Methods, systems, equipment, and media for dynamic risk assessment of chemical defense command based on multimodal data fusion

By integrating structured perception and audio-visual interaction data through multimodal data fusion technology, a multimodal feature set is constructed and input into a time-series risk model to generate command and decision instructions. This solves the problems of accuracy and data security in chemical defense risk assessment and realizes efficient emergency command and full-domain collaborative scheduling for chemical defense operations.

CN123117845APending Publication Date: 2026-11-17BEIJING ZHONGSHURUIZHI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611629008.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing chemical defense risk assessment technologies rely on single-modal data analysis, failing to effectively combine the inherent correlations between different types of data. This leads to biased risk level judgments, difficulty in accurately predicting the spread of toxic agents, and a disconnect between risk assessment and emergency command. Data storage is easily tampered with or lost, and the dispatching model struggles to balance rapid response with overall coordination.

Method used

A multimodal data fusion method is adopted to integrate the cross-modal feature correlation degree and temporal correlation feature of structured perception data and unstructured audio-visual interaction data to construct a multimodal fusion feature set. This set is then input into a temporal risk evolution model to generate command and decision instructions. Data storage and scheduling are achieved through a cloud-edge collaborative architecture.

Benefits of technology

Accurately reconstruct the evolution of chemical defense risks, improve the accuracy of predicting the spread of toxic agents, achieve linkage between risk assessment and emergency command, ensure data security and dispatch efficiency, and meet the needs of rapid response and overall coordination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117845A_ABST
    Figure CN123117845A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of chemical defense assessment and artificial intelligence technology, specifically providing a method, system, equipment, and medium for dynamic risk assessment of chemical defense command based on multimodal data fusion. The method includes: collecting and preprocessing structured and audio-visual interactive multi-source chemical defense data from the work area; extracting two types of features and calculating cross-modal feature correlation degrees according to preset risk calculation rules; mining cross-modal temporal correlation features based on the correlation degrees and integrating all features to construct a multimodal fusion feature set; inputting the feature set into a temporal risk evolution model to output regional and personnel risk levels and toxic agent spread prediction results; generating command instructions based on the prediction results; storing the prediction data, instructions, timestamps, and personnel identity information on the blockchain; and conducting chemical defense command and dispatch through a cloud-edge collaborative architecture. This invention constructs an integrated assessment and response solution, utilizing multimodal temporal fusion to improve the accuracy of dynamic risk assessment and meet the needs of chemical defense operation tracing and full-domain emergency control.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of chemical defense assessment and artificial intelligence technology, specifically to a method, system, equipment, and medium for dynamic assessment of chemical defense command risks based on multimodal data fusion. Background Technology

[0002] Chemical defense operations fall under the category of high-risk safety production. Accidents involving the leakage and spread of toxic agents at the work site are highly unpredictable and have a wide-ranging impact, directly threatening the personal safety of workers and potentially triggering regional environmental risks. Therefore, there is a rigid demand for accurate dynamic risk assessment, efficient emergency command and response, and comprehensive accountability throughout the entire process. With the development of the Internet of Things and intelligent analytics, the industry has gradually adopted data-driven approaches to assist in chemical defense risk assessment. However, existing technical solutions still have key shortcomings and are difficult to adapt to the comprehensive management and control requirements of complex scenarios.

[0003] Existing chemical defense risk assessment methods suffer from core technical deficiencies: they largely rely on single-modal data for analysis. Even when multi-source data is introduced, key parameters are extracted independently for each data type, without considering the actual risk patterns at the chemical defense site to analyze the inherent correlations between different data types or linking the interconnected changes in various data over time. This makes it difficult for existing methods to accurately reflect the overall process of the gradual development of on-site risks and the continuous spread of toxic agents. Furthermore, the assessment of risk levels is prone to significant deviations, resulting in insufficient accuracy in predicting toxic agent spread trends and failing to provide reliable decision-making basis for subsequent emergency command.

[0004] Furthermore, existing technologies fail to effectively integrate risk assessment, emergency command, data retention, and on-site dispatch, resulting in a disconnect between these processes. On one hand, the system can only output risk analysis results and cannot generate emergency response instructions based on the assessment results, leading to a gap between risk analysis and on-site operations. On the other hand, the industry generally uses a centralized approach to data storage, which exposes data to the risk of tampering or loss. Important information such as risk conclusions, response instructions, data collection times, and relevant personnel information is not uniformly archived, making it difficult to conduct post-incident review and assign responsibility in the event of a safety incident. Meanwhile, traditional early warning and dispatch models also have shortcomings: relying entirely on remote platforms leads to response delays, while operating only locally cannot coordinate the entire operational area, failing to simultaneously meet the needs of rapid emergency response and large-scale collaborative handling, thus lowering the overall safety management capabilities of chemical defense operations.

[0005] To address these issues, the present invention provides a method, system, equipment, and medium for dynamic risk assessment of chemical defense command based on multimodal data fusion. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method, system, equipment, and medium for dynamic risk assessment of chemical defense command based on multimodal data fusion, in order to solve the problems in existing technologies.

[0007] One embodiment of the present invention provides a method for dynamic risk assessment of chemical defense command based on multimodal data fusion, comprising the following steps: Collect multi-source data on chemical defense operations in the chemical defense area and perform data preprocessing. The multi-source data on chemical defense includes structured sensing data and unstructured audio-visual interaction data. Based on the preprocessed chemical defense multi-source data, structured perception features and audio-visual interaction features are extracted, and the cross-modal feature correlation degree of structured perception features and audio-visual interaction features is calculated based on the preset risk calculation rules. The cross-modal feature correlation degree is used to extract cross-modal temporal correlation features between structured perception features and audio-visual interaction features. The structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features are integrated to construct a multimodal fusion feature set. The multimodal fusion feature set is input into a pre-trained time-series risk evolution model to calculate the dynamic temporal correlation of risks and the nonlinear coupling relationship of multiple factors, and outputs prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. Based on the prediction results, corresponding command and decision instructions are generated. The prediction results, command and decision instructions, collection timestamps of chemical defense multi-source data, and identity information corresponding to the individual risk level of personnel are written into the chemical defense command alliance chain. Chemical defense command and dispatch are completed through a preset cloud-edge collaborative architecture.

[0008] This application also relates to a dynamic risk assessment system for chemical defense command based on multimodal data fusion, including: The data acquisition module is used to collect multi-source data on chemical defense operations in the chemical defense area and perform data preprocessing. The multi-source data on chemical defense includes structured sensing data and unstructured audio-visual interactive data. The feature extraction module is used to extract structured perception features and audio-visual interaction features based on the preprocessed chemical defense multi-source data, and to calculate the cross-modal feature correlation degree of structured perception features and audio-visual interaction features based on preset risk calculation rules. The feature integration module is used to extract cross-modal temporal correlation features between structured perception features and audio-visual interaction features using the cross-modal feature correlation degree, and to integrate structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features to construct a multimodal fusion feature set; The risk prediction module is used to input the multimodal fusion feature set into a pre-trained time-series risk evolution model, calculate the dynamic temporal correlation of risks and the nonlinear coupling relationship of multiple factors, and output prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. The command and dispatch module is used to generate corresponding command and decision instructions based on the prediction results. It writes the prediction results, command and decision instructions, the collection timestamps of chemical defense multi-source data, and the identity information corresponding to the individual risk level of personnel into the chemical defense command alliance chain, and completes chemical defense command and dispatch through a preset cloud-edge collaborative architecture.

[0009] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for dynamic risk assessment of chemical defense command based on multimodal data fusion.

[0010] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for dynamic risk assessment of chemical defense command based on multimodal data fusion.

[0011] The above embodiments provide a method, system, equipment, and medium for dynamic risk assessment of chemical defense command based on multimodal data fusion, which has the following beneficial effects: 1. By integrating structured perception data with unstructured audio-visual interactive data for comprehensive analysis, it is possible to uncover the inherent correlations and temporal linkages of different types of data, accurately reconstruct the complete process of on-site risk evolution and toxic agent diffusion, effectively reduce the bias in risk level judgment, improve the accuracy of toxic agent diffusion trend prediction, provide reliable assessment basis for emergency command, and overcome the shortcomings of insufficient accuracy in existing chemical defense risk assessment.

[0012] 2. This invention integrates the entire business chain of risk assessment, emergency command generation, data storage, and on-site dispatch, enabling linkage between risk assessment and on-site response; and through the chemical defense command alliance chain, it uniformly archives various core data, avoiding the problems of easy tampering and loss of centralized storage, and facilitating accident review and responsibility determination; at the same time, it utilizes a cloud-edge collaborative dispatch mode to balance rapid on-site response and overall coordination and disposal capabilities, solving multiple problems of existing technologies such as business fragmentation, low data credibility, and shortcomings in dispatch mode. Attached Figure Description

[0013] Figure 1 A flowchart illustrating a dynamic risk assessment method for chemical defense command based on multimodal data fusion, provided in an embodiment of the present invention; Figure 2This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0015] Reference Figure 1 One embodiment of the present invention provides a method for dynamic risk assessment of chemical defense command based on multimodal data fusion, comprising the following steps: S10. Collect chemical protection multi-source data of the chemical protection operation area and perform data preprocessing. The chemical protection multi-source data includes structured perception data and unstructured audio-visual interaction data. S20. Based on the preprocessed chemical defense multi-source data, extract structured perception features and audio-visual interaction features, and calculate the cross-modal feature correlation degree of structured perception features and audio-visual interaction features based on preset risk calculation rules. S30. Utilize the cross-modal feature correlation degree to extract cross-modal temporal correlation features between structured perception features and audiovisual interaction features, integrate structured perception features, audiovisual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features, and construct a multimodal fusion feature set; S40. Input the multimodal fusion feature set into the pre-trained time-series risk evolution model, calculate the dynamic temporal correlation of risk and the nonlinear coupling relationship of multiple factors, and output the prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. S50. Generate corresponding command and decision instructions based on the prediction results, write the prediction results, command and decision instructions, the collection timestamps of chemical defense multi-source data, and the identity information corresponding to the individual risk level of personnel into the chemical defense command alliance chain, and complete chemical defense command and dispatch through the preset cloud-edge collaborative architecture.

[0016] In this embodiment, step S10 is used to complete the full-domain collection and standardized preprocessing of multi-source raw data at the chemical defense operation site. The collected multi-source data is divided into two categories: structured sensing data and unstructured audio-visual interactive data. Structured sensing data consists of time-series numerical information collected by various sensors at the operation site, such as real-time concentration of toxic agents, regional wind speed and direction, personnel location coordinates, and ambient temperature and humidity. Unstructured audio-visual interactive data consists of multimedia images and audio materials collected by cameras and recording devices, including leakage noises, personnel calls for help, images of protective equipment being donned, and images of toxic agent leaks. These two types of data respectively represent on-site physical environment indicators and abnormal behavior information of personnel and equipment. In actual implementation, gas sensors deployed throughout the overall operation scenario are used... Basic data acquisition equipment such as sensors, positioning terminals, and meteorological data acquisition devices continuously collect structured sensing data. Simultaneously, monitoring and audio recording equipment are used to record audio and video materials carrying various on-site anomaly information as unstructured audio-visual interactive data, and all collected unstructured audio-visual interactive data are uniformly labeled with timestamps. Then, data processing such as missing value completion, anomaly error removal, and sampling frequency unification are performed on the structured data, and data processing such as audio noise reduction, invalid segment deletion, and video frame time sequence alignment are performed on the unstructured audio-visual interactive data to eliminate data noise and invalid information caused by equipment interference. This overcomes the problem of missing data dimensions caused by the existing solution of only collecting single-type sensor data, and avoids the defect of subsequent risk misjudgment caused by noise interference when directly using uncleaned raw data for analysis.

[0017] Step S20 is used to extract corresponding risk features from the two types of preprocessed multi-source data, and quantify the correlation between the two types of features according to the preset risk calculation rules adapted to the chemical defense scenario. Among them, the structured perception features are converted from the standardized sensor numerical data, mainly including parameter features reflecting changes in on-site physical risks such as toxic agent concentration gradient, meteorological changes, and personnel distance. The audio-visual interaction features are extracted from the processed audio and video materials, mainly including multimedia features characterizing sudden risk events on-site, such as dangerous semantic keywords, leakage-related voiceprint fluctuations, and abnormal personnel actions. The two types of features reflect the risks of chemical defense operations from the two dimensions of objective physical indicators and on-site abnormal events, respectively. In actual implementation, the structured perception features and audio-visual interaction features are first sliced ​​into asymmetric time windows with time delay offset. The process matches the physical delay of toxic agent propagation. Within a time window, it extracts the vectors of physical quantity mutations and the vectors of abnormal events, and then constructs a physical causal matrix for chemical defense risks. This matrix predefines the physical mapping relationship between the vectors of abnormal events and the vectors of physical quantity mutations under a preset operational scenario. Based on this matrix, the two sets of vectors are weighted by causal confidence. Corresponding association thresholds are assigned to two types of scenarios: dangerous event matching and environmental steady state. Finally, the cross-modal feature association degree is quantitatively output by combining the confidence score. This overcomes the problems of existing technologies that only use symmetrical time windows and general mathematical similarity to calculate feature association, do not consider the delay of toxic agent propagation, and lack chemical defense-specific physical causal constraints. It also avoids the association calculation logic from deviating from the actual risk evolution law on site, and can improve the matching degree between feature association results and actual risks.

[0018] Step S30 is used to mine the dynamic temporal correlation patterns of the two types of features based on the quantified cross-modal feature correlation, and integrate multi-dimensional risk characterization information to form a unified fusion feature set. Among them, the cross-modal temporal correlation feature is the core feature reflecting the dynamic coupling relationship between structured perception features and audio-visual interaction features in the time dimension. Specifically, it is manifested as the temporal response matching relationship between the evolution of physical quantities and the evolution of abnormal events. The four types of integrated features comprehensively cover chemical defense operation risk information from four dimensions: static physical indicators, on-site event characterization, cross-modal correlation strength, and dynamic temporal coupling. In actual implementation, the cross-modal feature correlation is used as the benchmark to perform temporal segmentation and alignment of structured perception features and audio-visual interaction features. Continuous time segments with correlation higher than a preset threshold are selected as risk evolution intervals. Within these intervals, structured perception features are extracted respectively. The physical quantity temporal evolution vector of the feature and the abnormal behavior temporal evolution vector of the audio-visual interaction feature are used to construct a cross-modal time delay coupling matrix. This matrix defines the dynamic coupling relationship between the two types of temporal evolution vectors under different preset time delay offset levels. Based on this matrix, the dynamic cross-correlation peak between the two types of vectors is calculated and determined as the cross-modal time-series correlation feature. Finally, the structured perception feature, audio-visual interaction feature, cross-modal feature correlation degree and cross-modal time-series correlation feature are integrated in a unified data format to form a multimodal fusion feature set. This overcomes the problems of existing technologies that simply splice multimodal features, fail to explore cross-modal time-series correlation rules, and lack dynamic coupling information in fused features. It also avoids the defects of single-dimensional features or unrelated integrated features that cannot fully represent the risk evolution process, and improves the completeness and effectiveness of the input features of the subsequent risk assessment model.

[0019] Step S40 involves inputting the integrated multimodal fusion feature set into a pre-trained temporal risk evolution model. The model analyzes the two core risk relationships and simultaneously outputs three different risk prediction results. The temporal risk evolution model is a dedicated prediction model with a multi-branch structure and incorporates constraints related to the dynamics of toxic agent diffusion. The risk dynamic temporal correlation represents the order of responses and evolution rate differences of different modal features over time. The multi-factor nonlinear coupling relationship represents the comprehensive impact of the synergistic effects of multiple features on the formation of on-site risks. Regional risk level, individual personnel risk level, and toxic agent diffusion trend correspond to the three risk assessment dimensions: the entire work area, on-site personnel, and spatial diffusion evolution, respectively. In practice, the multimodal fusion feature set is first input into the temporal risk evolution model. The model's internal temporal coding module performs feature extraction and temporal evolution calculations to generate coded feature vectors. Then, the temporal coding module captures the order of responses and evolution rate differences of each temporal component. The cross-modal feature correlation degree is used as the confidence threshold for the current time window to adjust the temporal response. The effective weights of the sequential relationship are used to quantify the dynamic temporal correlation of risks. Then, the synergistic correlation between structured perception features, audio-visual interaction features, cross-modal feature correlation degree, and cross-modal temporal correlation features are analyzed to construct feature coupling weights. The sequential relationship of temporal response and the difference in evolution rate are mapped to multi-source feature contributions according to the feature coupling weights. The multi-source feature contributions are fused through the gating mechanism of the temporal coding module to quantify the nonlinear coupling relationship of multiple factors. Finally, based on the encoded feature vector and the feature coupling weights, the calculation is completed through the three independent branches of the model. The corresponding hierarchical mapping rules are matched to output the regional risk level and the individual risk level of personnel. At the same time, the spatiotemporal diffusion trend is deduced to obtain the toxic agent diffusion trend within a preset time. This overcomes the problems of existing technologies that use single-branch temporal models without physical constraints, such as the inability to synchronously analyze the temporal response differences and multi-feature coupling effects, the inability to output the deduction results of regional, personnel risks and toxic agent diffusion in different dimensions, and the tendency of predicted values ​​to exceed the real physical boundaries of the site. This makes the prediction results of various risks fit the objective evolution law of chemical leakage scenarios.

[0020] This overcomes the problems of existing single-branch time series models without physical constraints, such as the inability to synchronously analyze differences in time series response and the coupling effect of multiple features, the inability to output regions in different dimensions, and the tendency for personnel risk and toxic agent diffusion simulation results and predicted values ​​to easily exceed the actual physical boundaries on site. This allows the prediction results of various risks to conform to the objective evolution law of chemical leakage scenarios.

[0021] Step S50 connects risk prediction results with on-site emergency response, completing command instruction generation, core data security storage, and full-domain collaborative scheduling, forming a complete process of assessment, instruction, execution, and storage. The command decision instructions are practical response plans matched to the on-site risk level and spread trend, including core elements such as evacuation route planning, decontamination area delineation, and emergency personnel deployment. The chemical defense command alliance chain is a distributed data storage architecture used to ensure that data throughout the risk response process is tamper-proof and traceable. The cloud-edge collaborative architecture consists of cloud-based coordination nodes and edge execution nodes, balancing full-domain coordination and rapid on-site response. In actual implementation, based on the regional risk level, individual personnel risk level, and the agent spread trend within a preset timeframe, a preset response plan library is matched to generate targeted command decision instructions, which are then used to generate the predicted data. The results, command and decision instructions, collection timestamps of multi-source chemical defense data, and identity information corresponding to individual personnel risk levels are encapsulated into standardized chemical defense operation data blocks. These blocks are written into the chemical defense command alliance chain for distributed storage, and data access permissions are controlled hierarchically according to roles. Finally, through a pre-set cloud-edge collaborative architecture, command and decision instructions are distributed to edge execution nodes. Simultaneously, the cloud dynamically corrects and reschedules command and decision instructions based on real-time feedback of on-site status data. This overcomes the problems of existing technologies, such as the lack of accurate risk matching in command instruction generation, reliance on centralized storage for core data which is prone to tampering and loss, and the difficulty in balancing overall coordination and on-site response speed in scheduling modes. It ensures the pertinence and timeliness of command instructions, achieves traceability of risk disposal data and delineation of responsibility, and improves the standardization and efficiency of emergency command in chemical defense operations.

[0022] In one feasible embodiment, it is assumed that: a chlorine gas leak occurs suddenly at a chemical protection operation site in a chemical industrial park. Two workers on site fail to detect the leak source in time. One of them smells the irritating odor first and shouts "Chlorine gas leak, evacuate quickly" through a walkie-talkie. At the same time, the sensing equipment near the leak point has detected the abnormal concentration. Step S10: The toxic gas sensors deployed on-site collect chlorine concentration data in real time (initial concentration 0.3). It rose to 1.2 within 5 minutes. ), wind speed in the area collected by meteorological sensors (2.3) The system recorded the coordinates of two workers in conjunction with the wind direction (northwest wind) and the personnel positioning terminal. The monitoring camera recorded a white fog-like diffusion image near the leak point. The audio pickup device simultaneously recorded the personnel's cries for help and abnormal leak sounds. All collected data were uniformly labeled with timestamps. Subsequently, missing values ​​were filled in for the concentration data (data from individual sensor momentary disconnections), abnormal errors were removed, and the audio and video were noise-reduced, and silent segments without abnormalities were deleted. The timing of the video frames and timestamps was aligned. Step S20: Extract features such as chlorine concentration gradient change (300% increase within 5 minutes) and wind speed and direction change from the preprocessed structured data; extract dangerous semantic keywords such as "chlorine leak" and "evacuation" and the voiceprint fluctuation features corresponding to the abnormal sound of leakage from the audio and video; first, perform asymmetric time window slicing with a 0.8-second time delay offset on the two types of features (matching the physical propagation speed of chlorine diffusion); extract the physical quantity mutation vector of exponential concentration increase and the abnormal event sudden vector of "call for help + abnormal sound" within the time window; based on the predefined physical causal matrix of chlorine leak scenario (already associated with the mapping relationship between "leak semantics + concentration surge" and high risk), perform causal confidence weighted calculation on the two sets of vectors, assign the first association threshold (confidence 0.92), and finally quantify the cross-modal feature association degree of 0.91; Step S30: Based on the cross-modal feature correlation degree of 0.91 (higher than the preset threshold of 0.7), select the continuous time segment of the past 5 minutes as the risk evolution interval; extract the time-series evolution vector of the physical quantity of continuously rising chlorine concentration and the time-series evolution vector of abnormal behavior of personnel calling for help + spreading images within this interval; calculate the dynamic cross-correlation peak value of the two types of vectors (0.88) through the cross-modal time delay coupling matrix, and determine it as the cross-modal time-series correlation feature; finally, integrate the structured perception feature, audio-visual interaction feature, cross-modal feature correlation degree (0.91) and cross-modal time-series correlation feature (0.88) to form a multimodal fusion feature set in a unified format; Step S40: Input the fused feature set into the pre-trained temporal risk evolution model. The model generates coded feature vectors through the temporal coding module, capturing the temporal response sequence of "soundprint fluctuation → semantic keywords → concentration increase" (abnormal sounds precede calls for help, and calls for help are synchronized with the concentration surge). The cross-modal feature correlation degree of 0.91 is used as the confidence gate to adjust the weights, and the dynamic temporal correlation of risk is quantified. At the same time, the synergistic correlation of the four types of features is analyzed (wind speed promotes concentration diffusion, and personnel calls for help verify the authenticity of the leak), feature coupling weights are constructed, and the contribution of multi-source features is quantified to obtain the nonlinear coupling relationship of multiple factors. Finally, the model outputs through three branches: the regional risk level is level one (extremely high risk), the individual risk level of the two workers is high risk (less than 10 meters from the leak source), and the toxic agent diffusion trend is preset to spread chlorine gas to a range of 50 meters in the southeast direction within 10 minutes. Step S50: Based on the Level 1 regional risk, the high-risk level of personnel, and the 10-minute spread trend, match the preset disposal plan library and generate a command decision instruction: "Immediately activate the Level 1 evacuation plan, plan an evacuation route from the work site to the northwest (against the wind), designate a 50-meter warning zone around the leak point, and dispatch 2 teams of emergency decontamination personnel with equipment to the site"; encapsulate the instruction, risk prediction results, original data collection timestamps, and the identity information of the 2 workers into a chemical defense operation data block and write it into the chemical defense command alliance chain (only emergency command personnel and safety supervision personnel have access rights); through the cloud-edge collaborative architecture, the instruction is sent to the edge execution nodes on site (handheld terminals of workers, on-site warning equipment), and the cloud receives real-time feedback on the personnel evacuation progress data from the site, dynamically corrects the evacuation route (avoiding temporary obstacles), and completes the overall collaborative dispatch.

[0023] It should be noted that the above scenarios and the related leakage types, data values, time parameters, and associated thresholds are all illustrative examples, used only to assist those skilled in the art in understanding the overall logic of the solution, and do not constitute any limitation on the present invention. In practical applications, the relevant parameters and thresholds can be flexibly adjusted according to the chemical defense operation scenario (such as different types of toxic agents, size of the operation area), equipment configuration, risk level classification standards, etc.

[0024] In one embodiment, step S20, which calculates the cross-modal feature correlation between structured perception features and audiovisual interaction features based on preset risk calculation rules, specifically includes: S201. Perform asymmetric time window slicing processing on the structured perception features and audio-visual interaction features, wherein the time window of the audio-visual interaction features has a preset time delay offset relative to the time window of the structured perception features, so as to match the physical time delay of the spread of toxic agents. S202. Within the asymmetric time window defined by the time delay offset, extract the abnormal event burst vector from the audio-visual interaction features and extract the physical quantity mutation vector from the structured perception features. S203. Construct a physical causal matrix for chemical defense risks, wherein the physical causal matrix for chemical defense risks predefines the physical mapping relationship between the sudden vector of abnormal events and the sudden change vector of physical quantities under a preset operation scenario; S204. Based on the aforementioned chemical defense risk physical causality matrix, perform a causal confidence weighted calculation on the sudden event vector and the physical quantity mutation vector to obtain a causal confidence score: If the abnormal event burst vector contains preset dangerous semantic keywords or voiceprint fluctuation features, and the physical quantity mutation vector shows an exponential increase in concentration, then the causal confidence score is the first association threshold. If the sudden vector of the abnormal event is environmental steady-state noise and the sudden vector of the physical quantity has no abnormal fluctuation, then the causal confidence score is the second correlation threshold. S205. Based on the causal confidence score, the cross-modal feature correlation degree is quantified.

[0025] In this embodiment, as described in steps S201~S205 above, specifically: Step S201 is used to perform temporal segmentation preprocessing on the extracted structured perception features and audiovisual interaction features. The asymmetric time window slicing process means that instead of using the same synchronized time window for both types of features, the time window corresponding to the audiovisual interaction features is offset by a preset duration relative to the time window of the structured perception features (i.e., time delay offset). This time delay offset is a fixed value pre-calibrated based on the diffusion speed of the target toxic agent and the spatial scale of the work area. The core purpose is to match the temporal dimension of the two types of features with the objective physical time delay of the toxic agent's diffusion from the leak source to the point of being captured by the acquisition equipment. For example, in a chemical plant area with a span of 50 meters, the diffusion speed of a certain toxic agent is approximately 0.8... If it takes 62.5 seconds for the leakage to spread to the edge acquisition device, then the time delay offset can be preset to 60 seconds. That is, the start time of the time window of the audio-visual interaction feature is delayed by 60 seconds compared with the start time of the time window of the structured perception feature. Then, the two types of features are continuously sliced ​​into time segments of a uniform length of 15 seconds each to ensure that the extracted structured perception feature segments (such as data on sudden changes in the concentration of toxic agent in a specific area) and audio-visual interaction feature segments (such as leak noises captured synchronously in the area) accurately correspond to the real physical process of toxic agent diffusion in the time dimension.

[0026] Step S202 involves extracting core vectors representing risk associations from two types of features within the asymmetric time window determined in step S201. The abnormal event outbreak vector is a high-dimensional data vector representing information related to sudden on-site risks, selected from the audio-visual interaction features. Specifically, it includes key information such as voice features corresponding to dangerous semantic keywords, leakage-related soundprint fluctuation features, and image features corresponding to abnormal personnel movements. The physical quantity mutation vector is a numerical vector extracted from the structured perception features reflecting significant changes in on-site physical environment indicators, specifically covering abnormal data such as surges in toxic agent concentration, sudden changes in wind speed, and rapid movement of personnel. For example, continuing the example from S201, within the 15-second time slice of the structured perception features (time interval 10:00:00-10:00:15) and the time-delayed slice corresponding to the audio-visual interaction features (time interval 10:01:00-10:01:15), the abnormal event outbreak vector containing the semantic keyword "chlorine leak" and leakage abnormal soundprints is extracted from the audio-visual interaction features. The vector also extracts the chlorine concentration from 0.5... Upgraded to 1.0 The concentration abrupt change data, wind speed from 2.0 Rising to 3.5 Meteorological abrupt change data are combined to form a physical quantity abrupt change vector.

[0027] Step S203 is used to construct a mapping rule carrier for the associated abnormal event outbreak vector and physical quantity mutation vector. The chemical defense risk physical causal matrix is ​​a two-dimensional matrix constructed based on common chemical defense operation scenarios (such as risk events caused by different toxic agent leaks and equipment failures), the physical laws of toxic agent diffusion, and historical disposal data. The row dimension of the matrix corresponds to the core feature dimension of the abnormal event outbreak vector (such as the type of danger semantics, the intensity of sound wave fluctuations, and the category of abnormal actions), and the column dimension corresponds to the key indicator dimension of the physical quantity mutation vector (such as the concentration change amplitude, the range of meteorological parameter fluctuations, and the personnel displacement distance). The elements in the matrix predefine the physical correlation relationship corresponding to different feature combinations of the two types of vectors (such as "leakage semantics + exponential increase in concentration" corresponds to strong correlation, and "environmental noise + no concentration fluctuation" corresponds to no correlation). For example, for preset operation scenarios such as chlorine leak and ammonia leak, the matrix has pre-labeled the mapping relationship of "semantic keywords of toxic agent leak + concentration increase within 5 minutes ≥ 200%" as high-risk strong correlation, and the mapping relationship of "equipment operating noise + concentration fluctuation ≤ 5%" as no correlation. Subsequently, the physical correlation attributes of the two types of vectors can be quickly matched directly based on this matrix.

[0028] Step S204 is used to quantify the correlation between the extracted abnormal event outbreak vector and the physical quantity mutation vector based on the chemical defense risk physical causal matrix constructed in step S203, and obtain the corresponding causal confidence score. The causal confidence weighted calculation refers to a calculation method that combines the predefined physical mapping relationship in the matrix to score the matching degree of the core features of the two types of vectors. The first correlation threshold is a preset high-confidence quantification score (usually ranging from 0.8 to 1.0), corresponding to risk scenarios where the two types of vectors are strongly correlated. The second correlation threshold is a preset low-confidence quantification score (usually ranging from 0.1 to 0.3), corresponding to steady-state scenarios where the two types of vectors are not correlated. Specifically, the core features of the abnormal event outbreak vector (such as semantic keywords and voiceprint type) and the key indicators of the physical quantity mutation vector (such as concentration change) are first compared. Substituting the trend and fluctuation range into the chemical defense risk physical causal matrix, the corresponding correlation type is matched. If the matching result is "the abnormal event sudden vector contains preset dangerous semantic keywords such as toxic agent leakage and evacuation, or contains preset leakage-type acoustic wave characteristics such as pipeline rupture and liquid splash, and the physical quantity mutation vector shows that the toxic agent concentration increases exponentially within a unit time (e.g., concentration increase ≥300% within 10 minutes)", then the first correlation threshold (e.g., 0.92) is directly assigned as the causal confidence score. If the matching result is "the abnormal event sudden vector only contains environmental steady-state noise such as normal equipment operation noise and environmental background noise, and the physical quantity mutation vector shows that indicators such as toxic agent concentration, wind speed, and personnel location do not have obvious abnormal fluctuations (e.g., concentration fluctuation range ≤5%)", then the second correlation threshold (e.g., 0.15) is assigned as the causal confidence score.

[0029] Step S205 is used to directly convert the causal confidence score obtained in step S204 into the final quantitative result representing the degree of correlation between structured perception features and audiovisual interaction features, namely, cross-modal feature correlation. Cross-modal feature correlation is the core quantitative indicator of the strength of correlation between the two types of features in the risk representation dimension, with a value range of 0-1. The closer the value is to 1, the stronger the correlation between the two types of features; the closer it is to 0, the weaker the correlation. In specific quantification, no additional complex calculations are required; the causal confidence score is directly used as the final value of the cross-modal feature correlation. If the causal confidence score obtained in S204 is the first correlation threshold (…), then… For example, if the causal confidence score is 0.92, then the corresponding cross-modal feature correlation degree is 0.92. If the obtained causal confidence score is the second correlation threshold (e.g., 0.15), then the corresponding cross-modal feature correlation degree is 0.15. For example, continuing with the aforementioned chlorine leak scenario, the abnormal event sudden vector in S204 contains the dangerous semantic keyword "chlorine leak", and the physical quantity mutation vector shows that the concentration increases by 300% (exponential increase) within 5 minutes, resulting in a causal confidence score of 0.92 (first correlation threshold). In this step, this score is directly quantified as a cross-modal feature correlation degree of 0.92, which intuitively reflects the strong correlation between the two types of features.

[0030] In one embodiment, step S30, which involves extracting cross-modal temporal correlation features between structured perception features and audiovisual interaction features using the cross-modal feature correlation degree, specifically includes: S301. Based on the cross-modal feature correlation degree, the structured perception feature and the audio-visual interaction feature are temporally segmented and aligned, and continuous time segments with correlation degree higher than a preset threshold are extracted as risk evolution intervals. S302. Within the risk evolution range, extract the physical quantity temporal evolution vector of the structured perception feature and the abnormal behavior temporal evolution vector of the audio-visual interaction feature, respectively. S303. Construct a cross-modal time delay coupling matrix, wherein the cross-modal time delay coupling matrix defines the dynamic coupling relationship between the physical quantity time-series evolution vector and the abnormal behavior time-series evolution vector under different preset time delay offset levels; S304. Based on the cross-modal time delay coupling matrix, calculate the dynamic cross-correlation peak value between the physical quantity time-series evolution vector and the abnormal behavior time-series evolution vector, and determine the dynamic cross-correlation peak value as the cross-modal time-series correlation feature.

[0031] In this embodiment, as described in steps S301~S304 above, specifically: Step S301 uses the cross-modal feature correlation obtained in step S205 as a basis to perform time-dimension segment matching on structured perception features and audio-visual interaction features, and filters out core time segments that can centrally reflect risk evolution. Temporal segment alignment refers to dividing the two types of features into several time segments with the same time granularity (e.g., 10 seconds / segment, 30 seconds / segment), and then aligning the segments with matching correlation in the two types of features according to the cross-modal feature correlation of each segment in the corresponding time dimension. The preset threshold is a pre-set correlation judgment standard (usually ranging from 0.6 to 0.8) used to distinguish between effective risk-related segments and unrelated segments. The risk evolution interval refers to the continuous time segments in the two types of features where the correlation is consistently higher than the preset threshold, which is the core data range for subsequent extraction of dynamic evolution vectors. In specific implementation, a unified time segment granularity (e.g., 30 seconds) is first set. The structured perception features and audiovisual interaction features are divided into several continuous time segments. The cross-modal feature correlation degree of each corresponding time segment is calculated. Then, the correlation degree of each group is compared with a preset threshold (e.g., 0.7). Time segments with a correlation degree higher than 0.7 are selected. If there are three or more consecutive time segments with a correlation degree higher than the preset threshold, the complete time range corresponding to the consecutive combination is determined as the risk evolution interval. For example, continuing with the chlorine leak scenario mentioned above, the cross-modal feature correlation degrees are 0.85 (0-30 seconds), 0.92 (30 seconds-1 minute), 0.88 (1 minute-1 minute 30 seconds), and 0.65 (1 minute 30 seconds-2 minutes). The preset threshold is 0.7. The correlation degree of the three consecutive time segments corresponding to 0-1 minute 30 seconds is higher than the threshold. Therefore, the continuous time range of 0-1 minute 30 seconds is determined as the risk evolution interval.

[0032] Step S302 is used to extract core vectors reflecting the dynamic changes in the time dimension from two types of features within the risk evolution interval determined in step S301. The physical quantity time-series evolution vector is a set of dynamic changes in structured perception features over time within the risk evolution interval. Specifically, it includes time-series data of physical indicators such as toxic agent concentration, wind speed and direction, and personnel location (e.g., concentration records every 5 seconds, coordinate updates every 10 seconds), which can intuitively reflect the evolution trend of the physical environment. The abnormal behavior time-series evolution vector is a set of dynamic changes in audio-visual interaction features within the risk evolution interval. Specifically, it covers the timing of the appearance of dangerous semantic keywords, the time change of voiceprint fluctuation intensity, and the time-series frame data of abnormal personnel actions, which can clearly reflect the development process of abnormal events on site. For example, continuing the 0-1 minute 30 second risk evolution interval of the aforementioned chlorine leak scenario, chlorine concentration data (0 seconds 0.3 seconds 0.5 ... 5 seconds 0.45 ... 90 seconds 1.2 The wind speed and direction data, collected every 10 seconds, are combined to form a physical quantity time-series evolution vector. The time point of occurrence of the semantic keyword "chlorine leak" (25 seconds), the temporal change of the intensity of the abnormal leakage sound (the sound intensity increases from 0.3 to 0.8 from 20 to 40 seconds), and the video frame time-series data of the person calling for help are extracted from the audio-visual interaction features and combined to form an abnormal behavior time-series evolution vector.

[0033] Step S303 is used to construct a carrier representing the dynamic correlation rules between the physical quantity temporal evolution vector and the abnormal behavior temporal evolution vector under different time delay scenarios. The cross-modal time delay coupling matrix is ​​a two-dimensional matrix constructed based on the temporal law of chemical agent diffusion in chemical defense scenarios and the feature dimensions of the two types of evolution vectors. The row dimension of the matrix corresponds to the time step dimension of the physical quantity temporal evolution vector (e.g., a continuous sequence with one time step every 5 seconds), and the column dimension corresponds to the time step dimension of the abnormal behavior temporal evolution vector. The preset time delay offset level is a pre-divided time delay interval (e.g., 0-0.5 seconds, 0.5-1 seconds, 1-1.5 seconds, etc.) used to quantify the degree of time offset between the two types of vectors. The dynamic coupling relationship refers to the degree of correlation between the physical quantity evolution trend and the abnormal behavior development process under different time delay offset levels (e.g., strong coupling, medium coupling, weak coupling). Specifically, during construction, the risk evolution interval is first determined... The time length and sampling frequency of the two types of vectors are divided into preset time delay offset levels (e.g., 0-0.5 seconds, 0.5-1 seconds, 1-1.5 seconds, and 1.5-2 seconds). Then, each time step of the physical quantity time evolution vector is paired with each time step of the abnormal behavior time evolution vector according to different time delay offset levels. The coupling strength is predefined in the corresponding cell of the matrix. For example, continuing the chlorine leak scenario mentioned above, the cross-modal time delay coupling matrix has predefined "the time step of concentration surge in the physical quantity time evolution vector (30 seconds) and the time step of acoustic peak in the abnormal behavior time evolution vector (30.6 seconds), corresponding to the 0.5-1 second time delay offset level, the coupling relationship is strong coupling" and "the time step of concentration stabilization (10 seconds) and the time step corresponding to environmental noise (10.2 seconds), corresponding to the 0-0.5 second time delay offset level, the coupling relationship is weak coupling".

[0034] Step S304 is used to quantify the peak correlation between the physical quantity time-series evolution vector and the anomalous behavior time-series evolution vector during dynamic changes, based on the cross-modal time-delay coupling matrix constructed in step S303, thereby determining the cross-modal time-series correlation features. The dynamic cross-correlation peak value refers to the maximum correlation value (range 0-1) obtained after cross-correlation operations on the two types of vectors at different preset time-delay offset levels. This value directly reflects the strongest correlation between the two types of vectors in the time dimension. The cross-modal time-series correlation feature is the quantization result corresponding to this peak value, which is the core dynamic feature for subsequently constructing a multimodal fusion feature set. Specifically, the time-step data of the physical quantity time-series evolution vector and the anomalous behavior time-series evolution vector are paired one by one according to the predefined time-delay offset levels of the cross-modal time-delay coupling matrix. For each pair of paired data, cross-correlation is performed (calculating the similarity of the changing trends of the two types of vectors) to obtain the cross-correlation number under the corresponding time delay level. Then, the coefficient with the largest value is selected from the cross-correlation numbers corresponding to all time delay levels as the dynamic cross-correlation peak value. Finally, this peak value is directly determined as the cross-modal time series association feature. For example, continuing the risk evolution interval of 0-1 minute 30 seconds in the aforementioned chlorine leak scenario, the physical quantity time series evolution vector (concentration, wind speed time series data) and the abnormal behavior time series evolution vector (soundprint intensity, semantic occurrence time series data) are paired and operated according to time delay offset levels such as 0-0.5 seconds and 0.5-1 seconds. The cross-correlation numbers of each level are 0.65, 0.88, 0.72, and 0.58, respectively. The largest coefficient of 0.88 is the dynamic cross-correlation peak value, which is determined as the cross-modal time series association feature.

[0035] In one embodiment, step S40, the construction and training of the time-series risk evolution model, specifically includes: S401. Obtain multiple sets of historical multimodal sequence data from historical chemical defense operation data. Each set of historical multimodal sequence data corresponds to a historical leakage event or simulated event under a preset operation scenario, and is marked with the actual diffusion value and risk label of the corresponding event. S402. Perform data preprocessing and feature fusion processing on each set of historical multimodal sequence data to obtain several sets of historical multimodal fusion feature set samples; S403. Construct a multi-branch model, the model including: The temporal coding module is used to receive historical multimodal fusion feature set samples for feature extraction and temporal evolution, in order to calculate the dynamic temporal correlation of risk and the nonlinear coupling relationship of multiple factors, and output the encoded feature vector; The first branch is used to receive encoded feature vectors, perform regional risk trend aggregation processing, and output the regional risk level. The second branch is used to receive encoded feature vectors to perform individual risk assessment and output the individual risk level. The third branch is used to receive encoded feature vectors, perform spatiotemporal diffusion trend deduction processing, and output the toxic agent diffusion trend within a preset time period. S404. Define a training objective function based on the actual diffusion value and risk label, and regularize the training objective function with diffusion dynamics constraints to construct an optimization objective with physical constraints. S405. Using historical multimodal fusion feature set samples as input, train the constructed multi-branch model along the optimization objective with physical constraints, so that the regional risk level, individual risk level and toxic agent diffusion trend output by the model meet the preset error threshold and satisfy the physical boundary conditions under the preset operation scenario, and obtain the trained time-series risk evolution model.

[0036] In this embodiment, as described in steps S401-S405 above, specifically: Step S401 is used to obtain basic data samples for training the time-series risk evolution model, and to construct a training dataset containing real and simulated scenarios. Historical chemical defense operation data comes from past chemical defense disposal records in chemical industrial parks, archived emergency drill data, and laboratory toxic agent diffusion simulation data. Historical multimodal sequence data refers to each set of data containing structured historical sensor data (such as historical toxic agent concentration, meteorological parameters, and personnel location time-series data) and unstructured historical audio and video data (such as historical on-site monitoring footage, personnel voice recordings, and leak noise recordings) corresponding to the corresponding event, and all data are arranged in chronological order to form a continuous sequence. The preset operation scenarios cover common scenarios such as different types of toxic agents (such as chlorine, ammonia, hydrogen sulfide, etc.), different leak scales (small point leaks, medium-sized pipeline leaks, large equipment rupture leaks), and different operating environments (indoor workshops, outdoor equipment areas, confined spaces). Historical leak events refer to past... Chemical risk events that actually occur and have complete data records are considered real-world events. Simulated events refer to virtual risk events generated through simulation platforms and laboratories that conform to real physical laws. Actual diffusion values ​​are the actual diffusion parameters of each data set (such as actual diffusion range, peak concentration at different time points, and diffusion duration). Risk labels are the actual risk levels marked based on the actual consequences of the event or professional assessments (such as regional risk levels and individual personnel risk levels). For example, the historical multimodal sequence data corresponding to a medium-sized ammonia leak event recorded in a chemical industrial park in 20XX includes ammonia concentration time-series data every 5 seconds within 2 hours after the event, wind speed and direction data every 15 seconds, location trajectory data of 3 on-site workers, and synchronously recorded on-site monitoring footage (including the diffusion process of the leaking white mist) and emergency communication voice messages from personnel. It also marks the actual diffusion values ​​(maximum diffusion range of 80 meters within 2 hours, peak concentration of 2.0 in 1 hour). The data set, consisting of risk labels (regional risk level 1, and 3 individuals with high risk level), serves as a valid sample for model training.

[0037] Step S402 is used to standardize and integrate multi-dimensional features of each set of historical multimodal sequence data obtained in step S401, transforming it into sample data that the model can directly train. The data preprocessing logic is consistent with step S10, namely, performing missing value completion (filling in blank data from sensor disconnections in historical records), abnormal error removal (removing outrageous data caused by human error or equipment failure) and unifying sampling frequency (adjusting structured data from different sources to the same time granularity, such as unifying it to 10 seconds / sample) on historical unstructured audio and video data (such as historical surveillance footage and voice recordings), performing audio noise reduction (removing environmental noise), invalid segment deletion (deleting blank segments without abnormal events), and time alignment of video frames and time-series data (ensuring accurate matching of audio and video segments with structured data at corresponding time points). The feature fusion processing logic is consistent with steps S20 and S30, namely, first extracting historical structured perception features (such as historical concentration gradients and meteorological change features) from the preprocessed historical structured data, and then extracting features from historical audio and video data. Historical audio-visual interaction features (such as historical hazard semantic keywords and leak voiceprint features) are extracted. Based on preset risk calculation rules, the correlation degree of historical cross-modal features is calculated. Then, this correlation degree is used to extract historical cross-modal temporal correlation features. Finally, historical structured perception features, historical audio-visual interaction features, historical cross-modal feature correlation degrees, and historical cross-modal temporal correlation features are integrated in a unified data format. For example, for the aforementioned historical multimodal sequence data of a medium-sized ammonia leak event in 20XX, missing values ​​in the ammonia concentration data within two hours are first filled in (filling in missing values ​​for a certain voiceprint at the 35th minute). The data was collected by sampling at a uniform frequency of 10 seconds per data point (sensor disconnection data every 10 seconds). The video was denoised, and segments without abnormalities in the first 10 minutes were removed and time-aligned. Then, features such as the concentration gradient change (increase of 250% within 1 hour) and the semantic keyword "ammonia leak" were extracted. The historical cross-modal feature correlation degree was calculated to be 0.89 and the historical cross-modal temporal correlation feature degree was 0.83. Finally, the four types of features were integrated to form a set of historical multimodal fusion feature set samples. After multiple sets of historical event data were processed in the same way, several sets of standardized model training samples could be obtained.

[0038] Step S403 is used to construct a model architecture that can simultaneously output prediction results for three types of risks. This multi-branch model uses a temporal encoding module to uniformly process features and performs specialized operations on three independent branches to achieve simultaneous multi-dimensional risk assessment. The temporal encoding module is the core feature processing unit of the model. It adopts temporal model structures such as temporal convolutional networks (TCN) or long short-term memory networks (LSTM). After receiving historical multimodal fusion feature set samples, it extracts high-dimensional key features from the samples through multi-layer neural networks, while simulating the evolution of features over time, thereby analyzing the dynamic temporal correlation of risks (such as the temporal relationships of different features). The response sequence and the nonlinear coupling relationship with multiple factors (such as the synergistic effect of concentration and wind speed) are considered. The final output is an encoded feature vector that integrates all risk information (with a uniform dimension of 512 or 1024 for easy subsequent branch processing). The first branch is the regional risk assessment branch, which adopts a fully connected network structure. After receiving the encoded feature vector, it performs trend aggregation processing on the features related to regional risk (such as the global concentration distribution and the temporal characteristics of the diffusion range) (integrating regional risk characterization information from different time steps) and matches the preset regional risk classification standard (Level 1: Extremely High Risk, Level 2: High Risk, Level 3: High Risk, Level 4: High Risk, Level 5: High Risk, Level 6: High Risk, Level 7: High Risk, Level 8: High Risk, Level 9: High Risk, Level 1: High Risk, Level 1: High Risk, Level 1: High Risk, Level 1: High Risk, Level 1: High Risk, Level 2 ... The first branch is for regional risk assessment, with three levels: Level 3 (medium risk) and Level 4 (low risk). The second branch is for individual risk assessment, also based on a fully connected network. It focuses on extracting features related to individuals from the encoded feature vector (such as temporal changes in the distance between individuals and the leak source, temporal data of concentration at the individual's location, and characteristics of the individual's movement trajectory). This is combined with a toxic agent damage threshold model to assess the risk and output the individual risk level (high risk, medium risk, low risk). The third branch is for spatiotemporal diffusion trend prediction, using a spatiotemporal prediction network (such as ST-Former or ConvLSTM). After receiving the encoded feature vector, it mines... The spatiotemporal evolution patterns (such as the boosting effect of wind speed and direction on diffusion, and the hindering effect of terrain on diffusion) are used by time-series prediction algorithms to predict the diffusion range, concentration distribution, and diffusion speed of toxic agents within a preset time (such as 10 minutes, 30 minutes, or 1 hour), and output specific toxic agent diffusion trends. For example, continuing the sample processing of the aforementioned medium-sized ammonia leak event, after receiving the historical multimodal fusion feature set sample of the event, the time-series coding module extracts key features such as concentration increase, wind speed fluctuation, and the semantics of "ammonia leak," and calculates the dynamic time correlation relationship of "concentration rise precedes people's calls for help by 15 seconds" and "wind speed ≥ 2 The nonlinear coupling relationship of "concentration diffusion rate increases by 50%" outputs a 512-dimensional encoded feature vector; the first branch aggregates the global concentration distribution features in this vector, matches it to the first-level extremely high risk standard, and outputs it; the second branch extracts the temporal features of the distance between three people and the leak source (all ≤50 meters), and outputs the high-risk level; the third branch combines the wind speed 2.3 Based on characteristics such as northwest winds, it is estimated that ammonia gas will spread to a range of 80 meters within 1 hour, and the corresponding toxic agent diffusion trend is output.

[0039] Step S404 defines the optimization direction of model training. By combining data label error and physical constraints, it ensures that the model training closely matches real data without violating objective physical logic. The training objective function is a mathematical expression that measures the error between the model's predictions and the actual situation (actual diffusion values, risk labels), and its core consists of regression loss terms and classification loss terms. The diffusion dynamics constraint term is a regularization term designed based on the physical laws of toxic agent diffusion (such as Fick's law of diffusion, the law of conservation of mass, and fluid dynamics equations). It is used to limit the model's predictions from exceeding objective physical boundaries (such as diffusion speed not exceeding the maximum propagation speed of the toxic agent in the corresponding medium, and concentration changes conforming to the law of conservation of mass diffusion). Regularization refers to incorporating the diffusion dynamics constraint term into the training objective function with preset weights, so that the model minimizes prediction errors. While ensuring accuracy, the optimization objective is forced to meet physical laws, ultimately constructing a physically constrained optimization objective that balances data fitting accuracy and physical rationality. In practice, a regression loss term (using mean squared error, MSE) is first defined based on the actual diffusion value to calculate the error between the model's predicted diffusion trend (e.g., diffusion range, peak concentration) and the actual diffusion value. Then, a classification loss term (using cross-entropy loss) is defined based on the risk label to calculate the error between the model's predicted regional risk level, individual risk level, and the actual risk label. The two losses are then summed according to their weights (e.g., regression loss weight 0.6, classification loss weight 0.4) to obtain the basic training objective function. Subsequently, diffusion kinetics constraints are designed, such as constraining the model's predicted diffusion rate to not exceed the maximum diffusion rate of the toxic agent in the air (e.g., the maximum diffusion rate of chlorine is 3.2). The maximum diffusion velocity of ammonia is 3.5. The concentration change rate conforms to the mass conservation law (the increase in concentration in a certain area per unit time ≤ the diffusion input). This constraint term is multiplied by a regularization weight (usually 0.1-0.3) and added to the basic training objective function to form the final optimization objective with physical constraints. For example, continuing the training scenario of the aforementioned medium-sized ammonia leak event, the basic training objective function is "MSE (model-predicted 1-hour diffusion range - actual 80-meter diffusion range) × 0.6 + cross-entropy (model-predicted regional risk level - actual level 1 risk label) × 0.4)," and the diffusion kinetics constraint term is set as "if the model-predicted ammonia diffusion rate > 3.5..." "If the penalty loss is triggered, the constraint term will be incorporated into the basic objective function with a weight of 0.2. The constructed optimization objective will both promote the model to accurately predict the diffusion range and risk level, and limit the prediction results from exceeding the physical velocity boundary of ammonia diffusion."

[0040] Step S405 is used to drive model learning through historical sample data, completing the training iteration of the multi-branch model along a preset optimization objective with physical constraints, and finally obtaining a time-series risk evolution model with satisfactory performance. The preset error threshold is a pre-set standard for model prediction accuracy (the regression error threshold is usually set to a relative error between the predicted and true values ​​≤10%, and the classification error threshold is usually set to an accuracy ≥90%). The physical boundary conditions are the objective physical limitations of agent diffusion under the preset operational scenario (such as the maximum diffusion rate of different agents, the upper limit of concentration saturation in the environmental medium, and the diffusion range boundary caused by terrain). In specific implementation, the historical multimodal fusion feature set samples obtained in step S402 are first divided into training and validation sets in a 7:3 ratio. The training set samples are fed into the multi-branch model constructed in step S403, and the model parameters (convolutional kernel weights of the time-series coding module, parameters of fully connected layers in each branch, etc.) are iteratively updated along the optimization objective with physical constraints using a gradient descent algorithm (such as the Adam algorithm). In one iteration (traversing the training set once), the model output is validated using validation set samples. The regression error between the toxic agent diffusion trend and the actual diffusion value, and the classification accuracy between the regional / personnel risk level and the true label are calculated. Simultaneously, the model's predictions are checked to ensure they meet physical boundary conditions (e.g., whether the predicted diffusion rate is ≤ the corresponding maximum diffusion rate of the toxic agent, and whether the concentration is ≤ the saturation limit). If the model's prediction error exceeds a preset threshold or does not meet the physical boundary conditions, training continues iteratively with a dynamically adjusted learning rate (e.g., the learning rate decreases to 0.9 times its original value every 10 iterations) until the model's regression error on the validation set is ≤10%, the classification accuracy is ≥90%, and all predictions meet the physical boundary conditions. Training then stops, and the current model parameters are saved. For example, continuing the training scenario of the aforementioned medium-sized ammonia leak event, 1000 sets of historical multimodal fusion feature set samples, including those from this event, are used for training. The preset regression error threshold is 8%, the classification accuracy threshold is 92%, and the physical boundary condition is set as "ammonia diffusion rate ≤ 3.5". Concentration ≤ 3.0 After 200 rounds of iterative training, the model predicted that the ammonia leak would spread to a distance of 78 meters within one hour (a relative error of 2.5% compared to the actual 80 meters), with a regional risk level of Level 1 (accurate classification) and an individual risk level of high risk (accurate classification). The predicted spread rate was 2.8... Peak concentration 1.9 All physical boundary conditions are met, at which point the model reaches the preset training standard, and the trained temporal risk evolution model is obtained.

[0041] In one embodiment, step S40, which involves inputting the multimodal fusion feature set into a pre-trained temporal risk evolution model to calculate the dynamic temporal correlation of risk and the nonlinear coupling relationship of multiple factors, specifically includes: S411. Input the multimodal fusion feature set into the trained temporal risk evolution model; S412. Based on the temporal coding module in the temporal risk evolution model, feature extraction and temporal evolution are performed on the multimodal fusion feature set to generate a coded feature vector; S413. The temporal response sequence and evolution rate difference of each temporal component in the multimodal fusion feature set are captured along the time axis by the temporal coding module. The temporal response sequence reflects the response time offset relationship between the structured perception feature sequence and the audio-visual interaction feature sequence and the evolution trend of cross-modal temporal correlation features. The evolution rate difference reflects the relative deviation between the concentration change rate of the structured perception feature sequence and the abnormal activity change rate of the audio-visual interaction feature sequence. The cross-modal feature correlation degree is used as the confidence gate of the current time window and is used to weight and adjust the effective weight of the temporal response sequence to quantify the risk dynamic temporal correlation relationship. S414. Analyze the synergistic relationship between structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features, construct feature coupling weights, and map the temporal response sequence relationship and evolution rate difference into multi-source feature contributions according to the feature coupling weights. Through the gating mechanism of the temporal coding module, fuse the multi-source feature contributions to quantify the multi-factor nonlinear coupling relationship.

[0042] In this embodiment, as described in steps S411~S414 above, specifically: Step S411 is used to input the multimodal fusion feature set constructed in step S30 into the trained model as input data, providing a data foundation for subsequent feature encoding and risk correlation calculation. The trained temporal risk evolution model is a mature model with fixed parameters and meeting preset error thresholds and physical boundary conditions after training iterations in steps S401-S405. It has the ability to stably analyze multimodal features and output accurate risk prediction results. The input multimodal fusion feature set is a standardized data set integrating structured perception features, audio-visual interaction features, cross-modal feature correlation, and cross-modal temporal correlation features. The data format is... The samples are completely consistent with the historical multimodal fusion feature set used during model training (e.g., uniformly in 1024-dimensional vector format) to ensure that the model can directly identify and process them. For example, continuing with the aforementioned chlorine leak scenario, the multimodal fusion feature set finally constructed in step S30 includes structured perception features such as "chlorine concentration gradient, wind speed and direction", audio-visual interaction features such as "chlorine leak semantic keywords, leak abnormal soundprints", a cross-modal feature correlation degree of 0.92 and a cross-modal temporal correlation feature of 0.88. In this step, the standardized fusion feature set is directly input into the trained temporal risk evolution model to trigger feature extraction and temporal evolution operations within the model.

[0043] Step S412 is used by the core processing unit of the time-series risk evolution model to perform deep feature mining and time-dimensional evolution simulation on the input multimodal fusion feature set, generating an encoded feature vector that integrates all risk information. The time-series encoding module uses the structure from model training (such as a temporal convolutional network TCN or a long short-term memory network LSTM), possessing the dual ability to extract high-dimensional abstract features and capture temporal change patterns. Feature extraction refers to the time-series encoding module using convolution and activation operations of multi-layer neural networks to filter out key information crucial for risk assessment (such as concentration mutation magnitude, hazard semantic confidence, cross-modal correlation strength, etc.) from the multimodal fusion feature set, discarding redundant and noisy features. Time-series evolution refers to the time-series encoding module simulating the changes in various features over time based on the time-series attributes of the features. The process of change (such as the evolution trajectory of concentration from stable to surging, and voiceprint intensity from weak to strong) integrates static features and dynamic change patterns into a unified representation. The encoded feature vector is a standardized high-dimensional vector (usually 512-dimensional or 1024-dimensional) output after the above processing. It integrates the core information and time-series evolution patterns of structured perception features, audio-visual interaction features, and cross-modal correlation features, providing a foundation for subsequent calculation of risk correlation. For example, continuing with the aforementioned chlorine leak scenario, after receiving a multimodal fusion feature set containing information such as concentration gradient, leak semantics, and cross-modal correlation, the time-series coding module first extracts key features such as "concentration increases by 300% in 5 minutes," "chlorine leak semantic confidence level 0.95," and "cross-modal correlation level 0.92," and then simulates the time-series changes of these features (such as concentration from 0.3...). Rising to 1.2 The process (including the peak change of voiceprint intensity in 20-40 seconds) is ultimately integrated to generate a 1024-dimensional encoded feature vector, which fully preserves the core information and temporal evolution of various features.

[0044] Step S413 is used to deeply mine the temporal dimension patterns of the multimodal fusion feature set through the temporal coding module, capture two types of core temporal relationships, and combine them with confidence gating to weight them, ultimately quantifying the dynamic temporal correlation of risks. The temporal response sequence refers to the triggering order of the structured perception feature sequence and the audiovisual interaction feature sequence on the time axis, and the evolution trend of the associated features. Specifically, it is reflected in the response time offset of the two types of feature sequences (e.g., the time difference between the capture of audiovisual interaction features before or after the capture of structured perception features), and the trend of the strength of cross-modal temporal correlation features over time (e.g., gradually increasing from weak correlation to strong correlation). The evolution rate difference refers to the comparison of the dynamic change speed of the two types of feature sequences, specifically reflected in the relative difference between the change rate of the core indicator (toxin concentration) in the structured perception features (concentration increase or decrease per unit time) and the change rate of abnormal activity in the audiovisual interaction features (frequency of dangerous semantics, increase in voiceprint intensity, etc. per unit time). The confidence gating is the value obtained in step S205. The cross-modal feature correlation degree is used to adjust the weight of the temporal response sequence based on the strength of feature correlation (the higher the correlation degree, the greater the effective weight of the corresponding temporal relationship, and the more significant the impact on the final result). The risk dynamic temporal correlation relationship is a quantitative indicator (range 0-1) characterizing the correlation strength of various features in the time dimension; the higher the value, the tighter the temporal correlation. In specific implementation, the temporal coding module traverses the multimodal fusion feature set frame by frame along the time axis. It first records the trigger time points of the structured perception feature sequence (such as concentration temporal sequence) and the audio-visual interaction feature sequence (such as voiceprint and semantic temporal sequence), and calculates the response time offset between the two (e.g., the leakage abnormal sound is captured at 20 seconds, and the concentration begins to surge at 20.8 seconds, with a time offset of 0.8 seconds). At the same time, it tracks the evolution trend of cross-modal temporal correlation features (e.g., gradually increasing from 0.3 at 20 seconds to 0.88 at 40 seconds), forming the temporal response sequence. Then, it calculates the concentration change rate of the structured perception features (e.g., the concentration increases from 0.4 at 20-30 seconds). Rise to 0.7 The rate of change was 0.03. The abnormal activity change rate of the audio-visual interaction features (e.g., the voiceprint intensity increases from 0.3 to 0.7 in the 20th-30th second, with a change rate of 0.04 / second) has a relative deviation of 0.01, forming a difference in evolution rate. Subsequently, the cross-modal feature correlation degree (e.g., 0.92) is used as a confidence gate to adjust the temporal response sequence (correlation degree 0.92 corresponds to a weight coefficient of 0.92). Finally, the risk dynamic temporal correlation relationship is quantified through normalization operation (e.g., 0.89). For example, continuing the aforementioned chlorine leak scenario, the temporal coding module captures the temporal response sequence of "leakage abnormal sound (20 seconds) → chlorine leak semantics (25 seconds) → concentration surge (20.8 seconds)" (voiceprint precedes concentration change by 0.8 seconds, semantics lags concentration change by 4.2 seconds). The cross-modal temporal correlation feature increases from 0.3 at 20 seconds to 0.88 at 40 seconds; the concentration change rate is 0.02. The abnormal activity change rate is 0.03 / second, with a relative deviation of 0.01. After weighting with a cross-modal feature correlation degree of 0.92 as the confidence gate, the risk dynamic time correlation relationship of 0.89 is obtained, which intuitively reflects the strong correlation attributes of various features in the time dimension. It should be noted that the risk dynamic time correlation relationship is an intermediate representation quantity inside the time-series coding module and is not directly output as a business result. On the one hand, it is used to update the hidden state inside the time-series coding module and enrich the time-series representation information carried by the encoded feature vector. On the other hand, the quantization result will be used as an input parameter in the subsequent step S414 to construct the feature coupling weight.

[0045] Step S414 is used to analyze the interaction patterns of the four core features. Through weight allocation, contribution mapping, and fusion operations, the multi-factor nonlinear coupling relationship representing the risk of multi-factor synergistic influence is finally quantified. Among them, the synergistic association relationship refers to the interaction pattern between structured perception features, audio-visual interaction features, cross-modal feature association degree, and cross-modal temporal association features (such as the concentration surge verifying the authenticity of semantic keywords, cross-modal temporal association features strengthening the association strength of two basic features, and the evolutionary trend of wind speed change affecting the concentration diffusion rate and thus associating abnormal behavior). The feature coupling weight is a weight coefficient (with a value range of 0-1, and the total weight is 1) assigned based on the synergistic association strength of various features. The closer the synergistic association, the higher the weight. The larger the weight, the stronger the synergy between "concentration change features and cross-modal feature correlation," with a weight of 0.4; "audiovisual interaction features and cross-modal temporal correlation features" with a weight of 0.3; and the remaining weights for other feature combinations. Multi-source feature contributions are quantified by weighting the temporal response sequence and evolution rate differences obtained in step S413 according to feature coupling weights, reflecting the specific role of the two types of temporal relationships in multi-factor synergy. The gating mechanism is the unit in the temporal coding module that controls the feature contribution fusion ratio. It dynamically adjusts the fusion weights of each multi-source feature contribution through the sigmoid activation function to ensure a higher proportion of core contributions. The multi-factor nonlinear coupling relationship is the final quantified index obtained after fusing all multi-source feature contributions. (Value range 0-1), the higher the value, the stronger the degree of risk synergistic influence of multiple factors; in specific implementation, firstly, the synergistic relationship of the four types of features is analyzed through correlation analysis (e.g., the correlation coefficient between "chlorine concentration surge" and "chlorine leakage semantics" is 0.93, and the correlation coefficient between "cross-modal time-series correlation features" and "concentration change rate" is 0.87), and feature coupling weights are assigned based on this correlation (e.g., concentration feature 0.25, audio-visual interaction feature 0.2, cross-modal feature correlation degree 0.3, cross-modal time-series correlation feature 0.25); then, the time-series response sequence (quantized value 0.89) and evolution rate difference (quantized value 0.75) are mapped to the multi-source feature contribution according to the above weights (e.g., the contribution corresponding to the time-series response sequence is 0.89 × 100%. 0.55 = 0.49, the contribution corresponding to the difference in evolution rate is 0.75 × 0.45 = 0.34, where 0.55 and 0.45 are the comprehensive coupling weights of the two types of temporal relationships; then, through the gating mechanism of the temporal coding module, the contributions of multi-source features are dynamically fused (e.g., the core contribution of 0.49 is assigned a gating weight of 0.6, and the secondary contribution of 0.34 is assigned a gating weight of 0.4), and after normalization, the nonlinear coupling relationship of multiple factors is quantified; for example, continuing the aforementioned chlorine leak scenario, the analysis shows that "concentration features and cross-modal correlation" and "audio-visual interaction features and cross-modal temporal correlation features" are strong synergistic relationships, and the coupling weights are assigned to 0.35 and 0.3 respectively, and the temporal response sequence relationship is 0.89 and the evolution rate difference is 0.The 75 mapping contributes 0.41 and 0.29 to the multi-source features. After fusion through a gating mechanism, the final quantification yields a multi-factor nonlinear coupling relationship of 0.86, clearly reflecting the strong coupling attribute of multiple factors such as concentration, semantics, and cross-modal correlation in synergistically driving risk escalation. It should be noted that this multi-factor nonlinear coupling relationship is also an internal intermediate representation of the time-series coding module and is not directly output externally. It is used to further correct the encoded feature vector and, in conjunction with the aforementioned risk dynamic temporal correlation, generate the final version of the feature coupling weight. The encoded feature vector, along with this feature coupling weight, is output to subsequent prediction branches, participating in the calculation of regional risk level, individual risk level, and toxic agent spread trend.

[0046] In one embodiment, step S40, which outputs a prediction result including the regional risk level, the individual risk level of personnel, and the predicted trend of toxic agent spread within a preset time period, specifically includes: S421. Based on the encoded feature vector and feature coupling weight, calculate the regional risk trend aggregation value through the first branch of the time-series risk evolution model, and output the regional risk level according to the preset regional risk classification mapping rule. S422. Based on the encoded feature vector and feature coupling weight, calculate the individual risk assessment value of the personnel through the second branch of the time-series risk evolution model, and output the individual risk level of the personnel according to the preset individual risk classification mapping rule. S423. Based on the encoded feature vector and feature coupling weight, the spatiotemporal diffusion trend is deduced through the third branch of the time-series risk evolution model, and the toxic agent diffusion trend within a preset time period is output.

[0047] In this embodiment, as described in steps S421-S423 above, specifically: Step S421 is used to perform aggregate calculation and level determination of regional risk based on the core features and feature coupling weights after model encoding, and output accurate regional risk assessment results through the first branch. Here, the encoded feature vector is a high-dimensional vector of condensed multimodal risk information generated in step S412, and the feature coupling weights are weight coefficients constructed in step S414 that reflect the strength of synergistic association of various features (e.g., concentration feature weight 0.25, cross-modal association weight 0.3, etc.). The first branch is a preset regional risk assessment branch, adopting a fully connected network structure, and has the ability to extract regional risk-related features and aggregate trends. The regional risk trend aggregation value is a quantified value (range 0-1) obtained by weighting and aggregating features related to regional risk in the encoded feature vector (e.g., global toxic agent concentration distribution, temporal changes in diffusion range, and trends in environmental parameter influence) with the feature coupling weights. A higher value indicates a higher overall regional risk. The preset regional risk classification mapping rule is a predefined standard for the correspondence between aggregated values ​​and risk levels (e.g., aggregated value ≥ 0.9 corresponds to Level 1 extremely high risk, 0.7-0.89 corresponds to Level 2 high risk, 0.5-0.69 corresponds to Level 3 medium risk, and < 0.5 corresponds to Level 4 low risk). In specific implementation, the first branch first extracts regional risk-related features such as the peak concentration of the entire region, the growth rate of the diffusion range, and the temporal trend of cross-modal correlation strength from the encoded feature vector. Then, it performs weighted aggregation calculation on the temporal evolution trend of these features according to the feature coupling weight (e.g., the peak concentration feature is weighted at 0.25, the growth rate of the diffusion range at 0.2, and the cross-modal correlation strength at 0.3) to obtain the regional risk trend aggregated value. Subsequently, the aggregated value is compared with the preset regional risk classification mapping rule to match the corresponding risk level and output it. For example, continuing with the aforementioned chlorine leak scenario, the encoded feature vector contains a peak chlorine concentration of 1.2 across the entire region. Features such as a 5-minute diffusion range of 30 meters and a time-series peak cross-modal correlation strength of 0.92 are used to calculate the regional risk trend aggregation value of 0.93 by combining feature coupling weights (concentration feature 0.25, cross-modal correlation degree 0.3, and other related features 0.45). According to the preset hierarchical mapping rules (aggregate value ≥ 0.9 is Level 1 extremely high risk), the final output regional risk level is Level 1 (extremely high risk).

[0048] Step S422 is used to complete the individual risk quantification assessment and level determination for each worker based on the encoded feature vector and feature coupling weights through the second branch, outputting accurate individual risk results. The encoded feature vector contains core risk information related to the worker (such as real-time location coordinates, time-series data of toxic agent concentration at the location, worker movement trajectory characteristics, and changes in distance from the leak source). Weights in the feature coupling weights that are strongly correlated with individual risk (such as worker location feature weights and individual area concentration feature weights) will play a key role in the assessment process. The second branch is a pre-defined individual risk assessment branch, constructed using a fully connected network combined with an individual risk assessment model, focusing on extracting features directly related to worker safety. The individual risk assessment value is a quantified value obtained by combining the worker-specific features in the encoded feature vector with the feature coupling weights through weighted summation and risk factor superposition. The value ranges from 0 to 1, with higher values ​​indicating greater individual risk to personnel. The pre-defined individual risk grading mapping rule is a predefined standard corresponding to assessment values ​​and risk levels (e.g., assessment value ≥ 0.85 corresponds to high risk, 0.6-0.84 corresponds to medium risk, and < 0.6 corresponds to low risk). This rule is formulated in conjunction with factors such as the threshold of harm from toxic agents to the human body and the protection level of personnel protective equipment. In specific implementation, the second branch first extracts the unique features of each worker from the coded feature vector (e.g., real-time distance from the leak source, peak concentration of the toxic agent at the location, angle between the direction of movement and the direction of diffusion, duration of stay, etc.). Then, based on feature coupling weights (e.g., personnel distance feature weight 0.3, location concentration feature weight 0.3, movement trajectory feature weight 0.2, other auxiliary feature weight 0.2), these unique features are weighted and calculated, and a toxic agent harm threshold correction factor (e.g., concentration exceeding 0.8) is superimposed. Each individual risk assessment value is obtained by adding a risk increment of 0.15 over time. This assessment value is then compared with a preset individual risk grading mapping rule to match the corresponding risk level and output it in association with the personnel's identity information. For example, continuing with the aforementioned chlorine leak scenario, the encoded feature vector includes the unique features of two workers: Personnel A is 8 meters away from the leak source, and the chlorine concentration at their location is 1.2. Personnel B is moving away from the leak source; the distance between personnel B and the leak source is 6 meters; the chlorine concentration at their location is 1.1. 1. Not moved yet; combined with feature coupling weights (distance 0.3, concentration 0.3, movement trajectory 0.2, others 0.2), the individual risk assessment value of person A is 0.88 and that of person B is 0.91; according to the preset classification mapping rules (≥0.85 is high risk), the final output is that the individual risk level of the two workers is high risk, and their identity information is associated to clarify the risk subject.

[0049] Step S423 is used to mine the spatiotemporal evolution law and complete the prediction of the toxic agent diffusion trend based on the encoded feature vector and feature coupling weights through the third branch, and output the specific diffusion results within a preset time. The encoded feature vector contains core spatiotemporal information related to toxic agent diffusion (such as the location of the leak source, initial leak amount, wind speed and direction time series data, current diffusion range, concentration distribution characteristics, etc.). The weights in the feature coupling weights that are strongly correlated with the diffusion trend (such as environmental parameter feature weights, concentration evolution feature weights, and cross-modal time series correlation feature weights) will dominate the inference process. The third branch is a preset spatiotemporal diffusion trend inference branch, using a spatiotemporal prediction network structure (such as ST-Former, ConvLSTM), which has the ability to capture spatiotemporal correlation laws and perform dynamic inference. Spatiotemporal diffusion trend inference refers to combining the spatiotemporal features and feature coupling weights in the encoded feature vector, and simulating the propagation process of the toxic agent in the spatial dimension and the evolution trend in the temporal dimension through a time series prediction algorithm. The preset time is a pre-set inference duration (such as 10 minutes, 30 minutes, 1 hour), which can be flexibly configured according to actual emergency needs. The potential is the specific result of the extrapolation output, including quantitative information such as the boundary of the toxic agent diffusion range within a preset time, the predicted concentration distribution in different regions, the temporal change of diffusion speed, and the trend of diffusion direction shift. In specific implementation, the third branch first extracts spatiotemporal related features such as diffusion source parameters (location, initial leakage intensity), environmental impact features (wind speed and direction time series, terrain constraint features), and current diffusion status (spread range, concentration gradient distribution) from the encoded feature vector. Then, it performs weighted fusion of these features according to feature coupling weights (e.g., environmental parameter weight 0.3, concentration evolution feature weight 0.25, cross-modal temporal correlation feature weight 0.2, diffusion source parameter weight 0.25) and inputs it into the spatiotemporal prediction network. The network learns the diffusion patterns in historical samples, simulates the propagation process of the toxic agent under the action of airflow, and extrapolates the diffusion changes within a preset time by combining physical constraints such as mass conservation and Fick's diffusion law. The final output is a toxic agent diffusion trend containing specific values. For example, continuing with the aforementioned chlorine leak scenario, the encoded feature vector includes the leak source being located in the center of the device area, the initial leakage intensity being 0.5 kg / s, and the wind speed being 2.3. Based on features such as northwest wind and a current 10-minute diffusion range of 30 meters, combined with feature coupling weights (environmental parameter 0.3, concentration evolution 0.25, cross-modal temporal correlation 0.2, diffusion source parameter 0.25), the diffusion trend within 10 minutes (preset time 1) and 30 minutes (preset time 2) is extrapolated through the third branch. The output result is: "The diffusion range expands to 50 meters within 10 minutes, and the peak concentration in the northwest region is 1.5." The diffusion rate is 2.1. Within 30 minutes, the diffusion range expanded to 90 meters. Due to topographical influence, diffusion in the southeast direction was hindered, and the peak concentration reached 1.8%. The diffusion rate gradually decreased to 1.7. ".

[0050] In one embodiment, step S50 specifically includes: S501. Based on the risk level of the area, the risk level of the individual personnel, and the trend of toxic agent spread, match the preset disposal plan library to generate command and decision instructions that include evacuation route planning, delineation of decontamination areas, and deployment of emergency personnel. S502. The prediction results, command and decision instructions, collection timestamps of multi-source chemical defense data, and identity information corresponding to the individual risk level of personnel are encapsulated into chemical defense operation data blocks, written into the chemical defense command alliance chain for distributed storage, and data access permissions are subject to hierarchical control. S503 uses a pre-defined cloud-edge collaborative architecture to send command and decision instructions to edge execution nodes. At the same time, the cloud dynamically corrects and reschedules the command and decision instructions based on real-time feedback of on-site status data.

[0051] In this embodiment, as described in steps S501~S503 above, specifically: Step S501 is used to match the core risk assessment results output by the model with standardized emergency response plans to generate practical comprehensive command and decision-making instructions. Specifically, the regional risk level (e.g., Level 1, extremely high risk) clarifies the overall emergency response level, the individual risk level (e.g., high risk) identifies key protected individuals, and the toxic agent diffusion trend (e.g., diffusion range, direction, and concentration distribution within a preset timeframe) provides a spatiotemporal basis for developing response measures. The preset response plan library is a standardized set of plans built based on the "Guidelines for Emergency Response to Hazardous Chemical Accidents," historical successful response cases, and optimal simulation results. It includes evacuation strategies and decontamination markings corresponding to different risk levels and diffusion scenarios. The system includes rules for evacuation and personnel deployment. The solution database is categorized and stored using a three-dimensional index: "Regional Risk Level - Personnel Risk Level - Spread Trend Type," facilitating rapid matching. Command and decision-making instructions are execution instructions that integrate three core response needs, specifically including evacuation route planning (defining personnel evacuation routes and refuge location), delineation of decontamination areas (determining the scope, method, and type of disinfectant), and emergency personnel deployment (defining the number, deployment location, and operation sequence of rescue, medical, and decontamination teams). In specific implementation, the regional risk level (Level 1), individual personnel risk level (high risk), and the spread trend of the toxic agent (spreading to 50 meters within 10 minutes, northwest wind direction, peak concentration 1.5) are first determined. Using [condition] as the search criteria, the system matches corresponding basic response plans from the response plan database. Then, it fine-tunes the basic plans based on real-time on-site data (such as the current location of personnel, surrounding refuge resources, and the real-time standby status of emergency teams), optimizing evacuation routes to avoid high-concentration diffusion areas, accurately delineating decontamination areas to cover the core diffusion area and potential pollution boundaries, and rationally allocating emergency personnel to meet the core needs of "prioritizing the rescue of high-risk personnel and rapidly controlling the spread." Finally, it generates structured command and decision-making instructions. For example, continuing with the aforementioned chlorine leak scenario, it matches the basic plan corresponding to "Level 1 regional risk + high-risk personnel risk + northwestward diffusion" in the plan database, considering that two high-risk personnel are located east of the leak source. Real-time information regarding two refuge points and three emergency response teams within 10 meters to the south and 300 meters around the perimeter generates the following instructions: "Evacuation route planning: Two high-risk personnel will evacuate in the southeast direction, proceeding to Refuge Point No. 1 (150 meters away) via the east side passage of the factory area, avoiding the northwest spread coverage area along the way; Decontamination area delineation: High-pressure spray decontamination will be used in the core area with a radius of 20 meters centered on the leak source, and atomized disinfection will be used at the edge of the spread path (within 50 meters); Emergency personnel deployment: One rescue team (3 people) will immediately go to the scene to guide personnel evacuation, one medical team (2 people) will stand by at Refuge Point No. 1, and one decontamination team (4 people) will rush to the core decontamination area with chlorine-specific decontamination agent."

[0052] Step S502 is used to securely encapsulate and distribute the core data of the entire chemical defense operation process, while ensuring data access security through access control, achieving data traceability, immutability, and privacy protection. The prediction results include the regional risk level (Level 1, extremely high risk), individual personnel risk level (2 high-risk individuals), and the toxic agent diffusion trend (diffusion to 50 meters within 10 minutes, etc.). Command and decision-making instructions are the evacuation, decontamination, and personnel deployment instructions generated in S501. The collection timestamps of the multi-source chemical defense data refer to the specific time (accurate to the second, e.g., 20XX-XX-XX) of the collection of structured sensing data (such as concentration and meteorological data) and unstructured audio and video data. 14:25:30), used to ensure data time-series traceability; personnel identity information is a unique identifier for operators corresponding to individual risk levels (such as employee ID, name, and position), used to accurately associate risk subjects; chemical defense operation data blocks are independent data units that encapsulate the above four types of information in a preset structured format (such as JSON format), including a data header (certificate holder, time), data body (various core information), and digital signature (to ensure data integrity); the chemical defense command alliance chain is a distributed ledger network jointly constructed by emergency management departments, chemical industrial park management, emergency rescue teams, regulatory agencies, and other relevant parties, with each participating party acting as an alliance Chain nodes jointly maintain data consistency; distributed notarization refers to synchronously storing encapsulated data blocks to all nodes of the consortium blockchain, ensuring data immutability and traceability through blockchain hash algorithms and consensus mechanisms (such as PBFT algorithm), avoiding data loss due to single points of failure; hierarchical access control is based on data access rules set by the roles of participating parties, which are solidified and executed through smart contracts to ensure that different roles can only access data within their authorized scope; in specific implementation, data blocks are first encapsulated according to a preset format, for example: {"Data Header": {"Notarization Subject": "XX Chemical Industrial Park Emergency Command Center", "Notarization Time": "20XX-XX-XX}} 14:30:00”},“Prediction Result”: {“Area Risk Level”: “Level 1”,“Personnel Risk Level”: [{“Employee ID”: “001”,“Name”: “Zhang XX”,“Level”: “High Risk”}]},“Command Decision Instruction”: “Evacuation Route Planning: XXX”,“Collection Timestamp”: {“Concentration Data Collection Time”: “20XX-XX-XX 14:25:30”,“Audio / Video Collection Time”: “20XX-XX-XX 14:25:20”}};The data block is then hashed using SHA-256 to generate a data digest, which is digitally signed by the command center and submitted to the chemical defense command consortium blockchain;After each node of the consortium blockchain verifies the digital signature and data integrity, the data block is synchronously stored through the consensus mechanism;Finally, access permissions are configured as follows: Command Center (highest privileges, able to view all data), Rescue Team (only able to view command and decision-making instructions and relevant personnel identification information), Regulatory Agency (able to view all stored data for review and auditing), and Ordinary Employees (no access permissions). Permission changes require confirmation from a majority of nodes in the consortium blockchain. For example, continuing with the aforementioned chlorine leak scenario, the encapsulated data block contains the IDs of two high-risk individuals (001, 002), various data collection timestamps, Level 1 risk level, evacuation and decontamination instructions, etc. After submission to the consortium blockchain, each node simultaneously stores the evidence. The rescue team can only view the evacuation route and corresponding personnel information, and cannot modify the data. The regulatory agency can subsequently review the rationality of the emergency response process through the stored data.

[0053] Step S503 is used to achieve efficient distribution and dynamic optimization of command and decision-making instructions through a cloud-edge collaborative architecture, ensuring that emergency response instructions are aligned with the real-time situation on-site, and improving response timeliness and accuracy. The preset cloud-edge collaborative architecture is a collaborative system composed of a cloud-based emergency command platform (responsible for overall decision-making and data analysis) and edge-side execution nodes (close to the site, low-latency response). The cloud focuses on overall coordination and dynamic adjustment, while the edge focuses on local execution and data collection. The two achieve real-time data interaction through high-speed communication links such as 5G and Industrial Internet of Things (IIoT), avoiding instruction delays caused by long-distance transmission from the cloud. Edge-side execution nodes are various terminal devices and control units deployed on-site, including handheld smart terminals for rescue personnel, decontamination equipment control nodes, evacuation guidance indicator lights, and on-site sensor gateways, possessing the ability to receive instructions, execute operations, and collect real-time data. The real-time feedback on the on-site status data is the latest scene information dynamically collected by the edge-side execution nodes, such as personnel evacuation progress (number of evacuated, number of remaining personnel, etc.). The system monitors the current location, real-time concentration of toxic agents at the site, changes in wind speed and direction, and equipment operating status (whether the decontamination equipment is working properly). Dynamic correction and rescheduling refer to the cloud comparing real-time feedback data with the original prediction results (such as the original spread trend and the original personnel evacuation plan). If a deviation is found (such as the actual spread speed being faster than predicted or personnel evacuation being obstructed), the response plan is re-matched based on the new site conditions, the command and decision instructions are adjusted, and the data is reissued. In specific implementation, the command and decision instructions generated by S501 (such as evacuation routes, decontamination ranges, and personnel deployment) are first pushed to the corresponding edge-side execution nodes (rescue personnel terminals receive evacuation instructions, decontamination equipment nodes receive start instructions, and evacuation indicator lights receive guidance instructions) with low latency through the edge gateway of the cloud-edge collaborative architecture, ensuring rapid on-site response. At the same time, the edge-side execution nodes collect on-site status data in real time (such as the rescue terminal reporting "Personnel 001 has arrived at the refuge point, Personnel 002 is temporarily stuck 30 meters away due to temporary congestion on the east side of the factory area" and the on-site sensors reporting "Wind speed increased to 3.0"). The diffusion direction is predominantly northeast, with a current concentration of 1.3 ppm within a 50-meter radius. The data is transmitted back to the cloud via a high-speed communication link. After receiving the data, the cloud-based emergency command platform compares it with the original prediction (original spread direction northwest, all personnel to be evacuated within 15 minutes), and finds deviations (spread direction shift, personnel evacuation obstructed). Based on the new situation, it recalculates: adjusting the evacuation route (planning a backup passage on the west side for personnel 002), expanding the decontamination area to cover the newly spread area on the northeast side, and allocating one nearby rescue worker to assist in clearing congestion. A revised command decision is generated and reissued to the edge execution node through the cloud-edge collaborative architecture, completing dynamic rescheduling. For example, continuing with the aforementioned chlorine leak scenario, the original instruction required personnel 002 to evacuate along the southeast passage. The edge side reports that this passage is congested due to equipment failure, and the real-time wind speed has increased to 3.0. This caused chlorine gas to spread northeast. The cloud immediately corrected the instructions: "Personnel 002 should evacuate to refuge point 2 (120 meters away) via the backup passage on the west side. The decontamination area should be expanded to a 15-meter range on the northeast side. Rescue personnel Li XX (ID: 101) should be dispatched to the passage on the east side to alleviate congestion." After the corrected instructions were quickly received by the edge terminal, they were immediately adjusted and executed on site to ensure that the response measures were always adapted to the real-time scenario.

[0054] In one embodiment, a dynamic risk assessment system for chemical defense command based on multimodal data fusion is provided. This system corresponds to the dynamic risk assessment method for chemical defense command based on multimodal data fusion described in the previous embodiment. The dynamic risk assessment system for chemical defense command based on multimodal data fusion includes: The data acquisition module is used to collect multi-source data on chemical defense operations in the chemical defense area and perform data preprocessing. The multi-source data on chemical defense includes structured sensing data and unstructured audio-visual interactive data. The feature extraction module is used to extract structured perception features and audio-visual interaction features based on the preprocessed chemical defense multi-source data, and to calculate the cross-modal feature correlation degree of structured perception features and audio-visual interaction features based on preset risk calculation rules. The feature integration module is used to extract cross-modal temporal correlation features between structured perception features and audio-visual interaction features using the cross-modal feature correlation degree, and to integrate structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features to construct a multimodal fusion feature set; The risk prediction module is used to input the multimodal fusion feature set into a pre-trained time-series risk evolution model, calculate the dynamic temporal correlation of risks and the nonlinear coupling relationship of multiple factors, and output prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. The command and dispatch module is used to generate corresponding command and decision instructions based on the prediction results. It writes the prediction results, command and decision instructions, the collection timestamps of chemical defense multi-source data, and the identity information corresponding to the individual risk level of personnel into the chemical defense command alliance chain, and completes chemical defense command and dispatch through a preset cloud-edge collaborative architecture.

[0055] Specific limitations regarding a dynamic risk assessment system for chemical defense command based on multimodal data fusion can be found in the above description of the limitations of a dynamic risk assessment method for chemical defense command based on multimodal data fusion, and will not be repeated here. Each module in the aforementioned dynamic risk assessment system for chemical defense command based on multimodal data fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0056] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a dynamic risk assessment method for chemical defense command based on multimodal data fusion.

[0057] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a dynamic risk assessment method for chemical defense command based on multimodal data fusion.

[0058] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a method for dynamic assessment of chemical defense command risks based on multimodal data fusion.

[0059] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0060] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0061] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dynamic risk assessment method for chemical defense command based on multimodal data fusion, characterized in that, Includes the following steps: Collect multi-source data on chemical defense operations in the chemical defense area and perform data preprocessing. The multi-source data on chemical defense includes structured sensing data and unstructured audio-visual interaction data. Based on the preprocessed chemical defense multi-source data, structured perception features and audio-visual interaction features are extracted, and the cross-modal feature correlation degree of structured perception features and audio-visual interaction features is calculated based on the preset risk calculation rules. The cross-modal feature correlation degree is used to extract cross-modal temporal correlation features between structured perception features and audio-visual interaction features. The structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features are integrated to construct a multimodal fusion feature set. The multimodal fusion feature set is input into a pre-trained time-series risk evolution model to calculate the dynamic temporal correlation of risks and the nonlinear coupling relationship of multiple factors, and outputs prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. Based on the prediction results, corresponding command and decision instructions are generated. The prediction results, command and decision instructions, collection timestamps of chemical defense multi-source data, and identity information corresponding to the individual risk level of personnel are written into the chemical defense command alliance chain. Chemical defense command and dispatch are completed through a preset cloud-edge collaborative architecture.

2. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 1, characterized in that, The step of calculating the cross-modal feature correlation between structured perception features and audiovisual interaction features based on preset risk calculation rules specifically includes: The structured perception features and audiovisual interaction features are subjected to asymmetric time window slicing processing, wherein the time window of the audiovisual interaction features has a preset time delay offset relative to the time window of the structured perception features, in order to match the physical time delay of the spread of toxic agents. Within the asymmetric time window defined by the time delay offset, anomaly event burst vectors are extracted from the audio-visual interaction features, and physical quantity mutation vectors are extracted from the structured perception features. A physical causal matrix for chemical defense risks is constructed, wherein the physical causal matrix for chemical defense risks predefines the physical mapping relationship between the sudden vector of abnormal events and the sudden change vector of physical quantities under a preset operation scenario; Based on the aforementioned physical causal matrix of chemical defense risks, a causal confidence weighted calculation is performed on the sudden event vector of the abnormal event and the sudden change vector of the physical quantity to obtain a causal confidence score: If the abnormal event burst vector contains preset dangerous semantic keywords or voiceprint fluctuation features, and the physical quantity mutation vector shows an exponential increase in concentration, then the causal confidence score is the first association threshold. If the sudden vector of the abnormal event is environmental steady-state noise and the sudden vector of the physical quantity has no abnormal fluctuation, then the causal confidence score is the second correlation threshold. Based on the causal confidence score, the cross-modal feature correlation degree is quantified.

3. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 1, characterized in that, The step of extracting cross-modal temporal correlation features between structured perception features and audiovisual interaction features using the cross-modal feature correlation degree specifically includes: Based on the cross-modal feature correlation degree, the structured perception features and the audio-visual interaction features are temporally segmented and aligned, and continuous time segments with correlation degrees higher than a preset threshold are extracted as risk evolution intervals. Within the risk evolution range, the physical quantity temporal evolution vector of the structured perception features and the abnormal behavior temporal evolution vector of the audio-visual interaction features are extracted respectively. A cross-modal time delay coupling matrix is ​​constructed, which defines the dynamic coupling relationship between the physical quantity time-series evolution vector and the abnormal behavior time-series evolution vector under different preset time delay offset levels; Based on the cross-modal time delay coupling matrix, the dynamic cross-correlation peak value between the physical quantity time-series evolution vector and the abnormal behavior time-series evolution vector is calculated, and the dynamic cross-correlation peak value is determined as the cross-modal time-series correlation feature.

4. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 1, characterized in that, The construction and training of the time-series risk evolution model specifically includes: Acquire multiple sets of historical multimodal sequence data from historical chemical defense operation data. Each set of historical multimodal sequence data corresponds to a historical leakage event or simulated event under a preset operation scenario, and is marked with the actual diffusion value and risk label of the corresponding event. Data preprocessing and feature fusion processing are performed on each set of historical multimodal sequence data to obtain several sets of historical multimodal fusion feature set samples; Construct a multi-branch model, the model including: The temporal coding module is used to receive historical multimodal fusion feature set samples for feature extraction and temporal evolution, in order to calculate the dynamic temporal correlation of risk and the nonlinear coupling relationship of multiple factors, and output the encoded feature vector; The first branch is used to receive encoded feature vectors, perform regional risk trend aggregation processing, and output the regional risk level. The second branch is used to receive encoded feature vectors to perform individual risk assessment and output the individual risk level. The third branch is used to receive encoded feature vectors, perform spatiotemporal diffusion trend deduction processing, and output the toxic agent diffusion trend within a preset time period. Based on the actual diffusion value and risk label, a training objective function is defined, and the training objective function is regularized with diffusion dynamics constraints to construct an optimization objective with physical constraints. Using historical multimodal fusion feature set samples as input, the constructed multi-branch model is trained along the optimization objective with physical constraints, so that the regional risk level, individual risk level and toxic agent diffusion trend output by the model meet the preset error threshold and satisfy the physical boundary conditions under the preset operation scenario, thus obtaining the trained time-series risk evolution model.

5. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 4, characterized in that, The step of inputting the multimodal fusion feature set into a pre-trained time-series risk evolution model and calculating the dynamic temporal correlation and multi-factor nonlinear coupling relationship of risk specifically includes: The multimodal fusion feature set is input into the trained temporal risk evolution model; Based on the temporal coding module in the temporal risk evolution model, feature extraction and temporal evolution are performed on the multimodal fusion feature set to generate coded feature vectors; The temporal response sequence and evolution rate difference of each temporal component in the multimodal fusion feature set are captured along the time axis by the temporal coding module. The temporal response sequence reflects the response time offset relationship between the structured perception feature sequence and the audio-visual interaction feature sequence, as well as the evolution trend of cross-modal temporal correlation features. The evolution rate difference reflects the relative deviation between the concentration change rate of the structured perception feature sequence and the abnormal activity change rate of the audio-visual interaction feature sequence. The cross-modal feature correlation degree is used as the confidence gate of the current time window and is used to weight and adjust the effective weight of the temporal response sequence to quantify the risk dynamic temporal correlation relationship. The collaborative correlation between structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features is analyzed, feature coupling weights are constructed, and the temporal response sequence and evolution rate difference are mapped to multi-source feature contributions according to the feature coupling weights. The multi-source feature contributions are fused through the gating mechanism of the temporal coding module to quantify the multi-factor nonlinear coupling relationship.

6. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 5, characterized in that, The steps for outputting the predicted results of regional risk level, individual personnel risk level, and the trend of toxic agent spread within a preset time period specifically include: Based on the encoded feature vector and feature coupling weight, the regional risk trend aggregation value is calculated through the first branch of the time-series risk evolution model, and the regional risk level is output according to the preset regional risk classification mapping rule. Based on the encoded feature vector and feature coupling weights, the individual risk assessment value is calculated through the second branch of the time-series risk evolution model, and the individual risk level is output according to the preset individual risk classification mapping rules. Based on the encoded feature vector and feature coupling weights, the spatiotemporal diffusion trend is deduced through the third branch of the time-series risk evolution model, and the diffusion trend of toxic agents within a preset time period is output.

7. The method for dynamic risk assessment of chemical defense command based on multimodal data fusion as described in claim 1, characterized in that, The steps of generating corresponding command and decision instructions based on the prediction results, writing the prediction results, command and decision instructions, the collection timestamps of multi-source chemical defense data, and the identity information corresponding to the individual risk level of personnel into the chemical defense command alliance chain, and completing the chemical defense command and dispatch through a preset cloud-edge collaborative architecture, specifically include: Based on the regional risk level, individual risk level, and toxic agent spread trend, a pre-set disposal plan library is matched to generate command and decision instructions that include evacuation route planning, delineation of decontamination areas, and deployment of emergency personnel. The prediction results, command and decision instructions, collection timestamps of multi-source chemical defense data, and identity information corresponding to the individual risk level of personnel are encapsulated into chemical defense operation data blocks, written into the chemical defense command alliance chain for distributed storage, and data access permissions are subject to hierarchical control. Through a pre-defined cloud-edge collaborative architecture, command and decision instructions are sent to edge execution nodes, while the cloud dynamically corrects and reschedules the command and decision instructions based on real-time feedback of on-site status data.

8. A dynamic risk assessment system for chemical defense command based on multimodal data fusion, used to implement the steps of the dynamic risk assessment method for chemical defense command based on multimodal data fusion as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source data on chemical defense operations in the chemical defense area and perform data preprocessing. The multi-source data on chemical defense includes structured sensing data and unstructured audio-visual interactive data. The feature extraction module is used to extract structured perception features and audio-visual interaction features based on the preprocessed chemical defense multi-source data, and to calculate the cross-modal feature correlation degree of structured perception features and audio-visual interaction features based on preset risk calculation rules. The feature integration module is used to extract cross-modal temporal correlation features between structured perception features and audio-visual interaction features using the cross-modal feature correlation degree, and to integrate structured perception features, audio-visual interaction features, cross-modal feature correlation degree and cross-modal temporal correlation features to construct a multimodal fusion feature set; The risk prediction module is used to input the multimodal fusion feature set into a pre-trained time-series risk evolution model, calculate the dynamic temporal correlation of risks and the nonlinear coupling relationship of multiple factors, and output prediction results including regional risk level, individual risk level of personnel and the trend of toxic agent spread within a preset time. The command and dispatch module is used to generate corresponding command and decision instructions based on the prediction results. It writes the prediction results, command and decision instructions, the collection timestamps of chemical defense multi-source data, and the identity information corresponding to the individual risk level of personnel into the chemical defense command alliance chain, and completes chemical defense command and dispatch through a preset cloud-edge collaborative architecture.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the dynamic risk assessment method for chemical defense command based on multimodal data fusion as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the dynamic risk assessment method for chemical defense command based on multimodal data fusion as described in any one of claims 1-7.