Intelligent analysis system and method based on traces of human thermal injury in fire scene

Abstract

This invention relates to an intelligent analysis system and method for thermal injury traces on the human body in fire scenes, belonging to the interdisciplinary field of forensic identification and artificial intelligence. The system collects images of burn traces on the bodies of victims in fires, constructs a labeled database containing multiple types of thermal injury traces, extracts multimodal features from the images using convolutional neural networks, and trains a deep learning model to achieve: automatic identification and classification of thermal injury traces (including flame damage, incandescent body contact damage, gas explosion damage, oxidizer-related damage, mechanical superposition damage, and near-death burns); inference of the cause and process of injury; determination of the cause and time of death (distinguishing between ante-death burns and post-death cremation); analysis of the manner of death (normal / self-immolation / homicide); and comprehensive assessment of the nature of the fire (accident / malicious fire / arson). This invention solves the accuracy and efficiency problems caused by the reliance on subjective experience in traditional fire investigations, providing scientific and efficient technical support for case solving.

Description

Technical Field

[0001] This invention relates to an intelligent analysis system and method based on thermal injury traces of the human body in a fire, belonging to the interdisciplinary field of forensic identification and artificial intelligence. Background Technology

[0002] In fire-related case investigations, accurately identifying the following elements is crucial for solving the case: the cause and process of injury of the injured; the cause and time of death of the deceased (distinguishing between death by burning before death and death by burning of the body after death); analysis of the manner of death (natural / suicide / homicide); and comprehensive assessment of the nature of the fire (accident / malicious fire / arson).

[0003] However, traditional identification methods rely heavily on forensic experience, which has three major limitations: First, manual identification is greatly influenced by the subjective factors of the examiners. Different examiners may interpret the same trace or evidence differently, making it difficult to guarantee the accuracy and consistency of the results. Second, fire scenes are often complex and changeable. After being burned at high temperatures, the characteristics of injuries and traces on the human body may be blurred, confused, or difficult to identify, increasing the difficulty and uncertainty of manual identification. Furthermore, traditional identification methods are inefficient when dealing with large amounts of complex fire evidence, failing to meet the demands of modern society for rapid detection and efficient handling of fire-related cases.

[0004] Currently, in the field of fire-related cases, there is no system capable of intelligently assessing traces of thermal injury to the human body through deep learning. In particular, there is a lack of integrated solutions covering key identification tasks such as inferring the cause and process of injury; determining the cause and time of death (distinguishing between death by burning before death and death by cremation after death); analyzing the manner of death (normal / self-immolation / homicide); and comprehensively assessing the nature of the fire (accident / accidental / arson). Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent analysis system and method based on thermal injury traces of the human body in a fire, particularly a system and method for intelligently analyzing the causes of death and injury and the nature of the fire based on these traces. Its core innovation lies in: data acquisition → constructing a multimodal injury feature database → training a cascaded deep learning model → achieving intelligent trace comparison → generating an interpretable assessment report. This solves the accuracy and efficiency problems caused by the reliance on subjective experience in traditional fire investigations.

[0006] To address the above problems, the technical solution adopted by the present invention is as follows:

[0007] The intelligent analysis system based on thermal injury traces of human bodies in fires is a system that intelligently analyzes the causes of death and injury and the nature of fires based on thermal injury traces of human bodies in fires. Specifically, it includes the following modules:

[0008] Image acquisition module: used to acquire digital images of burn marks on the bodies of the dead and injured in the fire;

[0009] Human burn scar database module: used to store labeled burn image datasets;

[0010] Feature extraction module: Employs a convolutional neural network (CNN) to extract multimodal morphological features from the input image;

[0011] Intelligent judgment module: Based on a cascaded deep learning model architecture, it executes sequentially as follows:

[0012] (a) Identification and localization of thermal damage trace types: Localizing regions with specific thermal damage types through spatial attention mechanisms;

[0013] (b) Injury process reconstruction: The dynamic process of injury formation is simulated based on temporal recursive networks;

[0014] (c) Determination of death status: Combining respiratory carbon deposition, cardiac blood CO-Hb concentration spectrum analysis and skin tissue elasticity characteristics, output the probability value of burning to death before life / cremation after death;

[0015] (d) Determining the manner of death: Generate a probability matrix for suicide, homicide, or accident;

[0016] (e) Comprehensive determination of fire nature: Combine the output results of (a)-(d) to generate a probability assessment of arson, accident, and natural disaster.

[0017] Furthermore, the annotation dimensions of the human burn scar database module include:

[0018] Thermal damage type tags: flame thermal damage, incandescent body contact thermal damage, gas explosion thermal damage, oxidizer contact thermal damage, oxidizer deflagration radiation thermal damage, oxidizer thermal damage healing marks, mechanical superposition damage, near-death burns;

[0019] Anatomical localization labels for the injury site: 86 key areas defined by international anatomical terminology; Labels relating injury time and severity: divided into stages I (congestion) to IV (carbonization) based on histological changes.

[0020] Furthermore, the feature extraction module includes:

[0021] The fractal dimension of texture in carbonized areas of the skin;

[0022] Edge gradient distribution of bubble morphology;

[0023] Color space differences between wound and normal tissue;

[0024] Infrared thermal radiation characteristics of deep tissue damage.

[0025] Furthermore, both the oxidizer contact-type thermal damage and the oxidizer deflagration radiation-type thermal damage include: oxidizer residual morphology identification and deflagration radiation range modeling, providing strong support for fire cause analysis.

[0026] Furthermore, the death status determination in the intelligent judgment module is achieved by analyzing carbon deposits in the respiratory tract, carboxyhemoglobin concentration in the cardiac blood, and skin tissue elasticity.

[0027] Furthermore, the hardware deployment of the intelligent analysis system also includes:

[0028] Edge computing nodes: used to perform on-site image preprocessing;

[0029] Central server: Used to run multi-task inference models;

[0030] Forensic interactive terminal: used to display timeline simulation animations.

[0031] An intelligent analysis method based on thermal injury traces of the human body in a fire, which employs the aforementioned intelligent analysis system based on thermal injury traces of the human body in a fire, includes the following steps:

[0032] S1: Data Acquisition and Standardization Processing

[0033] Collect images of no fewer than 100 forensic burn cases and label the injury type, anatomical location, and injury mechanism for each image;

[0034] Thin-plate spline transformation was performed to correct cadaver posture distortion, and a U-Net network was used to segment adhesion and injury areas to create an anatomical localization reference.

[0035] S2: Construction of a multi-dimensional feature database:

[0036] Images are labeled according to 9 types of thermal injury, 86 anatomical regions, and 4 levels of injury severity.

[0037] A graph database is used to store feature relationships; a human burn scar database module is constructed.

[0038] S3. Multimodal Feature Extraction:

[0039] The texture fractal dimension of the skin carbonization area, the edge gradient distribution of blister morphology, the color space difference between the wound and normal tissue, and the infrared thermal radiation characteristics of deep tissue damage were extracted by using a convolutional neural network (CNN).

[0040] S4. Training of Cascaded Deep Learning Models:

[0041] The training set, validation set, and test set are divided into three groups according to a 70%:15%:15% ratio.

[0042] Train sequentially and execute in cascades:

[0043] Thermal damage trace type identification and localization, i.e. damage type classification: using the EfficientNetB4 model to output 9 probability distributions;

[0044] Injury process reconstruction: Time-intensity curves were generated using an LSTM network;

[0045] Mortality status determination: A Transformer model is used to fuse three types of biomarkers;

[0046] Method of death determination: Generate a probability matrix for suicide, homicide, or accident;

[0047] S5. Intelligent Analysis and Report Generation:

[0048] Input a new case image and match feature templates using a cosine similarity greater than 0.85;

[0049] Generate a comprehensive report, including:

[0050] Animation demonstrating the injury mechanism;

[0051] 95% confidence level time interval for death;

[0052] Arson suspicion index: 0-100 points;

[0053] Generate a heatmap of key evidence attribution;

[0054] S6. Interpretable output:

[0055] Generate a decision path tree diagram. The path is: carbonization fractal dimension > 0.7 → water bubble gradient variance < 2.5 → CO-Hb > 60% → determine ante-death by burning.

[0056] Furthermore, in step S4, the cascaded deep learning model must meet the following conditions:

[0057] When the confidence level of the "mechanical superposition damage" category in the probability distribution of the thermal damage trace type identification and localization output is higher than the preset threshold, the spatiotemporal correlation module of bone fracture and surface burn is activated for analysis.

[0058] When the probability distribution of the thermal damage trace type identification and localization output shows that any of the categories of oxidizer contact thermal damage, oxidizer deflagration radiation thermal damage, or oxidizer thermal damage healing traces has a confidence level higher than a preset threshold, a cross-modal attention mechanism is activated to associate the image with the chemical spectrum for analysis.

[0059] Furthermore, the spectral data and microscopic image data of the cardiac CO-Hb concentration spectral analysis and respiratory carbon deposition analysis combined in the death status determination in step S4 are derived from forensic toxicology and pathology examination reports.

[0060] Furthermore, in step S5, the generated key evidence attribution heatmap is generated using the Grad-CAM algorithm; in step S6, a decision path tree diagram of key decision factors is generated based on the SHAP value.

[0061] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0062] This invention solves the accuracy and efficiency problems caused by the reliance on subjective experience in traditional fire investigations, providing scientific and efficient technical support for case solving. It reconstructs the causal chain from microscopic traces to the macroscopic nature of the fire (accident / malicious / arson), providing quantitative technical evidence for forensic identification. Attached Figure Description

[0063] Figure 1 This is a technical flowchart of the present invention, illustrating the entire process of data acquisition, database construction, trace feature extraction, deep learning model training, intelligent trace comparison and application. Detailed Implementation

[0064] 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.

[0065] Example 1

[0066] like Figure 1 As shown in this embodiment, an intelligent analysis system based on thermal injury traces of the human body in a fire is a system for intelligently analyzing the causes of death and injury and the nature of the fire based on thermal injury traces of the human body in a fire. Specifically, it includes the following modules in sequence:

[0067] Image acquisition module: used to acquire digital images of burn marks on the bodies of the dead and injured in the fire;

[0068] Human burn scar database module: used to store annotated burn image datasets; the annotation dimensions of the human burn scar database module include:

[0069] Thermal damage type labels: flame thermal damage, incandescent body contact thermal damage, gas explosion thermal damage, oxidizer contact thermal damage, oxidizer deflagration radiation thermal damage, oxidizer thermal damage healing marks, mechanical superposition damage, and near-death burns; among them, the oxidizer contact thermal damage and oxidizer deflagration radiation thermal damage both include: oxidizer residual morphology identification and deflagration radiation range modeling, providing strong support for fire cause analysis.

[0070] Anatomical localization labels for the injury site: 86 key areas defined by international anatomical terminology; Labels relating injury time and severity: divided into stages I (congestion) to IV (carbonization) based on histological changes.

[0071] Feature extraction module: Employs a convolutional neural network (CNN) to extract multimodal morphological features from the input image; the feature extraction module includes:

[0072] The fractal dimension of texture in carbonized areas of the skin;

[0073] Edge gradient distribution of bubble morphology;

[0074] Color space differences between wound and normal tissue;

[0075] Infrared thermal radiation characteristics of deep tissue damage.

[0076] Intelligent judgment module: Based on a cascaded deep learning model architecture, it executes sequentially as follows:

[0077] (a) Identification and localization of thermal damage trace types: Localizing regions with specific thermal damage types through spatial attention mechanisms;

[0078] (b) Injury process reconstruction: The dynamic process of injury formation is simulated based on temporal recursive networks;

[0079] (c) Death status determination: Combining respiratory carbon deposition, cardiac CO-Hb concentration spectrum analysis and skin tissue elasticity characteristics, output the probability value of burning to death before death / cremation after death; The death status determination in the intelligent judgment module is achieved by analyzing respiratory carbon deposition, cardiac carboxyhemoglobin concentration and skin tissue elasticity.

[0080] (d) Determining the manner of death: Generate a probability matrix for suicide, homicide, or accident;

[0081] (e) Comprehensive determination of fire nature: Combine the output results of (a)-(d) to generate a probability assessment of arson, accident, and natural disaster.

[0082] In this embodiment, the hardware deployment of the intelligent analysis system further includes:

[0083] Edge computing nodes: used to perform on-site image preprocessing;

[0084] Central server: Used to run multi-task inference models;

[0085] Forensic interactive terminal: used to display timeline simulation animations.

[0086] Example 2

[0087] This embodiment is an intelligent analysis method based on thermal damage traces of the human body in a fire. It adopts the intelligent analysis system based on thermal damage traces of the human body in a fire from Embodiment 1 above, and includes the following steps:

[0088] S1: Data Acquisition and Standardization Processing

[0089] Collect images of no fewer than 100 forensic burn cases and label the injury type, anatomical location, and injury mechanism for each image;

[0090] Thin-plate spline transformation was performed to correct cadaver posture distortion, and an I-Net network was used to segment adhesion and injury areas to create an anatomical localization reference.

[0091] S2: Construction of a multi-dimensional feature database:

[0092] Images are labeled according to 9 types of thermal injury, 86 anatomical regions, and 4 levels of injury severity.

[0093] A graph database is used to store feature relationships; a human burn scar database module is constructed.

[0094] S3. Multimodal Feature Extraction:

[0095] The texture fractal dimension of the skin carbonization area, the edge gradient distribution of blister morphology, the color space difference between the wound and normal tissue, and the infrared thermal radiation characteristics of deep tissue damage were extracted by using a convolutional neural network (CNN).

[0096] S4. Training of Cascaded Deep Learning Models:

[0097] The training set, validation set, and test set are divided into three groups according to a 70%:15%:15% ratio.

[0098] Train sequentially and execute in cascades:

[0099] Thermal damage trace type identification and localization, i.e. damage type classification: using the EfficientNetB4 model to output 9 probability distributions;

[0100] Injury process reconstruction: Time-intensity curves were generated using an LSTM network;

[0101] Death status determination: The Transformer model is used to fuse three types of biomarkers; the cardiovascular CO-Hb concentration spectral analysis and respiratory carbon deposition analysis combined in the death status determination in step S4 are based on spectral data and microscopic image data from forensic toxicology and pathology examination reports.

[0102] Method of death determination: Generate a probability matrix for suicide, homicide, or accident;

[0103] The cascaded deep learning model must meet the following conditions:

[0104] When the confidence level of the "mechanical superposition damage" category in the probability distribution of the thermal damage trace type identification and localization output is higher than the preset threshold, the spatiotemporal correlation module of bone fracture and surface burn is activated for analysis.

[0105] When the probability distribution of the thermal damage trace type identification and localization output shows that any of the categories of oxidizer contact thermal damage, oxidizer deflagration radiation thermal damage, or oxidizer thermal damage healing traces has a confidence level higher than a preset threshold, a cross-modal attention mechanism is activated to associate the image with the chemical spectrum for analysis.

[0106] S5. Intelligent Analysis and Report Generation:

[0107] Input a new case image and match feature templates using a cosine similarity greater than 0.85;

[0108] Generate a comprehensive report, including:

[0109] Animation demonstrating the injury mechanism;

[0110] 95% confidence level time interval for death;

[0111] Arson suspicion index: 0-100 points;

[0112] Generate key evidence attribution heatmaps using the Grad-CAM algorithm;

[0113] S6. Interpretable output:

[0114] A decision path tree diagram is generated based on the SHAP value. The path is: carbonization fractal dimension > 0.7 → water bubble gradient variance < 2.5 → CO-Hb > 60% → determined to have died from burns before death.

[0115] Example 3 (Arson Case Analysis)

[0116] Case Background: A fire broke out in a residential building, resulting in the deaths of two people. Police initially suspected arson and needed to analyze the bodies to obtain evidence.

[0117] Data Acquisition and Standardization: Images of burn cases from two corpses at the fire scene were acquired (a total of 15 high-resolution photos from different angles and locations). Each image was labeled with its injury type (e.g., deep burns, superficial burns), anatomical location (e.g., face, limbs, trunk), and injury mechanism (preliminarily suspected to be flame burns). Thin-plate spline transformation was used to correct postural distortions of the corpses caused by the fire scene environment (e.g., the corpses being curled up). A U-Net network was used to segment adhesion injury areas (e.g., areas where skin adheres to clothing after burns) to create anatomical positioning references, clearly showing the relative positional relationship between the injuries and normal tissues of each part of the corpse.

[0118] Multi-dimensional feature database construction: Based on the 9 types of thermal injuries, 86 anatomical regions, and 4-level injury severity staging standards in Example 2, the cadaver images in this case were annotated in detail. For example, in the facial anatomical region (one of the 86 anatomical regions), superficial burns (one of the 9 types of thermal injuries) were found, with a severity level of 1 (mild). These annotated data were used to store feature relationships in a graph database, constructing a human burn trace database module to facilitate rapid retrieval of relevant feature data during subsequent model training and feature extraction.

[0119] Multimodal feature extraction: Convolutional neural networks (CNNs) are used to extract the texture fractal dimension of charred skin areas (e.g., the fractal dimension of charred facial skin areas is 0.85), the edge gradient distribution of blister morphology (e.g., the edge gradient variance of blister morphology on the limbs is 2.0), the color space difference between wounds and normal tissues (e.g., the difference between wounds and normal tissues on the trunk in RGB color space is 45), and the infrared thermal radiation characteristics of deep tissue damage (e.g., the peak infrared thermal radiation intensity of muscle tissue is 320).

[0120] Cascaded deep learning model training: The previously collected large amount of data, including arson and non-arson cases, was divided into training, validation, and test sets at a ratio of 70%:15%:15%. First, a thermal damage trace type recognition and localization model was trained using the training set. The EfficientNetB4 model was then used to identify the body images from the case, outputting a probability distribution of 9 categories. The probability of the "accelerant contact thermal damage" category was 0.8 (higher than the preset threshold of 0.7). At this point, the spatiotemporal correlation module for bone fractures and surface burns was activated for analysis, examining the temporal and spatial correlation between bone damage and surface burns to further search for arson evidence.

[0121] Intelligent Analysis and Report Generation: Upon inputting the image of the new case, the system matches it with arson case feature templates in the previous database using a cosine similarity greater than 0.85 (similarity is 0.89). The generated comprehensive report includes an animation demonstrating the injury mechanism, showing the process of fire spreading from the arson point and causing burns to different parts of the body; it provides a 95% confidence time interval for death as 10-15 minutes after the fire; and an arson suspicion index of 92 points (0-100). A key evidence attribution heatmap generated using the Grad-CAM algorithm shows that thermal damage traces on the face and hands have a crucial impact on determining arson, with burn traces on these areas highly correlated with the splashing of accelerants and direct contact with flames during arson.

[0122] Interpretable output: A decision path tree diagram is generated based on the SHAP value. The path is: carbonization fractal dimension > 0.7 (facial carbonization fractal dimension is 0.85) → blister gradient variance < 2.5 (limb blister gradient variance is 2.0) → CO-Hb > 60% (cardiac blood test CO-Hb concentration is 68%) → determined to be burned before death. Combined with evidence such as on-site detection of accelerant residues, this further supports the determination of arson.

[0123] Example 4 (Accident Identification)

[0124] Case Background: A fire broke out in a factory workshop, resulting in injury to one worker. It is necessary to determine whether this fire was caused by a production accident and whether the worker's injuries are consistent with an accident.

[0125] Data Acquisition and Standardization: Images of the injured worker's burn case (10 photos from different angles and locations) were collected, and the injury type (e.g., partial-thickness burn), anatomical location (e.g., arm, leg), and injury mechanism (preliminarily suspected to be flame burns caused by high-temperature equipment) were labeled. Thin-plate spline transformation was performed to correct postural distortions caused by irregular postures in the factory environment (e.g., the twisted posture of the worker when avoiding a fire source). U-Net network was used to segment adhesion injury areas (e.g., the parts of the skin that adhered to the factory protective clothing after the burn) to create an anatomical positioning reference, clearly showing the relative positions of the worker's injuries and normal tissues at various parts of the body.

[0126] Multi-dimensional feature database construction: The worker's cadaver image was labeled according to the classification criteria in Example 2. For example, partial-thickness burn (one of 9 types of thermal injury) was found in the arm anatomical section (one of 86 anatomical sections), with a damage level of grade 2 (moderate). The feature association relationship was stored through a graph database to construct a human burn trace database module, which facilitates subsequent model calls and analysis.

[0127] Multimodal feature extraction: Convolutional neural networks (CNNs) are used to extract the texture fractal dimension of charred skin areas (e.g., the fractal dimension of charred skin areas on the arm is 0.78), the edge gradient distribution of blister morphology (e.g., the edge gradient variance of blister morphology on the leg is 3.0), the color space difference between wounds and normal tissues (e.g., the difference between wounds and normal tissues on the arm in RGB color space is 38), and the infrared thermal radiation characteristics of deep tissue damage (e.g., the peak infrared thermal radiation intensity of subcutaneous adipose tissue is 280).

[0128] Cascaded Deep Learning Model Training: The previously collected data containing various fire cases was divided into training, validation, and test sets at a 70%:15%:15% ratio. The training set was used to train a thermal injury trace type recognition and localization model. The EfficientNetB4 model was used to identify the worker's image, outputting nine probability distributions. The confidence score for the "mechanical superimposed injury" category was 0.6 (below the preset threshold of 0.7), so the correlation analysis module was not activated; while the confidence score for the "flame burn" category was 0.85 (above the preset threshold of 0.7), and further analysis was conducted. The injury process reconstruction model used an LSTM network to generate time-intensity curves, showing the change in fire intensity from ignition to worker injury. The death status determination model used a Transformer model to fuse three types of biomarkers (including cardiac CO-Hb concentration, blood inflammation markers, and tissue damage markers), combined with spectral data and microscopic image data, to determine if the worker was currently injured but not dead. The death method determination model generated a probability matrix for suicide, homicide, or accident, showing an accident probability of 98%.

[0129] Intelligent Analysis and Report Generation: Upon inputting the image of the new case, the system matches it to a production accident fire feature template in the database (similarity = 0.88) with a cosine similarity greater than 0.85. The generated comprehensive report includes an animation demonstrating the injury mechanism, showing the process of workers being burned by flames while operating equipment during a fire in the factory workshop; it provides a 95% confidence interval for the injury timeframe as 5-8 minutes after the fire; the arson suspicion index is 8 points (0-100 points), essentially ruling out arson. A key evidence attribution heatmap generated using the Grad-CAM algorithm shows that the burn marks on the worker's arms and legs are key evidence for determining it to be a production accident fire, as the burns on these areas highly match the height and spread path of the flames when the equipment caught fire in the factory workshop.

[0130] Interpretable output: A decision path tree diagram is generated based on the SHAP value. The path is as follows: carbonization fractal dimension > 0.7 (arm carbonization fractal dimension is 0.78) → blister gradient variance < 2.5 (leg blister gradient variance is 3.0, which is slightly higher than the threshold, but it is still judged as an accident based on other features) → CO-Hb is 45% (the upper limit of the normal range is generally 5%, but the worker did not die and was at the fire scene, so there was some carbon monoxide inhalation) → it is determined to be an injury caused by a production accident fire, which provides a strong basis for subsequent accident liability determination and workers' compensation.

[0131] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the structure of the present invention. The arrangement and quantity of the present invention are not limited to this example and can be optimized according to actual engineering conditions. Any modifications, equivalent changes, and decorations made to the above embodiments based on the technical principles of the present invention, without departing from the scope of the present invention, are still within the scope of the present invention.

Claims

1. An intelligent analysis system based on traces of human thermal injury in a fire scene, which is a system for intelligently analyzing the cause of death and injury and the nature of a fire based on traces of human thermal injury in a fire scene, characterized in that, Specifically, the system comprises the following modules in sequence: An image acquisition module for obtaining digital images of burn marks on dead and injured persons at a fire scene; A human burn mark database module for storing annotated burn image datasets; A feature extraction module for extracting multi-modal morphological features from input images using a convolutional neural network (CNN); The feature extraction module includes the following features: Texture fractal dimension of skin charring area; Edge gradient distribution of blister morphology; Color space difference between wound and normal tissue; Infrared thermal radiation features of deep tissue damage; An intelligent research and judgment module based on a cascaded deep learning model architecture, which performs the following functions in sequence: (a) Thermal damage mark type identification and positioning: locating specific thermal damage type areas through a spatial attention mechanism; (b) Wound process reconstruction: simulating the damage formation dynamics process using a time-series recurrent network; (c) Death state determination: combining respiratory tract carbon deposition, CO-Hb concentration spectral analysis of heart blood, and skin tissue elasticity features to output pre-burn death and post-burn corpse probability values; the respiratory tract carbon deposition analysis and CO-Hb concentration spectral analysis of heart blood used in death state determination are based on forensic toxicology and pathology test reports; (d) Death mode determination: generating probability matrices for suicide, homicide, or accident; 2. The intelligent analysis system based on traces of human thermal injury in a fire according to claim 1, characterized in that, (e) Comprehensive fire nature determination: integrating the output results of (a)-(d) to generate probability assessments for arson, accident, and natural disaster; The cascaded deep learning model must satisfy the following conditions: When the confidence level of the "mechanical superimposed damage" category in the probability distribution output by the thermal damage mark type identification and positioning module is higher than a preset threshold, activate the spatiotemporal correlation module for skeletal fracture and surface burn to perform analysis; When any of the confidence levels of the fuel additive contact type thermal damage, fuel additive deflagration radiation type thermal damage, or fuel additive thermal damage post-healing trace categories in the probability distribution output by the thermal damage mark type identification and positioning module is higher than a preset threshold, activate the cross-modal attention mechanism to correlate images and chemical spectra for analysis.

3. The intelligent analysis system based on traces of human thermal injury in a fire according to claim 2, characterized in that, The annotation dimensions of the human burn mark database module include:

4. The intelligent analysis system based on traces of human thermal injury in a fire according to claim 3, characterized in that, Thermal damage type labels: flame thermal damage, hot body contact thermal damage, gas explosion thermal damage, fuel additive contact type thermal damage, fuel additive deflagration radiation type thermal damage, fuel additive thermal damage post-healing trace, mechanical superimposed damage, and pre-death state burn; 5. The intelligent analysis system based on traces of human thermal injury in a fire according to claim 1, characterized in that, Anatomical positioning labels for damage sites: 86 key regions based on the International Anatomical Terminology; Damage time and extent correlation labels: divided into Ⅰ hyperemia stage to Ⅳ charring stage based on histological changes. Both the fuel additive contact type thermal damage and the fuel additive deflagration radiation type thermal damage include fuel additive residue morphology identification and deflagration radiation range modeling, providing strong support for fire cause analysis. The death state determination in the intelligent research and judgment module is achieved by analyzing respiratory tract carbon deposition, CO-Hb concentration in heart blood, and skin tissue elasticity.

6. An intelligent analysis method based on traces of thermal injury of human body in fire scene, which adopts the intelligent analysis system based on traces of thermal injury of human body in fire scene according to any one of claims 1-5, characterized in that, The hardware deployment of the intelligent analysis system also includes: An edge computing node for implementing on-site image preprocessing; A central server for running multi-task inference models; A forensic interaction terminal for displaying time axis deduction animations. The system comprises the following steps: S1: Data collection and standardization processing: Collect no less than 100 cases of burn images for forensic identification, label the damage type, anatomical location, and injury mechanism of each image; Perform thin plate spline transformation to correct the distortion of corpse posture, use U-Net network to segment the adhesion damage area; form an anatomical positioning reference map; S2: Multi-dimensional feature database construction: Label images according to 9 types of thermal damage, 86 anatomical partitions, and 4 levels of damage severity; Use graph database to store feature association; build a human burn trace database module; S3. Multi-modal feature extraction: Extract the texture fractal dimension of the carbonized skin area, the edge gradient distribution of the blister shape, the color space difference between the wound and normal tissue, and the infrared thermal radiation features of the deep tissue damage through the convolutional neural network CNN; S4. Cascade deep learning model training: Divide the training set, validation set, and test set according to the ratio of 70:15:15; Train and cascade execute in sequence: Thermal damage trace type identification and positioning, i.e. damage type classification: use EfficientNetB4 model to output 9 probability distributions; Injury process reconstruction: use LSTM network to generate time-intensity curve; Death state discrimination: use Transformer model to fuse three types of biomarkers; Death mode determination: generate probability matrix of suicide, homicide, or accident; S5. Intelligent research and report generation: Input new case images, match feature templates through cosine similarity greater than 0.85; Generate a comprehensive report, including: Injury mechanism animation; 95% confidence interval of death time; 0-100 score of arson suspect index; Generate key evidence attribution heat map; S6. Explainable output: Generate decision path tree diagram, path: carbonized fractal dimension > 0.7 → blister gradient variance < 2.5 → CO-Hb > 60% → determine antemortem burning.

7. The intelligent analysis method based on traces of human thermal injury in a fire scene according to claim 6, characterized in that, In step S5, the generated key evidence attribution heat map is generated by Grad-CAM algorithm; in step S6, the decision path tree diagram of key decision factors is generated based on SHAP value.

Images