fire site trace features and fire point trace features;

[0030] Three-dimensional to two-dimensional mapping modeling module: for projecting three-dimensional data through multi-view rendering to generate a set of two-dimensional images, and extracting texture features, color changes and edge contours using a convolutional neural network;

[0031] Feature fusion and correlation modeling module: for deep fusion of three-dimensional spatial features and two-dimensional image features, and establishing a cross-modal feature correlation model;

[0032] Machine learning analysis module: for training the collected data using deep learning algorithms to construct a fire site and fire point analysis large model;

[0033] Image intelligent comparison module: based on a twin neural network architecture combined with an attention mechanism, for comparing on-site images with database images, calculating similarity and determining the fire point location;

[0034] Model dynamic optimization module: introducing incremental learning mechanism, knowledge distillation technology and Bayesian optimization algorithm for dynamic adjustment and optimization of the model.

[0035] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the fire site and fire point tracing method described above.

[0036] After adopting the above technical solution, the present application has at least one of the following beneficial effects compared with the prior art:

[0037] The present application collects fire-related case event site data and laboratory simulation experiment data through three-dimensional scanning, constructs a multi-dimensional and multi-modal fire site database, and covers fire site, fire cause, spread direction, fire site, fire point, fire time and geographic information and other elements. The data is classified and managed using a hierarchical label system and technical naming method to ensure the systematicness, standardization and traceability of the information. On this basis, a large model system based on deep learning is developed, which breaks through the limitations of traditional reliance on manual experience analysis and judgment, and promotes the intelligentization and scientization of fire-related case event investigation work.

[0038] The technical process of this invention includes data acquisition and database construction, fire scene trace feature extraction based on three-dimensional spatial correlation, two-dimensional mapping modeling, intelligent image comparison, model optimization, and practical testing. The three-dimensional spatial trace feature extraction focuses on analyzing the fire spread direction, ignition location, and ignition point characteristics. Specific methods include: using three-dimensional reconstruction and image processing technology to extract smoke trace features, analyzing the characteristics and distribution patterns of heat loss traces from combustible and non-combustible materials, establishing the correlation between traces and heating conditions and processes, and inferring the fire spread path and direction; based on this, combined with trace comparison and correlation analysis, it provides a scientific basis for tracing the ignition location and ignition point.

[0039] In terms of accurate fire ignition point determination, this invention, after determining the direction of fire spread and the location of the fire, meticulously extracts typical local features of thermal damage traces of combustible and non-combustible materials at the ignition point. Through 3D-to-2D mapping modeling, the 3D features are projected onto a 2D image. Combining deep convolutional neural networks and Siamese network structures, intelligent image comparison and feature similarity calculation are performed to achieve accurate tracing of the fire ignition point. Simultaneously, the system introduces dynamic incremental learning and hyperparameter optimization mechanisms to continuously improve the model's accuracy and generalization ability, ensuring the system's reliability in different fire scenarios.

[0040] This invention establishes a scientific, efficient, and scalable fire origin tracing technology system through systematic data construction, refined feature extraction, deep learning model analysis, and dynamic optimization and updating. This system not only significantly improves the scientific rigor and accuracy of fire investigations but also provides strong technical support to public security, fire departments, emergency management, and housing and construction departments, making a positive contribution to public safety and judicial fairness. Attached Figure Description

[0041] Figure 1 This is a technical flowchart of the present invention, which shows the entire process of data acquisition, feature extraction, mapping modeling, comparison and judgment and optimization. Detailed Implementation

[0042] The following is in conjunction with the appendix Figure 1 The present invention will be further described in detail below to facilitate a clear understanding of the invention, but these descriptions do not constitute a limitation thereof.

[0043] In the description of this invention, it should be noted that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0044] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0045] Embodiment 1

[0046] This embodiment takes a residential fire scene as an example to illustrate a fire starting site fire starting point tracing method based on spatial correlation and fire scene trace image intelligent analysis of the present application, comprising the following steps:

[0047] S 1. Data acquisition and database construction, acquire three-dimensional scanning data and image data of fire scene, combine with laboratory simulation fire to generate standard data set, establish fire starting point database containing fire starting place, fire starting reason, fire starting site, fire starting point, time multidimensional and multimodal information, and carry out standardization processing and labeling management;

[0048] Specifically, the data sources include:

[0049] Three-dimensional scanning data of fire-related cases: use three-dimensional laser scanning equipment to obtain the spatial data of the real fire scene, and record the information of smoke, burning and deformation completely; in this embodiment, 10000 fire scene data are collected, covering various scenes such as residential, warehouse, forest, etc., to generate information including smoke trace, wood carbonization area, metal discoloration area, building material structure feature, burning residue feature, thermal radiation and convection feature, physical and chemical change feature, etc.

[0050] Laboratory simulation data: build experimental models of different fire starting reasons and situations in a controllable environment, and use three-dimensional scanning technology to record the fire starting process and feature information at each stage; in this embodiment, scenes such as electrical short circuit and flammable liquid ignition are simulated in a sealed combustion cabin, to generate three-dimensional data set with annotations, including fire starting point position (accurate to centimeter level), smoke diffusion path (annotating diffusion direction and speed), wood carbonization depth (annotating carbonization depth at different temperatures), metal discoloration area (annotating discoloration range under high temperature) and other features.

[0051] Data preprocessing: MeshLab, CloudCompare three-dimensional processing software is used to remove noise, reconstruct discrete point cloud and standardize the collected data, and accurately label the features such as fire point coordinates, fire site, fire cause and other characteristic labels. At the same time, the fire point and important feature position are marked with multi-level and multi-dimensional label tags. Among them, the database includes fire places (such as residential buildings, entertainment places, storage warehouses, factories, etc.), fire causes (electrical short circuit, open fire ignition, human arson, equipment failure, etc.), fire sites (living room, bedroom, warehouse corner, etc.), fire point position (socket, line, etc.), time information (day / night), geographical location (province, city, township, etc.). The case data is classified by using technical naming method, and through the hierarchical label system, the system can quickly locate and analyze the fire characteristics in a specific scene, which is convenient for subsequent quick retrieval and model training. In this embodiment, the first level is the fire place (such as residential building, entertainment place, factory, etc.), the fire cause (such as electrical short circuit, human arson, etc.), the fire site (such as living room, kitchen, etc.), and the fire point (such as socket, wire, etc.). The second level is time (a certain year, month, day, specific to hour), geographical information (province, city, township, specific to street).

[0052] S2. Fire scene trace feature extraction in three-dimensional space, including spread direction trace feature extraction, fire site local feature extraction and fire point local trace feature extraction; through a deep learning model, fire-related spatial features are extracted from three-dimensional reconstruction data. The fire scene trace feature extraction in three-dimensional space specifically includes:

[0053] Spread direction trace feature extraction: identify the charring, smoking and melting trace features of combustible and non-combustible surfaces, and deduce the fire spread path; specifically, comprehensively survey the combustible materials in the fire scene, determine their types, distribution positions and shapes, observe the burning traces on the surface of the combustible materials, and consider the influence of factors such as heat conduction, convection and radiation; analyze the thermal discoloration phenomenon, melting trace distribution and flow direction on the surface of non-combustible materials, and the traces left by the surrounding combustible materials burning on the non-combustible materials, so as to deduce the fire spread direction.

[0054] Fire site trace feature extraction: analyze the material properties of the fire site, local deformation damage signs, and establish a preliminary judgment of the fire location; in this embodiment, based on the determination of the fire spread direction, local features of the fire site are extracted, including the shape, size, structural characteristics of the fire site, etc., such as whether the wall has cracks, whether the doors and windows are intact, whether the electrical lines are exposed, etc. These features may be closely related to the cause of the fire. At the same time, the characteristics of the building materials, decorative materials, etc. of the fire site are analyzed, and the burning performance, heat conduction performance, etc. of different materials will affect the spread of the fire and the trace features of the fire site. Finally, the extracted trace features of the fire site are compared and correlated with known trace features in the database to find similar feature patterns, providing a reference basis for determining the cause of the fire.

[0055] Fire point trace feature extraction and comprehensive determination: refine the trace features of the fire point such as ash accumulation and electrical failure traces, and determine the location of the fire point based on the comprehensive fire scene characteristics. In this embodiment, the trace features of the fire point are further extracted within the fire site, such as the shape of the ash accumulation, the distribution center of the residual liquid traces, and the failure points of electrical equipment. These trace features usually have strong directionality and can directly indicate the location of the fire point. Combining the trace features of the spread direction, the fire site, and the fire point, comprehensive analysis and determination are performed to determine the final location of the fire point, and the cause of the fire is inferred based on the trace features of the fire point, providing a scientific basis for the investigation and handling of fire accidents.

[0056] S3. Construct a mapping model from three-dimensional space to two-dimensional image, generate a set of two-dimensional images by multi-view rendering projection of three-dimensional data, and extract texture features, color changes and edge contour features using convolutional neural network; feature fusion and correlation modeling, deep fusion of three-dimensional space features and two-dimensional image features, and establishment of cross-modal feature correlation model;

[0057] Among them, the mapping modeling of two-dimensional image adopts the method of combining orthogonal projection and perspective projection, which converts three-dimensional space features into multi-view two-dimensional images, preserves the spatial topological relationship, and extracts and enhances the features of the mapped two-dimensional images through convolutional neural network CNN, ensuring the integrity and robustness of the two-dimensional images in spatial representation ability.

[0058] Specifically, in this embodiment, the following methods are included:

[0059] Multi-view rendering projection: advanced multi-view rendering technology is used to project three-dimensional data from multiple angles to generate a rich set of two-dimensional images, revealing the details of the fire scene from different perspectives.

[0060] Maintaining data integrity: Throughout the projection process, ensure that the topological relationship and texture information of the original three-dimensional data are not lost, and the key features of the original data are completely preserved, providing accurate data basis for subsequent analysis.

[0061] Convolutional neural network feature extraction: Use convolutional neural network (CNN) to extract texture features, color changes, edge contours and other important features from the generated two-dimensional images, fully leverage the advantages of CNN in image feature extraction, and obtain discriminative image features.

[0062] Feature fusion and correlation modeling: Deeply fuse the extracted three-dimensional spatial features and two-dimensional image features at the feature level, establish a cross-modal feature correlation model, effectively correlate three-dimensional spatial features and two-dimensional image features, break down the barriers between data dimensions, and provide unified and comprehensive feature description for subsequent image comparison and analysis.

[0063] S4. Machine learning analysis, using deep learning algorithms to train a large amount of fire scene data and laboratory simulation data collected, building a fire site fire point analysis model, by inputting a large number of labeled three-dimensional spatial feature data and two-dimensional image data, the model learns the feature patterns and rules of fire site fire points in different fire scenes, so as to accurately analyze and judge new fire scene data; based on the twin neural network architecture combined with attention mechanism for image intelligent comparison, compare the scene image with the database image, calculate the similarity and determine the fire point position;

[0064] Among them, image intelligent comparison uses twin network combined with attention mechanism to perform high-dimensional feature matching and fine-grained similarity determination on target image and database image; similarity determination adopts a comprehensive determination method of weighted fusion of cosine similarity and Euclidean distance, so as to improve the accuracy and stability of fire point position matching. Specifically as follows:

[0065] Network architecture design: Based on the powerful Siamese neural network (Siamese Network) architecture, combined with attention mechanism, build an image comparison network. The target scene image and the sample image in the database are input into the network at the same time, and the attention mechanism is used to automatically focus on the key areas in the image, such as the trace features near the fire point, highlighting the key information and improving the accuracy of comparison.

[0066] Similarity calculation: By calculating the cosine similarity and Euclidean distance between images, and using weighted fusion method, the matching degree between images is obtained comprehensively, and the similarity between the target scene image and the database image is quantified, providing reliable numerical basis for judgment.

[0067] Determination and output: a similarity threshold is set, and when the calculated similarity exceeds the threshold, it is determined that a matching fire point location in the database is found, and the three-dimensional coordinates of the fire point and the matching confidence are output, providing intuitive and accurate fire point positioning information for fire investigators to assist in fire accident investigation and handling. In this embodiment, when the similarity is higher than the threshold (such as 0.85), it is determined that it is a matching fire point, and the three-dimensional coordinates of the fire point and the confidence (such as 92%) are output.

[0068] S5. Model optimization and dynamic learning, continuously introduce new data through dynamic incremental learning mechanism, optimize model performance, improve accuracy and generalization ability.

[0069] The model optimization specifically includes:

[0070] Incremental learning: introduce newly collected fire case data, continuously update the deep neural network with small steps, and enhance the model's adaptability to new types of fire scene; Specifically, such as introducing incremental learning mechanism every month, continuously incorporating new fire scene data, dynamically adjusting and updating the model. By dynamically adjusting the depth network parameters, the model can continuously adapt to new fire scenes and feature changes, improve the model's generalization ability in complex and variable environments, and ensure the long-term effectiveness and accuracy of the model in actual application.

[0071] Knowledge distillation: by migrating the knowledge of large and complex models to lightweight sub-models, the model deployment efficiency is improved while maintaining the accuracy of the determination, and the terminal device operation requirements are adapted; Specifically, using knowledge distillation technology, the rich knowledge and experience learned by the large model is migrated to the lightweight model, realizing the lightweight and efficient of the model. This makes the lightweight model more convenient to deploy and apply on site devices while retaining key knowledge, expanding the actual application scenarios and application scope of the system.

[0072] Automatic optimization of hyperparameters: use Bayesian optimization algorithm to automatically optimize key hyperparameters such as learning rate, regularization coefficient, and loss function weight, improve the overall training efficiency and prediction performance of the model. Specifically, use Bayesian optimization algorithm to automatically adjust learning rate, loss function weight and other hyperparameters, dynamically optimize the value of hyperparameters according to the performance of the model in the training and verification process, continuously improve the overall performance and generalization ability of the model, so that the model can maintain the best operating state in different fire scenes, and provide more reliable fire point tracing results for fire investigation.

[0073] In actual fire investigation applications, the system operates through the following processes:

[0074] On-site data collection: In actual fire investigation, first of all, high-quality three-dimensional scanning data and image data are collected on the fire scene to obtain comprehensive information of the fire scene, providing a data basis for subsequent analysis.

[0075] Automatic feature extraction: The system automatically processes the collected data, extracts the fire spread direction and preliminary fire site features, quickly narrows down the investigation range, and determines the key analysis area.

[0076] Machine learning analysis: Using deep learning algorithms, the system trains on the collected fire scene data and laboratory simulation data. By inputting a large amount of labeled three-dimensional spatial feature data and two-dimensional image data, a large model for fire site and ignition point analysis is constructed. The model has learned the feature patterns and rules of fire site and ignition point in different fire scenarios. This enables the system to accurately analyze and judge new fire scene data, providing strong support for subsequent image comparison and ignition point determination.

[0077] Intelligent image comparison: Through intelligent image comparison technology, the images collected on the scene are quickly and accurately matched with the images in the database. Using pre-trained models and algorithms, the ignition point is accurately locked in a large number of sample data, improving the efficiency and accuracy of fire investigation.

[0078] Result output and application: The final output is the spatial coordinates of the fire site and ignition point, as well as the inference result of the fire cause, providing scientific and accurate basis for fire responsibility identification and related law enforcement work, assisting relevant departments in fair and reasonable handling of fire accidents and responsibility determination.

[0079] Example 2

[0080] A fire site and ignition point tracing system based on spatial correlation and intelligent analysis of fire scene trace images, which adopts the fire site and ignition point tracing method of the above-mentioned example 1, comprising:

[0081] Data acquisition device: for collecting three-dimensional scanning data, image data and laboratory simulation data of the fire scene;

[0082] Data storage unit: for constructing a fire point database containing multi-dimensional data such as fire site, fire cause, fire site, fire point, time information and geographical location;

[0083] Trace feature extraction module: for extracting fire-related spatial features from three-dimensional reconstruction data, including combustible and non-combustible spread direction trace features, fire site trace features and ignition point trace features;

[0084] Three-dimensional to two-dimensional mapping modeling module: for generating a set of two-dimensional images by multi-view rendering projection of three-dimensional data, and extracting texture features, color changes and edge contours using a convolutional neural network;

[0085] Feature fusion and association modeling module: for deep fusion of three-dimensional spatial features and two-dimensional image features, and establishing a cross-modal feature association model;

[0086] Machine learning analysis module: for training the collected data using deep learning algorithms to build a fire site analysis large model;

[0087] Image intelligent comparison module: based on a twin neural network architecture combined with an attention mechanism, for comparing on-site images with database images, calculating similarity and determining the fire point location;

[0088] Model dynamic optimization module: introducing incremental learning mechanism, knowledge distillation technology and Bayesian optimization algorithm for dynamic adjustment and optimization of the model.

[0089] Embodiment 3

[0090] This embodiment verifies the accuracy and efficiency of the system through comparative experiments, and compares the performance with traditional methods.

[0091] 1. Experimental design

[0092] Data set preparation:

[0093] Use a scanner to collect three-dimensional data of 100 fire scenes; simulate 20 different fire scenes in the laboratory to generate a labeled three-dimensional data set.

[0094] Experimental objective: to verify the accuracy and efficiency of the system in determining the fire point in complex fire scenes.

[0095] 2. Experimental steps

[0096] Data preprocessing: remove noise and feature labeling (technical nomenclature) from the collected three-dimensional data.

[0097] Feature extraction: use the PointNet++ algorithm to extract key spatial features from the preprocessed three-dimensional data, which can effectively capture local and global features in three-dimensional data, providing accurate feature description for subsequent image comparison.

[0098] Image comparison: match the target on-site image with the database image through a twin network combined with an attention mechanism. Input the target on-site image (two-dimensional) and the database image (two-dimensional), and output the similarity score and the fire point location.

[0099] Result evaluation: Compare the accuracy (the proportion of correct determination of fire points), recall rate (the proportion of actual fire points determined), and processing time (the time required for single processing) of the traditional method (such as experience judgment) and the method of the present invention.

[0100] 3. Experimental results

[0101] The experimental results show that the method of the present invention has significantly improved accuracy and recall rate compared to traditional methods, indicating that the system can more accurately determine the fire point; at the same time, the single processing time of the method of the present invention is significantly shortened, reflecting its advantage in processing efficiency.

[0102] 4. Result analysis

[0103] Advantages: The system of the present invention combines three-dimensional spatial features and two-dimensional image features, fully utilizes the advantages of deep learning technology, and can more accurately capture key features of fire scenes, significantly improving the accuracy and efficiency of fire point determination in fire sites, effectively overcoming the limitations of traditional methods in subjectivity and efficiency, and is particularly suitable for fire point tracing in complex fire scenes.

[0104] Limitations: In extremely complex scenarios (such as multiple fire points), the system may need to further optimize model parameters to improve performance.

[0105] 5. Comparison verification

[0106] Traditional method: relies on the experience of investigators, strong subjectivity, low efficiency;

[0107] Method of the present invention: based on deep learning and three-dimensional spatial analysis, strong objectivity, high efficiency, suitable for complex fire scenes.

[0108] Example 4

[0109] A computer-readable storage medium of this embodiment has a computer program stored thereon, which, when executed by a processor, implements the fire point tracing method of the fire site of the above-mentioned embodiment 1.

[0110] The above is only a preferred embodiment of the present invention, and does not limit the structure of the present invention in any form. The arrangement and number of uses of the present invention are not limited to this example, and can be optimized and selected according to engineering practice. Any modification, equivalent change and decoration of the above-mentioned embodiment according to the technical principle of the present invention, without departing from the technical solution of the present invention, are still within the scope of the technical solution of the present invention.

Claims

1. A method for tracing the ignition point of a fire, based on spatial correlation and intelligent analysis of fire scene trace images, characterized in that, Includes the following steps: S1. Data Acquisition and Database Construction: Collect 3D scanning data and image data of the fire scene, combine with laboratory fire simulation to generate a standard dataset, establish a fire point database containing multi-dimensional and multi-modal information on the fire location, cause of fire, fire location, fire point, and time, and perform standardized processing and tagging management. S2. Extraction of fire trace features in three-dimensional space, including trace features of spread direction, local features of ignition location, and local trace features of ignition point; S3. Construct a mapping model from 3D space to 2D images, generate a set of 2D images from 3D data through multi-view rendering projection, and use a convolutional neural network to extract texture features, color changes and edge contour features; feature Fusion and correlation modeling deeply integrates three-dimensional spatial features with two-dimensional image features to establish a cross-modal feature correlation model; S4. Machine learning analysis: Using deep learning algorithms, a large model for analyzing the ignition points of fire locations is built by training a large amount of collected fire scene data and laboratory simulation data. By inputting a large amount of labeled 3D spatial feature data and 2D image data, the model learns the characteristic patterns and rules of ignition points of fire locations under different fire scenarios. Based on the Siamese neural network architecture combined with the attention mechanism, intelligent image comparison is performed to compare on-site images with database images, calculate similarity, and determine the location of the ignition point. S5. Model optimization and dynamic learning: Through a dynamic incremental learning mechanism, new data is continuously introduced to optimize model performance and improve accuracy and generalization ability.

2. The method for tracing the ignition point of a fire location according to claim 1, characterized in that: In S1, the data sources include: 3D scanning data of fire incidents: Spatial data of real fire scenes are obtained using 3D laser scanning equipment, and information on smoke, combustion, and deformation at the scene is fully recorded. Laboratory simulation data: Experimental models of different causes and scenarios of fire were built in a controlled environment, and 3D scanning technology was used to record the fire process and characteristic information of each stage; Data preprocessing: MeshLab and CloudCompare 3D processing software were used to remove noise, reconstruct discrete point clouds and standardize the collected data. At the same time, multi-level and multi-dimensional labels were added to the ignition point and important feature locations.

3. The method for tracing the ignition point of a fire location according to claim 2, characterized in that: In step S2, the extraction of fire scene trace features in three-dimensional space specifically includes: Fire spread direction trace feature extraction: Identify the charring, smoke, and melting trace features on the surfaces of combustible and non-combustible materials to infer the fire spread path; Fire ignition site feature extraction: Analyze the material properties and local deformation and damage signs of the fire ignition site to establish a preliminary judgment of the fire location; Ignition point trace feature extraction and comprehensive judgment: refine the ignition point trace features and determine the location of the ignition point by combining fire scene features.

4. The method for tracing the ignition point of a fire location according to claim 3, characterized in that: In S2, the detailed characteristics of the ignition point traces include ash accumulation and electrical fault traces.

5. The method for tracing the ignition point of a fire location according to claim 4, characterized in that: In S3, the mapping modeling of the two-dimensional image adopts a combination of orthogonal projection and perspective projection to convert the three-dimensional spatial features into a multi-view two-dimensional image, preserve the spatial topological relationship, and use a convolutional neural network (CNN) to extract and enhance features of the mapped two-dimensional image, ensuring the integrity and robustness of the two-dimensional image in terms of spatial representation capability.

6. The method for tracing the ignition point of a fire location according to claim 5, characterized in that: In S4, the intelligent image comparison uses a twin network combined with an attention mechanism to perform high-dimensional feature matching and fine-grained similarity determination on the target image and the database images. The similarity determination adopts a comprehensive determination method that combines cosine similarity and Euclidean distance weighted fusion, thereby improving the accuracy and stability of the fire point location matching.

7. The method for tracing the ignition point of a fire location according to claim 6, characterized in that, In S5, the model optimization specifically includes: Incremental learning: Introducing newly collected fire case data, the deep neural network is continuously updated in small steps to enhance the model's adaptability to new types of fire scenarios; Knowledge distillation: By transferring knowledge from large and complex models to lightweight sub-models, we can improve model deployment efficiency while maintaining decision accuracy and adapting to the operational needs of terminal devices. Automatic hyperparameter optimization: Bayesian optimization algorithm is used to automatically tune key hyperparameters such as learning rate, regularization coefficient, and loss function weights to improve the overall training efficiency and prediction performance of the model.

8. A system for tracing the ignition point of a fire based on spatial correlation and intelligent analysis of fire scene trace images, which adopts the method for tracing the ignition point of a fire as described in claim 7, characterized in that, include: Data acquisition device: used to collect three-dimensional scanning data, image data, and laboratory simulation data of the fire scene; Data storage unit: used to build a fire point database containing multi-dimensional data such as fire location, cause of fire, fire location, ignition point, time information and geographical location; Trace feature extraction module: used to extract fire-related spatial features from 3D reconstruction data, including trace features of the spread direction of combustible and non-combustible materials, trace features of the fire location, and trace features of the ignition point; 3D to 2D mapping modeling module: used to generate a set of 2D images from 3D data through multi-view rendering projection, and to extract features such as texture features, color changes and edge contours using convolutional neural networks; Feature fusion and association modeling module: used to deeply fuse 3D spatial features with 2D image features to establish a cross-modal feature association model; Machine learning analysis module: Used to train the collected data using deep learning algorithms to build a large model for analyzing the ignition point of the fire location; Image intelligent comparison module: Based on the Siamese neural network architecture combined with the attention mechanism, it is used to compare on-site images with database images, calculate similarity and determine the location of the fire point; Model dynamic optimization module: Introduces incremental learning mechanism, knowledge distillation technology and Bayesian optimization algorithm for dynamic adjustment and optimization of model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for tracing the ignition point of a fire location as described in any one of claims 1 to 7.

Images