A virtual simulation training method and system for fire-related case investigation based on a large language model

Abstract

This invention discloses a virtual simulation training method and system for fire-related case investigation based on a large language model (LLM), relating to the fields of public security criminal investigation training and artificial intelligence technology. The method includes: constructing a virtual scene and case metadata for fire-related cases; modeling trainee competency profiles; protecting the scene; conducting on-site investigation; patrolling the scene; preliminary examination; detailed examination; inspection and identification; training evaluation and dynamic difficulty adjustment; and archiving training data and competency tracking. This invention, through the deep integration of large language models and VR virtual simulation technology, fully covers the six legally mandated stages of fire scene investigation, recreating the real fire scene environment and supporting risk-free repeated training. The LLM-driven intelligent evaluation quantifies the completeness of the investigation logic and provides targeted improvement suggestions. The dynamic difficulty adjustment mechanism ensures that the training challenge matches the trainees' abilities, effectively solving the problems of limited training scenarios, subjective evaluation, low efficiency, and fragmented stages in traditional training.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of public security criminal investigation training and artificial intelligence, and in particular to a virtual simulation training method and system for fire-related case investigation based on a large language model. Background Technology

[0002] Fire accident investigation is an important area of ​​collaboration between fire and rescue agencies and criminal investigation departments of public security organs, involving key aspects such as determining the cause of the fire, tracing liability, and securing evidence. According to relevant regulations on fire accident investigation, fire scene investigations must strictly follow legal procedures such as protecting the scene, conducting on-site investigations, patrolling the scene, preliminary examination, detailed examination, and inspection and identification.

[0003] At present, the training for fire case investigation mainly has the following problems: (1) The existing training mainly adopts offline practical training and case teaching methods. Traditional training has problems such as low scene restoration, high training cost, high safety risk and unreproducible cases. The real fire scene is dangerous and irreversible, making it difficult to train repeatedly; (2) The existing VR fire training system is mainly focused on scene restoration and operation simulation, lacking intelligent reasoning and dynamic feedback capabilities. It does not involve the deep integration of large language models and investigation processes, still relies on preset scripts, and cannot dynamically adjust the case and clues according to the trainees' behavior; (3) The application of large language models in the field of fire rescue is concentrated on official document writing, emergency plan formulation, fire hazard identification, etc., and has not yet involved case investigation logic training, especially lacking systematic training on the six statutory links and data linkage between links; (4) The existing training evaluation mainly relies on the operation standardization score, lacks quantitative evaluation of the integrity of investigation logic, the ability to build evidence chains, and the accuracy of legal application, and the evaluation of each link is independent, making it impossible to achieve the closed-loop simulation of "the operation of the previous link affecting the evidence status of the next link", which makes it difficult to objectively measure the training effect. (5) The training sessions in the existing system are independent of each other and cannot simulate the causal relationship between the sessions such as "improper on-site protection leading to evidence contamination" and "on-site investigation information guiding the key points of the investigation" in real investigations. Trainees find it difficult to understand the internal logical connection between the procedures.

[0004] Therefore, there is an urgent need to develop a virtual simulation training method that can fully cover the six statutory procedures for investigating fire-related cases, possess intelligent reasoning and dynamic feedback capabilities, and achieve data linkage between the stages. Summary of the Invention

[0005] The purpose of this invention is to provide a virtual simulation training method and system for fire-related case investigation driven by a large language model, which enables full-process training of six stages: site protection, site investigation, site inspection, preliminary examination, detailed examination, and inspection and identification.

[0006] The objective of this invention is achieved as follows: a virtual simulation training method and system for fire-related case investigation driven by a large language model, comprising the following steps: Step S1: Construction of virtual scene and case metadata for fire-related cases. Extract case metadata based on real fire-related case database, generate a three-dimensional virtual fire scene through VR engine, and dynamically adjust scene parameters according to training objectives using large language model. Step S2: Model the trainee's competency profile, collect basic trainee information and historical training records, construct the trainee competency graph, and generate personalized training path suggestions from the large language model; Step S3, the scene protection step, monitors the setting of the trainee's warning tape, wearing of protective equipment, and personnel entry control operations. When a violation is detected, the big language model dynamically triggers an evidence contamination event and records the evidence status change. The evidence status change is synchronously transmitted to the subsequent steps. Step S4, on-site investigation: The large language model drives the virtual witness and party roles, generating differentiated responses based on the trainees' questioning skills, and the information obtained from the investigation dynamically guides the focus of subsequent investigations. Step S5, the site inspection phase, captures the trainees' inspection trajectory and focus of vision, analyzes the survey coverage using a large language model, and dynamically adjusts the visibility of clues based on the on-site investigation information; Step S6, preliminary investigation stage: trainees conduct a macroscopic investigation of the fire scene to determine the location of the fire and the direction of its spread. The preliminary investigation conclusions serve as input conditions for the detailed investigation. Step S7, detailed inspection process, records the trainee's inspection route, physical evidence location and photo records. The quality and completeness of the evidence collected during the inspection directly affect the credibility of the identification results in the inspection and identification process. Step S8, the inspection and identification process, verifies the representativeness of the sampling location and the standardization of the sampling method of the physical evidence submitted by the trainee. The big language model generates a differentiated identification report based on the operation quality, and evaluates the credibility of the conclusion by combining the operation records of the first five steps. Step S9, Training Assessment and Dynamic Difficulty Adjustment: The large language model summarizes the operational data from the six stages to generate a multi-dimensional assessment report, and dynamically adjusts the subsequent training difficulty parameters based on the scoring algorithm. Step S10, Training Data Archiving and Competency Tracking, stores training process data and evaluation results in training archives to support cross-institutional competency certification.

[0007] Preferably, the evidence contamination events in step S3 include: virtual crowds intruding and damaging footprints, lack of protective equipment leading to blurred traces, and insufficient warning range leading to the movement of physical evidence. Changes in the status of evidence simultaneously affect the visibility of clues and the identification results in the on-site inspection in step S5 and the inspection and identification process in step S8.

[0008] Preferably, the differentiated response in step S4 includes: obtaining detailed information through open-ended questions, obtaining ambiguous answers through closed-ended leading questions, triggering changes in the character's emotions through contradictory follow-up questions, and automatically converting the inquiry content into a written record and having the completeness of the inquiry evaluated by a large language model; the information obtained in step S4 dynamically guides the focus of subsequent investigations, including increasing the weight of electrical circuit investigation when a witness mentions abnormal sounds, increasing the weight of trace investigation when a witness mentions suspicious persons, and increasing the weight of combustion accelerant investigation when a witness mentions abnormal odors.

[0009] Preferably, the dynamic adjustment of clue visibility in step S5 adopts a condition triggering mechanism. When the trainee's gaze stays on the key trace area for more than a preset time and the exploration path covers the key area, the next level of clues is triggered to appear. When the trainee misses the key area, the large language model asks guiding questions in the role of a virtual instructor instead of directly telling the answer.

[0010] Preferably, the preliminary investigation conclusion in step S6 serves as the input condition for the detailed investigation in step S7. If the preliminary investigation makes an incorrect judgment about the location of the fire, the difficulty in discovering key evidence in the detailed investigation will increase, simulating the difficulty in solving the case caused by directional deviation in a real investigation.

[0011] Preferably, the differential identification report in step S8 includes: generating accurate identification conclusions through standardized operations, generating misleading data due to sampling location deviations, generating contamination labeling reports due to non-standard sampling methods, and simulating the uncertainty of real laboratory testing; the large language model in step S8 comprehensively evaluates the credibility of the conclusions based on the operation records of the first five steps, including: determining that the conclusions are insufficiently supported if key evidence is missing due to violations of on-site protection regulations, determining that the investigation is incomplete if on-site investigation information is not fully utilized, and determining that the conclusions are unreliable if the chain of evidence is incomplete.

[0012] Preferably, the multi-dimensional evaluation report in step S9 includes scores for on-site protection standardization, investigation skills proficiency, patrol coverage, inspection integrity, evidence chain logic, and identification and interpretation accuracy, with the weights of each dimension being configurable.

[0013] Preferably, the scoring algorithm dynamically adjusted in step S9 adopts the Elo scoring algorithm, which calculates the ability value based on the trainee's historical training performance. The higher the ability value, the higher the degree of concealment of clues in subsequent training cases, the greater the time pressure, the more interference information, and the higher the complexity of link linkage.

[0014] Preferably, in step S10, the training data archiving adopts a federated learning architecture, where each training institution stores student data locally and only uploads encrypted capability graph parameters to the central server, thus ensuring data privacy and security.

[0015] A virtual simulation training system for fire-related case investigation includes: The VR interaction module includes a head-mounted virtual reality display, a gesture sensor, an eye-tracking module, and a voice acquisition port; The large language model inference engine includes a domain fine-tuning module, a retrieval enhancement generation module, a multi-role dialogue management module, and a process linkage control module; The six-step process control module includes the following sub-modules: site protection, site investigation, site inspection, preliminary investigation, detailed investigation, and inspection and appraisal. The assessment and feedback module includes a multi-dimensional scoring submodule, a dynamic difficulty adjustment submodule, and an assessment report generation submodule; The data management module includes a trainee competency graph database, a case meta-database, a training record database, and an evidence status tracking database. The large language model inference engine has been fine-tuned using professional corpus of fire-related cases. The training data includes fire investigation textbooks, typical case files, physical evidence identification reports, court judgments, and interrogation transcripts. The fine-tuning adopts LoRA low-rank adaptation technology. The six-step process control module supports step jumps and repeated training. Trainees can choose to return to the previous step and repeat the operation after any step is completed. The system records the differences in operation of each attempt and includes them in the evaluation, but the evidence contamination state is irreversible. The evidence status tracking database records the status change history of each piece of evidence, including the initial status, pollution event, pollution cause, and impact stage, supporting the tracing of evidence change processes during training and review.

[0016] Compared with the prior art, the advantages of the present invention are: This invention fully covers the six core aspects of fire scene investigation requirements: site protection, on-site investigation, site inspection, preliminary examination, detailed examination, and inspection and identification, ensuring that the training content is consistent with legal procedures. The design incorporates a six-stage data linkage mechanism, where the quality of each stage directly impacts the evidence status and investigation difficulty of the next. For example, violations of scene protection regulations can lead to evidence contamination, obscuring key clues during site inspections; on-site investigation information guides the focus of the investigation; preliminary investigation conclusions influence the direction of the detailed investigation. This linkage mechanism simulates the causal relationships of real investigations, cultivating trainees' procedural awareness and holistic thinking. The system drives the dynamic generation and intelligent display of case clues through a large language model, avoiding the rigidity of pre-set scripts. Each training session presents different clue distributions and evidence statuses, achieving personalized training for each individual. When trainees fail to set the warning line correctly or do not wear protective equipment, the system dynamically triggers evidence disappearance or status change events, intuitively demonstrating the legal consequences of violations. Based on the standardization of the sampling locations and techniques used by the trainees, the large language model generates different identification report contents, simulating detection failures caused by improper operation in real cases, thereby training the trainees' rigor and professional judgment. The on-site investigation phase employs a large language model-driven virtual witness, generating differentiated responses based on trainees' questioning skills. This simulates the difficulty of obtaining information in real questioning, cultivating trainees' questioning strategies and contradiction identification abilities. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0018] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0019] It should be noted that in the description of this invention, the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. These terms are used only for the convenience of describing the invention and for 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 limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. The terms "horizontal," "vertical," and "suspended," etc., do not indicate that the component must be absolutely horizontal or suspended, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0020] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] like Figure 1 As shown, a virtual simulation training method and system for fire-related case investigation based on a large language model includes the following steps: Step S1: Construction of Virtual Scene and Case Metadata for Fire-Related Cases Based on a database of real fire-related cases, case metadata (including building structure, distribution of burn marks, location of physical evidence, characteristics of the ignition point, spread path, and information on relevant personnel) is extracted. A 3D virtual fire scene is generated using a VR engine. A large-scale language model dynamically adjusts scene parameters according to training objectives (electrical fires, arson with oxidizers, spontaneous combustion fires, and fires caused by careless use of fire), generating differentiated case configurations, including initial information such as ignition time, location of the fire, suspected individuals, and eyewitnesses. The training method for the large-scale language model includes constructing training datasets for multiple types of fire-related cases and a phased fine-tuning strategy. The training dataset covers typical scenarios such as electrical fires (short circuits, overloads, poor contact), arson with oxidizers (flammable liquids such as gasoline and alcohol), spontaneous combustion fires (oxidation and exothermic reactions of accumulated materials), and fires caused by careless use of fire, and includes complete case descriptions, on-site investigation records, chains of evidence, and expert conclusions. During model training, general semantic understanding is first pre-trained, and then fine-tuned through instructions to enhance the task capabilities of "case generation—evidence reasoning—conclusion output." By combining reinforcement learning based on human feedback to optimize the rationality and consistency of reasoning, and setting differentiated parameter adjustment rules for different fire types, including evidence weight allocation (e.g., increasing the weight of circuit melting marks and load anomalies in electrical fires, increasing the weight of accelerant residue and human behavior in arson cases, and increasing the weight of environmental temperature and humidity and material accumulation characteristics in spontaneous combustion fires), clue generation probability (controlling the explicitness and implicitness of key evidence), proportion of interfering information (introducing misleading testimony or irrelevant traces), and ignition point distribution strategy, the authenticity, complexity, and trainability of case generation in virtual scenarios are achieved. Step S2: Modeling Trainee Competency Profiles The system collects basic information about trainees (job level, years of experience, professional background), historical training records, and competency assessment data to construct a trainee competency map. A large-scale language model analyzes competency gaps (such as on-site protection standardization, investigation completeness, and interrogation skills) to generate personalized training path suggestions and dynamically recommend case types of appropriate difficulty. A multi-dimensional feature vectorization method is used to map trainees' various competency indicators into a unified vector space, and clustering and hierarchical algorithms are used to classify competency levels. The large-scale language model performs causal analysis and pattern recognition based on historical behavioral data and assessment results to pinpoint competency gaps and, combined with training objectives, outputs personalized training paths and difficulty matching strategies to achieve dynamic adaptation. Step S3: Protecting the Scene Trainees perform actions such as setting up warning tapes, donning protective gear, controlling personnel entry, and initially securing evidence in a VR environment. The system monitors the compliance of these actions in real time through spatial coordinate detection, gesture recognition, and object collision detection. When violations are detected, the large language model dynamically triggers evidence contamination events (such as virtual crowds intruding and damaging footprints, lack of shoe covers causing blurred traces, and insufficient warning range leading to the movement of physical evidence). The change in evidence status is simultaneously transmitted to subsequent on-site inspections and examinations, affecting the visibility of clues and the identification results. Based on spatial boundary detection and behavioral rule matching, combined with key area coverage calculation, operation sequence verification, and action standard comparison, the system makes real-time judgments on trainee operations. When violations are detected, a corresponding evidence contamination model is generated through an event triggering mechanism, and the contamination propagation path and impact range are recorded. Step S4: On-site investigation Trainees engage in dialogues with virtual witnesses, parties involved, discoverers, and firefighters in a VR environment. A large language model drives the virtual characters, generating differentiated responses based on the trainees' questioning skills (open / closed-ended questions, timing of follow-up questions, contradiction identification, etc.). The questioning content is automatically converted into written records, and the large language model evaluates the completeness of the questioning and the acquisition of key information. The information obtained from the investigation dynamically guides the focus of subsequent investigations; for example, if a witness mentions "abnormal sounds before the fire," the weight of electrical circuit investigation is increased. The large language model's recognition mechanism, based on natural language understanding and dialogue state management, performs semantic classification and intent recognition on trainees' questions, determining the question type and logical structure. Consistent responses are generated through contextual memory and multi-turn dialogue reasoning, and a contradiction detection algorithm is used to compare logical conflicts between different testimonies, thereby evaluating the quality of the questioning and the completeness of the information. Step S5: Site Inspection Trainees wear VR headsets for panoramic patrols, with the system capturing patrol trajectories, gaze focus, and dwell time. A large language model analyzes patrol coverage and dwell time in key areas, dynamically adjusting clue visibility based on information gathered from on-site investigations (e.g., correctly identifying a "V"-shaped burn mark triggers the next clue; virtual instructors provide voice guidance when key points are missed). Standardized draft on-site investigation records are automatically generated. Patrol findings influence the priority of preliminary and detailed investigations. Statistical analysis of trainees' gaze trajectories, spatial paths, and dwell time calculates on-site coverage and focus on key areas. Dynamic adjustment logic, based on prior investigation information and current patrol behavior, controls the visibility of clues in real-time and assists trainees in identifying key areas through guidance mechanisms. Step S6: Preliminary Inspection Trainees conduct preliminary fire scene investigations to determine the fire's origin, spread direction, and key investigation areas. The system records the trainees' investigation paths and judgment criteria. A large-scale language model compares the trainees' judgments with the actual fire location; if the deviation is too large, a re-investigation is prompted. The preliminary investigation conclusions serve as input for detailed investigations; incorrect judgments will lead to deviations in the detailed investigation direction, increasing the difficulty of solving the case. Data is stored through time-series trajectory recordings and key operation node logs, including movement paths, stopping points, and decision-making criteria. The large-scale language model uses spatial deviation calculation and evidence chain matching algorithms to compare the trainees' judgments with a standard fire model, evaluating accuracy from both spatial location error and logical consistency perspectives. Step S7: Detailed Inspection Trainees conduct microscopic fire scene investigations, including the virtual extraction and packaging of physical evidence such as charcoal, soil, accelerant residues, and electrical melt marks; the system records the investigation path, evidence location, and photographic records; the quality and completeness of the extracted evidence directly affect the credibility of the identification results in the inspection and appraisal process; a large language model is used to evaluate the standardization and completeness of the investigation records; based on fire trace analysis and physical evidence analysis methods, the degree of charring is used to determine the combustion temperature, the morphology of melt marks is used to identify electrical fault characteristics, the detection of chemical residues is used to determine the use of accelerants, and the spatial distribution is combined to reconstruct the fire spread path; Step S8: Inspection and Identification Process The system verifies the representativeness of the sampling locations and the standardization of the sampling methods for the physical evidence submitted by trainees; the big data model generates differentiated identification reports based on operational quality (standard operations generate accurate reports, while non-standard operations generate contaminated / invalid results), simulating the feedback mechanism of a real laboratory; trainees need to interpret the identification data and form a final investigation conclusion; the big data model comprehensively evaluates the credibility of the conclusion by integrating the operational records of the first five stages, and if violations of on-site protection lead to the loss of key evidence, the conclusion is deemed insufficiently supported; the identification result generation logic is constructed based on the physical evidence quality score, sampling standardization, and contamination status, and identification reports with different levels of credibility are generated through a combination of rule constraints and probability models, simulating the uncertainty and delayed feedback mechanism in the real testing process; Step S9: Training Assessment and Dynamic Difficulty Adjustment The large language model aggregates operational data from six stages to generate a multi-dimensional evaluation report (including on-site protection standardization, investigative skills proficiency, patrol coverage, examination completeness, evidence chain logic, and accuracy of expert interpretation). Based on the Elo scoring algorithm, trainees' competency values ​​are calculated, dynamically adjusting subsequent training case complexity, clue concealment, time pressure parameters, and the amount of interfering information. The evaluation index system includes dimensions such as operational standardization score, information acquisition completeness, path coverage, evidence chain closure, and conclusion consistency. The Elo scoring algorithm uses the initial score as a baseline and dynamically adjusts competency values ​​based on trainee performance, incorporating a K-value adjustment mechanism to reflect changes in learning at different stages. The dynamic parameter adjustment rules automatically adjust case complexity, clue visibility, and interference intensity based on the score range, achieving adaptive training. Step S10: Training Data Archiving and Capability Tracking Training process data, evaluation results, and competency maps are stored in training archives to support subsequent tracking and comparative analysis. A federated learning architecture ensures data privacy and security, and supports cross-institutional competency certification. Training data can be used to optimize the reasoning capabilities of large language models and the quality of case generation. The federated learning architecture employs a distributed training model, where each institution completes model training locally, uploading only model parameters or gradients to the central server. Aggregation updates are performed using a federated averaging algorithm (FedAvg), achieving model optimization and cross-institutional collaboration without sharing original data.

[0022] Example 1: Training scenario for fire investigation in residential buildings 1. Construction of virtual scenarios and case metadata for fire-related cases Based on the training objective of "fire caused by careless use of fire," the large language model used a VR engine to construct a fire scene in a three-story residential building, including key areas such as cigarette butts, lighters, burned appliances, bedrooms, living rooms, and kitchens. Initial case information: alarm time 06:00, fire area 20 square meters, 1 fatality.

[0023] 2. On-site protection Trainees are required to: set up a cordon (covering the room where the fire started), wear protective equipment (shoe covers, gloves, head coverings), and initially secure easily perishable evidence (such as the condition of doors and windows, smoke traces); if the trainee's cordon is insufficient, the large language model can trigger a "media reporter intrusion and filming" event; if shoe covers are not worn, the system will mark "center site footprint contamination", and key footprint clues will not be visible in subsequent patrols.

[0024] 3. On-site investigation phase Students ask questions in the VR environment: Homeowner: "The deceased's smoking habits and lifestyle" Neighbor: "The situation when the fire was discovered" Property management: "Community security, fire safety facilities, resident information" A large language model drives virtual avatars, generating responses based on trainees' questioning techniques. For example, if a trainee asks, "Was there anything unusual before the fire?" (open-ended), the homeowner might answer, "There were sparks from the air conditioner socket a few days ago." If the trainee asks, "Has there been any instance of cigarettes or lighters igniting items?" (closed-ended leading question), the homeowner might hesitate or give a vague answer. Information obtained from the investigation dynamically guides subsequent inspections: if the homeowner mentions "burning holes in clothing collars and blankets while smoking," the inspection of careless use of fire is given higher weight.

[0025] 4. On-site inspection Trainees wearing VR headsets survey the site, with the system tracking their gaze. When a trainee focuses on the ash area on the ground near the west side of the bed for more than 5 seconds, "trace evidence" is triggered. If the trainee misses the headboard area, a virtual instructor provides a voice prompt: "Please check the burning and spread traces at the headboard area." After the survey is completed, the system automatically generates a draft investigation report. Preliminary findings from the survey indicate: the headboard was burned, with the western side more severely damaged than the eastern side; the inner side of the footboard was charred, while the outer side was relatively intact; and there was localized burn-through on the western side, thus pinpointing this area as the origin of the fire.

[0026] 5. Preliminary inspection stage The trainees conducted a macroscopic inspection and determined the location of the fire: the northwest bedroom; they also identified the key inspection area: the west side of the bed. The large language model compares the student's judgment with the actual location of the fire; if the student judges it to be the living room (incorrect), the system will prompt "The shape of the burn marks does not match the judgment, and it is recommended to re-inspect"; if the student insists on the incorrect judgment, the detailed investigation will increase the difficulty of solving the case (key evidence will be more difficult to find).

[0027] 6. Detailed inspection process Trainees conducted a microscopic examination, collecting cigarette butts, lighters, and residues of accelerants from the ground (to rule out arson); and photographed and recorded the location of all physical evidence. The system records the inspection route, the location of physical evidence, and photographic records. The quality and completeness of the evidence collected during the inspection directly affect the credibility of the identification results in the examination and appraisal process.

[0028] 7. Inspection and Appraisal Process The system verifies the sampling procedures, which include: placing evidence tags, photographing the evidence, and finally extracting the evidence; and placing the evidence in an evidence bag and sealing it with a label. After standardized operation, the large language model generates an identification report: "The combustion residue on the west side of the bed was extracted and sent for testing. The results showed no gasoline, kerosene, diesel, or heavy mineral oil." If the sample was contaminated, the report states: "The sample was contaminated, and no valid conclusion can be drawn." Trainees are required to interpret the identification data and form a final investigation conclusion; the large language model then evaluates the credibility of the conclusion based on the operational records of the first five steps.

[0029] 8. Assessment Report After the training, the large language model generates an evaluation report: Link Score Comments Protect the scene 80 points There were omissions in the wearing of protective equipment. On-site investigation 95 points Skilled in questioning and able to obtain sufficient key information. Site inspection 90 points 95% coverage, with one secondary area missing. Preliminary investigation 100 points Accurate identification of the fire location Detailed inspection 100 points The investigation records were complete and the physical evidence was collected in accordance with regulations. Inspection and identification 100 points Sampling procedures are standardized, and identification and interpretation are accurate. Overall Score 94 points Ability Level: Proficient The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept should fall within the protection scope of the present invention. All technical contents for which protection is sought in this invention are fully described in the claims.

Claims

1. A virtual simulation training method for fire-related case investigation driven by a large language model, characterized in that, Includes the following steps: Step S1: Construction of virtual scene and case metadata for fire-related cases. Extract case metadata based on real fire-related case database, generate a three-dimensional virtual fire scene through VR engine, and dynamically adjust scene parameters according to training objectives using large language model. Step S2: Model the trainee's competency profile, collect basic trainee information and historical training records, construct the trainee competency graph, and generate personalized training path suggestions from the large language model; Step S3, the scene protection step, monitors the setting of the trainee's warning tape, wearing of protective equipment, and personnel entry control operations. When a violation is detected, the big language model dynamically triggers an evidence contamination event and records the evidence status change. The evidence status change is synchronously transmitted to the subsequent steps. Step S4, on-site investigation: The large language model drives the virtual witness and party roles, generating differentiated responses based on the trainees' questioning skills, and the information obtained from the investigation dynamically guides the focus of subsequent investigations. Step S5, the site inspection phase, captures the trainees' inspection trajectory and focus of vision, analyzes the survey coverage using a large language model, and dynamically adjusts the visibility of clues based on the on-site investigation information; Step S6, preliminary investigation stage: trainees conduct a macroscopic investigation of the fire scene to determine the location of the fire and the direction of its spread. The preliminary investigation conclusions serve as input conditions for the detailed investigation. Step S7, detailed inspection process, records the trainee's inspection route, physical evidence location and photo records. The quality and completeness of the evidence collected during the inspection directly affect the credibility of the identification results in the inspection and identification process. Step S8, the inspection and identification process, verifies the representativeness of the sampling location and the standardization of the sampling method of the physical evidence submitted by the trainee. The big language model generates a differentiated identification report based on the operation quality, and evaluates the credibility of the conclusion by combining the operation records of the first five steps. Step S9, Training Assessment and Dynamic Difficulty Adjustment: The large language model summarizes the operational data from the six stages to generate a multi-dimensional assessment report, and dynamically adjusts the subsequent training difficulty parameters based on the scoring algorithm. Step S10, Training Data Archiving and Competency Tracking, stores training process data and evaluation results in training archives to support cross-institutional competency certification.

2. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, The evidence contamination events in step S3 include: virtual crowds intruding and damaging footprints, lack of protective equipment leading to blurred traces, and insufficient warning range leading to the movement of physical evidence. Changes in the status of evidence simultaneously affect the visibility of clues and the identification results in the on-site inspection in step S5 and the inspection and identification process in step S8.

3. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, The differentiated response in step S4 includes: obtaining detailed information through open-ended questions, obtaining ambiguous answers through closed-ended leading questions, triggering changes in the character's emotions through contradictory follow-up questions, and automatically converting the inquiry content into a written record, with the completeness of the inquiry being evaluated by a large language model; the information obtained in step S4 dynamically guides the focus of subsequent investigations, including increasing the weight of electrical circuit investigation when a witness mentions abnormal sounds, increasing the weight of trace investigation when a witness mentions suspicious persons, and increasing the weight of combustion accelerant investigation when a witness mentions abnormal odors.

4. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, In step S5, the dynamic adjustment of clue visibility adopts a conditional trigger mechanism. When the trainee's gaze stays on the key trace area for more than a preset time and the exploration path covers the key area, the next level of clues is triggered to appear. When the trainee misses the key area, the large language model asks guiding questions in the role of a virtual instructor instead of directly telling the answer.

5. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, The preliminary investigation conclusion in step S6 serves as the input condition for the detailed investigation in step S7. If the preliminary investigation makes an incorrect judgment on the location of the fire, the difficulty in discovering key evidence in the detailed investigation will increase, simulating the difficulty in solving the case caused by directional deviation in a real investigation.

6. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, The differential identification report in step S8 includes: generating accurate identification conclusions through standardized operations, generating misleading data due to sampling location deviations, generating contamination labeling reports due to non-standard sampling methods, and simulating the uncertainty of real laboratory testing; the large language model in step S8 comprehensively evaluates the credibility of the conclusions based on the operation records of the first five steps, including: determining insufficient support for the conclusions if key evidence is missing due to violations of on-site protection regulations, determining incomplete investigation if on-site investigation information is not fully utilized, and determining unreliable conclusions if the chain of evidence for on-site investigation is incomplete.

7. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, The multi-dimensional evaluation report in step S9 includes scores for on-site protection standardization, investigation skills proficiency, patrol coverage, inspection integrity, evidence chain logic, and identification and interpretation accuracy. The weights of each dimension are configurable.

8. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, In step S9, the scoring algorithm is dynamically adjusted using the Elo scoring algorithm. The ability value is calculated based on the trainee's historical training performance. The higher the ability value, the higher the concealment of clues in subsequent training cases, the greater the time pressure, the more interference information, and the higher the complexity of the linkage between links.

9. The virtual simulation training method for fire-related case investigation based on a large language model as described in claim 1, characterized in that, In step S10, the training data archiving adopts a federated learning architecture. Each training institution stores student data locally and only uploads encrypted capability graph parameters to the central server to ensure data privacy and security.

10. A virtual simulation training system for investigating fire-related cases, implementing the method of any one of claims 1-12, characterized in that, include: The VR interaction module includes a head-mounted virtual reality display, a gesture sensor, an eye-tracking module, and a voice acquisition port; The large language model inference engine includes a domain fine-tuning module, a retrieval enhancement generation module, a multi-role dialogue management module, and a process linkage control module; The six-step process control module includes the following sub-modules: site protection, site investigation, site inspection, preliminary investigation, detailed investigation, and inspection and appraisal. The assessment and feedback module includes a multi-dimensional scoring submodule, a dynamic difficulty adjustment submodule, and an assessment report generation submodule; The data management module includes a trainee competency graph database, a case meta-database, a training record database, and an evidence status tracking database. The large language model inference engine has been fine-tuned using professional corpus of fire-related cases. The training data includes fire investigation textbooks, typical case files, physical evidence identification reports, court judgments, and interrogation transcripts. The fine-tuning adopts LoRA low-rank adaptation technology. The six-step process control module supports step jumps and repeated training. Trainees can choose to return to the previous step and repeat the operation after any step is completed. The system records the differences in operation of each attempt and includes them in the evaluation, but the evidence contamination state is irreversible. The evidence status tracking database records the status change history of each piece of evidence, including the initial status, pollution event, pollution cause, and impact stage, supporting the tracing of evidence change processes during training and review.

Images