Emergency decision generation method based on large language model in education scene

Abstract

The invention discloses an emergency decision generation method based on a large language model in an education scene, and relates to the technical field of education emergency management, and the method comprises the steps: sensing and preprocessing multi-modal information, deploying heterogeneous equipment to collect data, and storing the data in an edge node through space-time alignment and feature fusion; constructing and finely adjusting a domain enhancement model; modeling a dynamic risk field, calculating a risk value in combination with event characteristics, and updating a grade in real time; generating a multi-objective decision, fusing cases and rule reasoning to generate a scheme, and optimizing the scheme; performing decision execution and feedback, pushing an analysis instruction to a terminal, and performing real-time monitoring and dynamic adjustment; carrying out three-dimensional situation visualization, and constructing a digital twinning presentation situation; and redisk and optimization: recording data redisk and incrementally training the model. According to the invention, the problems of weak dynamic response, poor adaptation and low efficiency of the traditional campus emergency decision are solved; and through multi-modal perception, dynamic risk modeling, age adaptation decision and closed-loop optimization, the recognition and decision accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of education emergency management, and in particular to an emergency decision generation method based on a large language model in an education scenario. BACKGROUND

[0002] Campus is a place with special crowd structure and personnel concentration. Fire, earthquake, stampede, and sudden illness occur frequently. The timeliness and scientificity of emergency disposal are directly related to the safety of teachers and students. With the advancement of education informatization, campus emergency management has gradually moved away from the traditional "manual patrol + fixed plan" mode, but the existing technical system still has many bottlenecks to be solved, which is difficult to adapt to the complex and changing campus emergency needs.

[0003] Traditional campus emergency decision-making relies too much on preset plans and manual experience, and lacks dynamic adaptation ability. The plan is mostly based on typical scenarios, such as fixed evacuation routes for fire and unified safety instructions for earthquake, but actual emergency events often show dynamic evolution characteristics - the spread speed of fire is affected by wind direction and flammable material distribution, and the risk of stampede changes with crowd density in real time. Preset plans cannot respond to these dynamic variables. Manual decision-making is limited by information acquisition lag and subjective judgment, for example, the process of teachers reporting abnormalities step by step takes several minutes, and the selection of evacuation routes may lead to secondary risks due to lack of information about remote congestion. More importantly, the differences in emergency needs of teachers and students of different ages are ignored. The young children group needs simple visual instructions, and the college students group can undertake part of the self-disposal tasks, but the existing decision-making scheme mostly adopts the "one-size-fits-all" mode, and the compliance rate and execution effect are greatly discounted.

[0004] The weak data perception and fusion capability further restricts the quality of emergency decision-making. Although cameras, temperature and humidity sensors and other devices are deployed in the campus, they mostly operate in isolation, lacking unified spatio-temporal alignment and feature fusion mechanism: temperature data and smoke data are not synchronized, leading to misjudgment in the early stage of fire, cameras can only capture visual information and miss abnormal sound precursors, and large positioning errors of personnel position cannot support accurate evacuation. Data sharing in multi-campus scenarios also faces privacy compliance problems. Sensitive data such as the location and health status of teachers and students in the campus cannot be transmitted at will, making it difficult to reuse cross-campus emergency experience, and the model training lacks sufficient and diversified data, and the response capability to rare extreme events (such as strong earthquake superimposed with dangerous chemical leakage) is almost blank.

[0005] There is a clear "scene adaptation gap" and "decision-making closed loop missing" in the application of large language models. The existing technology directly applies the general large language model without specialization and optimization for the campus field, and lacks understanding of professional terms such as "laboratory hazardous chemicals" and "teaching building evacuation layer". The output decision often deviates from the actual situation of the campus. The decision-making generation and execution process lacks an effective feedback mechanism and cannot dynamically adjust according to real-time situations such as evacuation congestion and equipment failure. Post-mortem only stays at the event recording level and does not form a closed loop of "data-decision-effect-optimization", leading to the repeated occurrence of similar errors. For example, in a stampede incident at a certain primary school, the preset evacuation route caused congestion at the stairwell because it did not take into account the slow movement of lower grade students. This problem was not included in the subsequent plan optimization, exposing the rigidity and lag of the existing system.

[0006] These problems overlap, leading to a chain reaction of "slow recognition, biased decision-making, chaotic execution, and weak post-mortem" in campus emergency response. According to statistics, under the existing technology, the average time for campus emergency decision-making is 5-8 minutes, and the accuracy rate of complex scene recognition is less than 80%. The low efficiency caused by poor decision-making adaptability prolongs the evacuation time by more than 30%. As the campus expands and emergency scenarios become more complex, traditional technology cannot meet the campus emergency needs of "second-level response, precise adaptation, and scientific disposal". It is urgent to build an intelligent emergency decision-making technology system that integrates multi-modal perception, dynamic risk modeling, specialized model decision-making, and full-cycle optimization to fill the technical gap in campus emergency intelligence. SUMMARY

[0007] The present application proposes an emergency decision-making generation method based on large language models in an educational scenario to solve the problems mentioned in the existing technology.

[0008] 2. To achieve the above-mentioned purpose, the present application adopts the following technical solution: an emergency decision-making generation method based on large language models in an educational scenario, comprising the following steps: Multi-modal emergency information real-time perception and space-time alignment. Deploy heterogeneous devices in key areas of the campus to collect multi-source data, and complete space alignment through coordinate system conversion matrix. After signal interference removal and feature extraction, compress into a unified feature vector by the Transformer encoder; Domain-enhanced large language model construction and fine-tuning. Construct a campus emergency case annotation corpus and fine-tune the 20B parameter base model in two stages: freeze 90% of the parameters to learn domain knowledge using LoRA technology; design a scene recognition Prompt template to input fused features and output classification results, and start the knowledge distillation mechanism for ambiguous scenes; Dynamic risk field modeling and level evolution. Based on the recognition results, call the event-specific risk assessment framework and calculate the initial risk value using the entropy weight-TOPSIS algorithm. Construct a risk evolution differential equation; Multi-target emergency decision-making intelligent generation, large language model retrieval decision knowledge base, adopts case and rule reasoning fusion strategy to generate initial scheme set; Construct an optimization function with evacuation time, casualty probability and resource consumption as the target, and solve the optimal solution by improved NSGA-III algorithm; Distributed decision-making instruction execution and feedback, the optimal solution is analyzed into multi-terminal instructions; Real-time monitoring of people flow state by millimeter wave radar, congestion or deviation triggers adjustment mechanism, large language model generates local optimization instructions; Three-dimensional situation visualization and interactive command, build a campus digital twin model, render the situation map through edge and cloud collaboration, and design an interactive interface to support touch to view details; Full-cycle data review and model evolution, automatically record the full-process data chain of the emergency, use knowledge graph to build the event timeline; Through comparative learning to mine decision-making failure modes and build an error case library, use incremental training and reinforcement learning to optimize the model.

[0009] Further, it also includes a cross-modal attention fusion mechanism, which constructs a temperature, sound, and image three-modal interactive attention matrix, mines the intra-modal spatio-temporal correlation through a self-attention module, calculates the inter-modal dependency coefficient through a cross-attention module, dynamically adjusts the feature weight value of each modality, and the temperature feature weight increases nonlinearly with the degree of deviation from normal temperature, the sound feature weight and the abnormal sound confidence present a quadratic function relationship, the image feature weight is adjusted according to the personnel density gradient, and the fused feature vector is reduced by principal component analysis.

[0010] Further, it also includes crowd flow dynamics prediction, based on continuum mechanics to build a crowd evacuation model, Where ρ(x,y,t) is the personnel density at coordinate (x,y) at time t, v(x,y,t) is the crowd flow velocity at that point, and ∇· is the divergence operator, which describes the conservation relationship of the crowd through the equation; Combined with the social force model to calculate the acceleration field, and use the finite difference method to solve the crowd distribution in the next 30s, and the large language model optimizes the evacuation route according to the prediction results to avoid potential congestion areas.

[0011] Further, it also includes an adaptive decision-making scheme generation mechanism, which takes into account the differences in cognitive and action ability among different age groups, builds a decision adaptation evaluation system, focuses on visual instruction proportion and step simplification in kindergarten scenarios, increases voice instruction clarity and guide personnel ratio in primary school scenarios, strengthens autonomous decision-making space in middle school scenarios, and focuses on professional instruction accuracy in university scenarios; Design a decision complexity adjustment algorithm to dynamically adjust the instruction density according to the real-time monitored average crowd movement speed.

[0012] Further, it also includes distributed emergency resource collaborative scheduling, builds a resource-demand dynamic matching network, abstracts fire fighting equipment, medical resources and manpower support as resource nodes, and quantifies attribute values according to the ability dimension; an improved ant colony algorithm is used to solve the optimal scheduling path, and a time attenuation factor is introduced, and when the total amount of resources is insufficient, a priority sorting mechanism is started; through federated learning, resource sharing among multiple campuses is realized.

[0013] Further, it also includes multi-modal multi-language instruction adaptation, for foreign teachers and students and special groups on campus, a full-modal instruction generation system is built, four language real-time conversion of Chinese, English, Japanese and Korean is supported for text instructions, emotion speech synthesis is adopted for voice instructions, visual instructions include dynamic icons and sign language animation, and vibration codes are output through smart bracelets for tactile instructions; an instruction effectiveness evaluation function is defined, wherein E is the comprehensive effectiveness, is the modal weight, is the recognition error rate of the modal.

[0014] Further, it also includes an adversarial scene enhancement training, for extreme events with insufficient training data coverage, a virtual scene library is built using a generative adversarial network, the generator generates extreme samples based on the real event feature distribution, and the discriminator distinguishes between real and virtual samples; the generated samples are mixed into the training set at a ratio of 1:3, and the large language model is adversarially fine-tuned.

[0015] Further, it also includes an edge and cloud collaborative reasoning mechanism, scene recognition, risk assessment and decision generation tasks are split according to computational complexity, real-time tasks are processed by edge nodes; complex tasks are processed by cloud servers, model compression technology is used to reduce the size of the large language model to 30% of the original, data transmission is ensured through 5G slicing technology; a dynamic task scheduling strategy is designed, when the network delay is >100ms, some cloud tasks are automatically sunk to the edge.

[0016] Further, it also includes emergency capability digital twin evaluation, an index system for evaluating the emergency capability of a campus is built, 12 specific indexes are set from the perception layer, the decision layer and the execution layer; based on the digital twin model, 20 typical scenarios are simulated, index data is collected to calculate the comprehensive score, and a causal inference algorithm is used to locate weak links.

[0017] Further, it also includes privacy-protected data review, federated learning framework is used to realize collaborative review of multi-campus data, each campus completes data cleaning and feature extraction locally, model parameters are exchanged through homomorphic encryption technology, and the aggregation server uses secure multi-party computation to fuse parameters; the data involving personal information is protected by differential privacy, and a cross-campus event correlation network is built using federated knowledge graph.

[0018] Compared with the prior art, the present application has the following advantages: The breakthrough of multi-modal perception and fusion technology lays a solid foundation for accurate decision-making. By deploying heterogeneous sensing devices, real-time collection of multi-dimensional data such as temperature, sound, and personnel density is achieved. Combined with spatio-temporal alignment and feature fusion mechanisms, the problem of traditional data isolation and asynchronization is solved. The cross-modal attention fusion mechanism can dynamically adjust the weight of each data, effectively capturing the correlation features of abnormal signals. Even in complex scenarios such as fire superimposed on earthquakes, high-accuracy scene recognition can be achieved, avoiding decision-making errors caused by incomplete information from the source.

[0019] Dynamic risk modeling and specialized decision generation significantly improve the adaptability and scientificity of decisions. By abandoning the traditional fixed plan mode, the risk evolution model tracks the event spread path and risk level in real time, enabling the decision to accurately respond to dynamic changes. After fine-tuning and instruction optimization of the large language model with campus domain corpus, its understanding ability for campus professional scenarios is significantly enhanced. Combined with the decision-making scheme generated by the multi-objective optimization algorithm, it can balance evacuation efficiency, safety risk, and resource consumption, etc. More importantly, the adaptation mechanism for different groups such as children and university students greatly improves the compliance rate and execution effect of decision instructions, solving the adaptability problem of "one-size-fits-all" decisions.

[0020] Distributed execution and real-time feedback mechanism ensures the efficiency of decision implementation. The decision is analyzed into multi-modal instructions such as voice, vision, and touch, which are adapted to the needs of different terminals and special groups, avoiding the delay caused by poor instruction transmission. The real-time feedback link constructed by devices such as millimeter wave radar can timely capture problems such as evacuation congestion and route deviation, triggering dynamic adjustment of decisions, realizing the dynamic closed loop of decision, execution, and correction, and effectively reducing secondary risks.

[0021] Digital twin visualization and collaborative optimization further enhance emergency command capabilities. The campus digital twin model visually presents information such as personnel location, risk range, and resource status, providing global situational awareness for command personnel and significantly improving multi-department collaboration efficiency. The edge and cloud collaborative reasoning mechanism solves the problem of decision continuity in unstable network scenarios, ensuring uninterrupted emergency response.

[0022] Model iteration and privacy protection mechanism realize the continuous evolution and compliant development of emergency capabilities. The full-cycle review system uses knowledge graph and contrastive learning to identify decision-making flaws, driving incremental optimization of the model, enabling the system to continuously learn from historical events. Federated learning and privacy protection technology enable multi-campus data collaborative training while ensuring the security of sensitive campus data, improving the model's ability to respond to rare scenarios and complying with data privacy regulations.

[0023] Overall, the application constructs a campus intelligent emergency decision-making system with accurate perception, dynamic decision-making, efficient execution and continuous evolution, converts the general intelligence of the large language model into special ability in the campus scene, effectively fills the short board of the prior art, provides strong technical support for campus emergency management, and effectively safeguards the life safety of teachers and students. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A schematic block diagram of an emergency decision generation method based on a large language model in an education scene is provided for the application. Figure 2 A comparison bar chart of identification accuracy for different emergency scenes; Figure 3 A comparison line chart of fire risk value dynamic evolution; Figure 4 A comparison combination chart of evacuation efficiency for different age groups; Figure 5 A comparison bar chart of cross-campus emergency resource response time; Figure 6 A model iteration optimization effect trend chart. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0026] In the description of the application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the application.

[0027] In addition, the terms "first", "second", etc. are used only for descriptive purposes and should not be construed as indicating or implying relative importance or an indicated number of technical features. Therefore, the features defined as "first", "second", etc. can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited. In addition, the terms "mounting", "connecting", "connecting" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail below with reference to the accompanying drawings.

[0028] 3. Refer to Figures 1 to 6 A method for generating emergency decision based on large language model in education scene, comprising the following steps: Real-time multi-modal emergency information perception and spatio-temporal alignment, deploy heterogeneous perception devices in key areas of the campus, infrared thermal imager collects environmental temperature field distribution at a frequency of 1Hz (resolution 640x512, temperature measurement range -20℃~150℃), array microphone collects 300-3000Hz band sound (sampling rate 44.1kHz, sensitivity -38dB±3dB), binocular camera generates personnel density heat map through stereo matching (update period 0.5s, error ≤3%), Internet of Things sensor monitors electrical equipment operating parameters (voltage fluctuation ±2%, current sampling frequency 50Hz); Time stamp interpolation method (accuracy 1ms) is used to realize time synchronization of multi-source data, coordinate system conversion matrix (error ≤0.2m) is used to complete spatial alignment, variational mode decomposition is used to remove device heat dissipation interference in temperature signal (signal-to-noise ratio improved to 40dB), abnormal sound features are extracted based on mel frequency cepstral coefficients (128-dimensional vector), compressed into unified feature vector (dimension 512) by Transform encoder, stored in distributed cache (response delay ≤80ms) of edge node; Domain-enhanced large language model construction and fine-tuning, build a labeled corpus containing 200,000 campus emergency cases (covering 18 types of events such as fire, earthquake, and stampede, each containing event characteristics, disposal process, and result evaluation triplets), fine-tune the basic large language model (parameter size 20B) in two stages, the first stage uses LoRA technology to freeze 90% of the parameters (rank value 32, learning rate 2e-5) to learn domain knowledge, and the second stage optimizes the output format through instruction fine-tuning (3000 campus-specific instructions) and adds a domain glossary containing 450 terms such as "laboratory hazardous chemicals" and "teaching building evacuation floor" (embedding dimension 768); design a scene recognition Prompt template, including event type, impact range, trigger condition, and other 10 key dimensions, the model outputs the classification result after inputting the fused features (Top-1 accuracy ≥ 96%, processing time ≤ 1.5s), and the knowledge distillation mechanism is started for ambiguous scenarios with confidence of 60%-80% (fusion of 3 lightweight model decisions); Dynamic risk field modeling and grade evolution, based on the recognition results, call the event-specific risk assessment framework, which includes personnel exposure coefficient (classroom 0.9, playground 0.6, laboratory 1.2), environmental vulnerability index (flammable material density 0-8 kg / m 2 corresponding to 0-1.0), and emergency resource accessibility (0-10 min corresponding to 1.0-0.2), use the entropy weight-TOPSIS combination algorithm to calculate the initial risk value (0-100); build a risk evolution differential equation, introduce an event diffusion rate parameter (fire 0.8 m / min, stampede 1.5 m / min), update the risk field distribution every 2s (spatial resolution 0.5m×0.5m), predict the influence of the spread path through the risk gradient direction, and generate dynamic risk levels (1-5) and key node warning times (error ≤ 10%); Multi-objective emergency decision-making intelligent generation, large language model retrieves decision knowledge base (including national campus emergency specifications and 5000+ historical cases) based on risk level, uses case reasoning and rule reasoning fusion strategy to generate initial scheme set (including evacuation route, rescue steps, resource allocation, etc.); build a multi-objective optimization function, with evacuation time, casualty probability, and resource consumption as optimization objectives (weights 0.4, 0.4, and 0.2 respectively), solve the Pareto optimal solution through improved NSGA-III algorithm (population size 100, iteration 50 generations), check the feasibility of the scheme (route capacity ≥ personnel quantity 1.3 times, rescue force ≥ casualty estimate 1.5 times), output 3 candidate schemes and decision basis (generation time ≤ 4s); Distributed decision-making instruction execution and feedback, the optimal solution is parsed into multi-terminal executable instructions, the broadcast system is converted into voice instructions (clarity ≥ 95 dB, coverage ≥ 50 m), electronic screen displays dynamic evacuation route map (update frequency 1 Hz), emergency indicator light output optical signal (red flash 2 Hz represents prohibition, green flash 1 Hz represents passage), Internet of Things gateway pushes instructions to faculty and staff intelligent terminals (transmission delay ≤ 40 ms); The millimeter wave radar deployed in the corridor (sampling frequency 15 Hz) monitors the flow rate in real time (error ≤ 0.08 m / s), and when congestion (density > 2.5 people / m 2 ) or deviation (deviation > 2 m) is detected, the decision adjustment mechanism is triggered, and the large language model generates local optimization instructions based on real-time data (response time ≤ 0.8 s); Three-dimensional situation visualization and interactive command, build a digital twin model of the campus (accuracy 0.1 m), integrate personnel location (GPS+UWB fusion positioning, error ≤ 0.3 m), risk diffusion range (grid display), resource status (fire / medical force real-time coordinates) and other information, through edge-cloud collaborative rendering (frame rate 30 fps) to generate dynamic situation map; Design command interaction interface, support touch circle selection area to view details (response delay ≤ 100 ms), provide risk trend prediction curve (future 60 s), resource scheduling suggestion list, evacuation progress bar and other visualization components, push key indicators (evacuation completion rate, risk reduction rate) to the command center; Full-cycle data review and model evolution, automatically record the emergency process data chain (original perception data, decision instructions, execution logs, disposal results), use knowledge graph technology to build event timeline (accuracy 1 s), extract key node deviation (such as predicted evacuation time and actual deviation rate); Through comparative learning to mine decision failure modes (identify errors, scheme defects), build error case library (every 50 cases accumulated trigger model update); Use incremental training strategy to optimize large language model (learning rate 1e-5, freeze 80% parameters), adjust decision weights combined with reinforcement learning (reward function based on disposal effect design), synchronize update knowledge base rules (new case pass rate ≥ 90%).

[0029] This invention also includes a cross-modal attention fusion mechanism, constructing a three-modal interactive attention matrix of temperature, sound, and image. The self-attention module mines the spatiotemporal correlation within each modality (8s time window, 5m spatial neighborhood), and the cross-attention module calculates the intermodal dependency coefficient (the Pearson coefficient for temperature and smoke concentration ≥0.8 enhances the weight). The weight values ​​of each modal feature are dynamically adjusted. The temperature feature weight increases non-linearly with the degree of deviation from normal temperature (25℃) (the weight increases to 1.5 times the initial value when the deviation is 10℃). The sound feature weight has a quadratic function relationship with the confidence level of abnormal sounds (0-1) (the weight reaches its peak when the confidence level is 0.8). The image feature weight is determined according to the personnel density gradient (0-4 people / m²). 2 Segmented adjustment (gradient > 2 people / m) 2 The time weight is increased by 20%, and the fused feature vector is reduced in dimensionality by principal component analysis (retaining 97% of the information), which improves the recognition accuracy of complex scenes (such as earthquakes accompanied by fires) by 10-15% and increases the feature processing efficiency by 25%.

[0030] This invention also includes crowd flow dynamics prediction, constructing a crowd evacuation model based on continuum mechanics. Where ρ(x,y,t) is the population density (people / m²) at coordinates (x,y) at time t. 2 v(x,y,t) represents the flow velocity of the crowd at that point (m / s), and ∇・ is the divergence operator. This equation describes the conservation relationship of the crowd. The acceleration field is calculated by combining the social force model (considering the repulsive force between people, the attractive force of the walls, and the driving force of the target). The finite difference method is used to solve the crowd distribution in the next 30 seconds (prediction error ≤12%). The large language model optimizes the evacuation route based on the prediction results, avoiding potential congestion areas (such as stairwells and corners), thereby improving the evacuation efficiency by 20-30%, especially for the path planning of multi-story evacuation in teaching buildings.

[0031] This invention also includes an adaptive decision-making mechanism. It constructs a decision-making suitability assessment system to address the differences in cognitive and behavioral abilities among different age groups (preschool / primary / secondary / university). For preschool scenarios, it emphasizes the proportion of visual instructions (≥60%) and the simplification of steps (≤3 steps). For primary school scenarios, it increases the clarity of voice instructions (vocabulary ≤5000 commonly used words) and the ratio of guides (1:10). For secondary school scenarios, it strengthens the space for autonomous decision-making (reserving 30% flexibility). For university scenarios, it emphasizes the accuracy of professional instructions (terminology accuracy ≥98%). A decision complexity adjustment algorithm is designed to dynamically adjust the instruction density (2-5 instructions per minute) based on the real-time monitored average movement speed of the crowd (preschool ≤0.8m / s, university ≥1.2m / s), thereby increasing the decision compliance rate of different groups by 15-20% and reducing the error rate to below 5%.

[0032] In the present application, distributed emergency resource cooperative scheduling is also included, a resource-demand dynamic matching network is constructed, fire-fighting equipment (fire extinguishers, fire hydrants), medical resources (AED, first aid kit, medical staff), human support (teachers, security guards) are abstracted as resource nodes, and the attribute values are quantified according to the ability dimension (fire-fighting capacity, first aid level, command authority); the improved ant colony algorithm is used to solve the optimal scheduling path (pheromone evaporation coefficient 0.1, heuristic factor 2.0), and a time decay factor is introduced (the resource response value decays exponentially with time, and the half-life is 30s), and when the total amount of resources is insufficient, a priority sorting mechanism is started (severe injuries > dense crowds > equipment protection, weight ratio 5:3:2); through federated learning, multi-campus resource sharing is realized (model parameters are transmitted in encrypted form, and local data is not leaked), which shortens the resource response time by 25-35%, and improves the resource utilization rate by 20-25%.

[0033] In the present application, multi-modal multi-language instruction adaptation is also included, for foreign teachers and students and special groups (deaf / blind) on campus, a full-modal instruction generation system is constructed, the text instruction supports real-time conversion of four languages (Chinese, English, Japanese and Korean) (translation accuracy ≥98.5%), the voice instruction adopts emotional speech synthesis (emergency degree is adjusted by speed / tone, speed range 120-200 words / minute), the visual instruction includes dynamic icons (the flickering frequency of the evacuation arrow is positively correlated with the emergency degree), sign language animation (covering 20 core instruction actions), and the tactile instruction outputs vibration coding through a smart bracelet (short vibration 1 time represents a set, continuous 3 times represent risk avoidance); define the instruction effectiveness evaluation function, wherein E is the comprehensive effectiveness (0-1), is the modal weight (text 0.2, voice 0.2, vision 0.3, touch 0.3), is the recognition error rate of the mode (all ≤5%), the modal combination is dynamically adjusted through real-time evaluation, and the instruction understanding rate of the whole population is improved to more than 96%.

[0034] In the present application, the training of the anti-scenario is also included, for the extreme events (such as strong earthquake superimposed dangerous chemical leakage) with insufficient training data coverage, a virtual scenario library is constructed using a generative adversarial network, the generator generates extreme samples (including temperature rise, crowd panic, equipment failure, etc.) based on the real event feature distribution (mean ± 3 sigma range), and the discriminator distinguishes between real and virtual samples (accuracy converges to 50%±5%); the generated samples are mixed into the training set at a ratio of 1:3, and the large language model is fine-tuned (the disturbance intensity is positively correlated with the rarity of the event, and the highest disturbance is 15% of the feature value), so that the decision success rate of the model in the unseen extreme scenario is improved by 18-25%, and the decision failure caused by data distribution deviation is avoided.

[0035] In the present application, edge, cloud collaborative inference mechanism is also included, which splits the scene recognition, risk assessment, decision generation and other tasks according to the calculation complexity, the edge node (5TOPS of computing power) processes real-time tasks (multimodal data preprocessing, simple scene recognition, local instruction pushing), and the response delay is less than or equal to 50ms; the cloud server (100TOPS of computing power) processes complex tasks (fuzzy scene recognition, multi-objective optimization, global situation construction), adopts model compression technology (knowledge distillation+quantization) to reduce the size of the large language model to 30% of the original (precision loss less than or equal to 2%), and guarantees data transmission (bandwidth greater than or equal to 100Mbps, packet loss rate less than or equal to 0.1%) through 5G slicing technology; a dynamic task scheduling strategy is designed, when the network delay is greater than 100ms, part of the cloud task is automatically sunk to the edge (such as the simplified version of risk assessment), so that the decision continuity of the system is improved by 30-40% when the network is unstable.

[0036] In the present application, emergency capability digital twin evaluation is also included, an evaluation index system of campus emergency capability is constructed, 12 specific indexes are set from three dimensions of perception layer (device coverage rate greater than or equal to 95%, identification accuracy greater than or equal to 90%), decision layer (generation time less than or equal to 5s, scheme feasibility greater than or equal to 90%), and execution layer (instruction response rate greater than or equal to 95%, resource availability rate greater than or equal to 90%); based on the digital twin model, 20 typical scenes (each type is repeated 10 times) are simulated, the index data is collected to calculate the comprehensive score (0-100), and the causal inference algorithm is used to locate the weak links (such as a certain area of perception blind area, a certain type of event decision delay); through reinforcement learning, the system parameters (such as sensor deployment density, model inference threshold) are optimized, so that the comprehensive score is improved by 15-20 points, and the disposal efficiency of key scenes is improved by 25%.

[0037] In the present application, privacy protection type data review is also included, and the federal learning framework is used to realize the collaborative review of multi-campus data, each campus completes data cleaning and feature extraction locally (without transmitting original data), exchanges model parameters through homomorphic encryption technology (encryption strength 2048 bits), and the aggregation server uses secure multi-party computation to fuse parameters (aggregation weight is positively correlated with campus size); the data related to personal information (such as student location, health status) is protected by differential privacy (adding Laplace noise, privacy budget ε=1.5), and the federal knowledge graph is used to construct a cross-campus event correlation network (only event features are retained, and identity is removed); through the mechanism, the model cross-scene adaptability is improved by 10-15% while meeting the compliance requirements of the Personal Information Protection Law, and the multi-campus collaborative disposal efficiency is improved by 20-30%.

[0038] The specific implementation of the system is further illustrated by two embodiments as follows: Embodiment 1: Primary school low-grade teaching building fire emergency decision generation (a public primary school, 6-storey teaching building, 1-3 grades, a total of 900 people) This embodiment is aimed at the characteristics of primary school students in lower grades (6-9 years old) with limited cognitive ability and slow action speed, focusing on the sudden fire scenario in the third-floor classroom of the teaching building (caused by electrical short circuit, initial burning range 2m×3m). The traditional emergency plan leads to chaotic evacuation due to complex instructions and fixed routes. The invention method realizes accurate decision-making, and the specific implementation is as follows.

[0039] 1. Multimodal information perception and preprocessing Deploy heterogeneous sensing devices on each floor of the teaching building: install 2 infrared thermal imagers on each floor corridor, deploy 4 array microphones in the classroom and stairwell, deploy 3 binocular cameras on each floor, and install Internet of Things sensors in the power distribution room and corridor.

[0040] After the fire is triggered, the infrared thermal imager collects temperature data at 1Hz, and the temperature on the third floor corridor rises from 25℃ to 68℃. After variational mode decomposition (decomposition layer number 4, penalty factor 200), the device heat interference is removed, and the signal-to-noise ratio is improved to 42dB; the array microphone captures the sound of glass breaking (frequency 1200Hz) and students shouting, and extracts 128-dimensional Mel frequency cepstral coefficients; the binocular camera generates a personnel density heat map, showing that the personnel density in the east classroom on the third floor is 3.2 people / m 2 , and the stairwell is 0.8 people / m 2 .

[0041] Align multi-source data through timestamp interpolation method (accuracy 1ms), unify device coordinates to campus geodetic coordinates (error 0.15m) through coordinate system conversion matrix, compress to 512-dimensional feature vector through Transformer encoder, and store in distributed cache of edge node, response delay 72ms.

[0042] 2. Domain-enhanced model construction and scenario identification Use a 20B parameter-based large language model, fine-tune based on a corpus containing 12,000 primary school fire emergency cases: in the first stage, freeze 90% of the parameters using LoRA technology (rank value 32, learning rate 2e-5), learn domain knowledge such as "low-age student evacuation guidance" and "teaching building low-floor fire disposal"; in the second stage, optimize the output through 3000 specialized instructions (such as "describe low-grade fire escape actions in 3 steps").

[0043] Add a glossary of 300 new terms such as "teaching building evacuation floor" and "child safety evacuation distance", design a Prompt template containing 10 dimensions such as event type, impact range, and personnel characteristics, and after inputting the fused features, the model outputs the classification result in 1.2s: "teaching building 3rd floor electrical fire, impact range 15m×10m, involving 120 low-age students", confidence 97%, no need to start the knowledge distillation mechanism.

[0044] 3. Risk field modeling and grade evolution Call fire-specific risk assessment framework: 3-layer classroom personnel exposure coefficient 0.9, corridor 0.7, stairway 0.8; burning area flammable material density (wooden desks and chairs) 6 kg / m 2 , environmental vulnerability index 0.8; the nearest fire hydrant is 8 m away (2 min access time), emergency resource accessibility 0.7. Use the entropy weight-TOPSIS algorithm to calculate the initial risk value of 82 points.

[0045] Build a risk evolution differential equation, introduce a fire spread rate of 0.8 m / min, update the risk field every 2 s (spatial resolution 0.5 m x 0.5 m), and the risk gradient direction shows a westward spread to the classroom. After 30 s, the affected area expands to 20 m x 15 m, generating a risk level of 4, and warning of the risk of stairway congestion (possibly 2.5 people / m 2 after 15 s).

[0046] 4. Multi-objective decision-making generation Large language model retrieves decision-making knowledge base, matches 32 similar cases of "low-grade teaching building fire", and generates an initial solution set of 3 sets in combination with the "Guidelines for Emergency Evacuation of Primary and Secondary Schools and Kindergartens". Build a multi-objective optimization function (evacuation time weight 0.4, casualty probability 0.4, resource consumption 0.2), and improve the NSGA-III algorithm (population 100, iteration 50 generations) to solve the optimal solution: select 2 staircases on the east side as the main evacuation channel, and 1 on the west side as the standby channel, assign 1 teacher to lead each classroom, and control the evacuation speed at 0.6 m / s.

[0047] Feasibility check shows that the east stair capacity is 4 people / s, meeting the evacuation demand of 120 people (120 ÷ 4 = 30 s), with 3 security guards + 2 school doctors assigned to support, and the rescue force meets the standard, with a decision-making generation time of 3.8 s.

[0048] 5. Decision execution and feedback The solution is analyzed as a multi-terminal instruction: the broadcast system outputs child-friendly voice ("Children follow the teacher, bend over and cover your nose to walk east stairs", speed 140 words / minute, clarity 96 dB); the corridor electronic screen displays a green dynamic arrow (update frequency 1 Hz); the emergency light flashes green 1 Hz to guide the direction; the teacher's smart bracelet pushes the class and route.

[0049] The 3-layer corridor millimeter wave radar monitors the east stairway to a human flow density of 2.3 people / m 2 , close to the congestion threshold, triggering the adjustment mechanism immediately. The large language model generates instructions based on real-time data: "Class 302 switches to the west standby staircase, and Class 303 delays 10 s to start", with a response time of 0.7 s, and the density drops to 1.8 people / m 2 after adjustment.

[0050] 6. Three-dimensional situation visualization Unity built campus digital twin model (accuracy 0.1m), integrated UWB positioning of students (error 0.25m), fire spread range (rasterized red display), real-time coordinates of school doctors / security guards. Edge-cloud collaborative rendering (frame rate 30fps) generates situation maps, command center interface displays evacuation completion rate (40% in 30s), risk reduction rate (15 minutes per minute), and supports touch circle selection of 302 class to view evacuation progress.

[0051] 7. Review and model optimization Automatic recording of full-process data chain: original temperature curve, microphone audio, decision instruction log, etc. Knowledge graph constructs event timeline (1s accuracy), extracts key deviations: actual evacuation time 42s, 12s longer than predicted, due to delay caused by student falling in 301 class.

[0052] By contrast learning, this case is classified into the "evacuation of young students in case of accident" category, and after 50 cases, incremental training of the model is started (learning rate 1e-5, 80% parameters frozen), and the learning reward function is adjusted based on the "quick guidance after falling" scenario, and the decision library rules are updated simultaneously: "one guide for every 10 students in low-grade evacuation".

[0053] 8. Execution of other modules Cross-modal fusion: temperature feature weight increased to 1.6 times the initial value due to deviation from normal temperature of 68°C, sound feature weight set to 0.35 with confidence of 0.95, image feature weight set to 0.4 due to density gradient of 2.4 people / m, and recognition accuracy after fusion is 97%. 2

[0054] Adversarial training: generate "fire + student trampling" virtual samples, mix into training set, and model decision success rate for this extreme scenario is improved.

[0055] Privacy protection: student location data is added with Laplace noise (ε=1.5), and identity is removed during review, only keeping grade and location features.

[0056] Example 2: Emergency decision-making for hazardous chemical leakage in university chemical laboratory (a certain university of science and technology, 3 campuses, 5-story laboratory in chemical building, 200 people) This example is aimed at the scenario of hazardous chemical (ethanol) leakage in a university chemical laboratory (5-story organic laboratory, pipe rupture leading to 10L leakage, volatile range diffusion), involving multi-campus resource scheduling and professional disposal requirements. Traditional solutions lack professional instructions and weak cross-campus collaboration, leading to delayed disposal. This invention method realizes efficient decision-making, and the specific implementation is as follows.

[0057] 1. Multi-modal information perception and preprocessing ​Dedicated sensing equipment is deployed in the chemical building: infrared thermal imagers (FLIRT1040) and combustible gas sensors (MQ-2, detection range 0-10000ppm) are installed in the laboratories; array microphones and binocular cameras are deployed in the corridors and stairwells; and pressure sensors (PX409-015G5V) are installed in the pipelines.

[0058] After the leak was triggered, the combustible gas sensor detected an ethanol concentration of 800 ppm (the lower explosive limit of 3.3% corresponds to 1650 ppm), the infrared thermal imager showed a leak point temperature of 22℃ (no abnormal temperature rise), and the pressure sensor showed that the pipeline pressure dropped from 0.3 MPa to 0.05 MPa; the array microphone captured the sound of a pipe rupture (frequency 800 Hz); and the binocular camera showed a personnel density of 1.2 people / m² in the laboratory and corridor. 2 .

[0059] The data is denoised by variational mode decomposition (5 layers, penalty factor 250), and after Mel coefficient extraction, the Transformer encoder compresses it into 512-dimensional features, which are stored on the edge node (Huawei Atlas800, computing power 64 TOPS) with a response latency of 78ms.

[0060] 2. Domain Augmentation Model Construction and Scene Recognition The 20B parameter model is finely tuned based on 20,000 cases of hazardous chemical leaks in universities: LoRA rank value 32, learning rate 2e-5, learning knowledge such as "ethanol leak handling" and "laboratory ventilation control"; 3,000 specialized instructions include professional content such as "chemical laboratory graded protection instructions", and 450 new terms have been added, such as "hazardous chemical emergency spraying" and "explosion-proof tool use".

[0061] After inputting the Prompt, the model outputs the following result in 1.4 seconds: "Ethanol leak in a 5-story organic laboratory, 10L leak, 800ppm concentration, involving 25 laboratory personnel", with a confidence level of 96%, no need to initiate knowledge distillation.

[0062] 3. Risk Field Modeling and Hierarchical Evolution Hazardous chemical spill risk framework: Laboratory personnel exposure factor 1.2, corridor 0.8; ethanol evaporation rate 0.5 m / min, environmental vulnerability index 0.7; emergency station distance on campus 300 m (access time 3 min), resource accessibility 0.6. Entropy weight-TOPSIS calculates initial risk value of 78 points.

[0063] The risk evolution equation introduces a volatilization rate of 0.5 m / min, updates the risk field every 2 seconds, predicts that the concentration will reach 1200 ppm after 30 seconds, spread to a range of 10 m in the corridor, with a risk level of 3, and may reach the lower explosive limit after 5 minutes of warning.

[0064] 4. Multi-objective decision generation Model retrieval 5000+ cases and "University Laboratory Safety Emergency Guide", generate 3 sets of initial scheme. Multi-objective optimization (evacuation time 0.3, casualty probability 0.4, resource consumption 0.3) optimal scheme: immediately start the laboratory ventilation system (exhaust rate 10 m / s), personnel wear protective masks along the west stairs, the school emergency station carries anti-explosion tools, absorption cotton disposal, and calls East Campus AED and first aid personnel standby.

[0065] Feasibility check: Stair capacity 6 people / s, 25 people can be evacuated in 5s, emergency resources arrive within 3min, scheme generation time 4.2s.

[0066] 5. Decision execution and feedback Instruction analysis: Laboratory broadcast outputs professional voice ("Immediately turn off the power, wear yellow protective mask, evacuate from the west side"); electronic screen displays evacuation route + protection diagram; teacher's bracelet pushes "count the experimental personnel" instruction; emergency light green flashes 1Hz.

[0067] Millimeter wave radar monitors that the west stair entrance causes the flow rate to drop to 0.8 m / s due to carrying equipment, density 1.8 people / m 2 , model generates adjustment instruction: "open the east standby stair, experimental equipment is placed in the safety area first", response time 0.9s, speed increases to 1.1m / s.

[0068] 6. Three-dimensional situation visualization Digital twin model integrates laboratory equipment location, gas concentration distribution (blue gradient display), and cross-campus resource coordinates. The command interface displays the evacuation completion rate (100% in 10s), concentration drop rate (200ppm per minute), and supports viewing the east campus first aid personnel's travel route.

[0069] 7. Review and model optimization Data chain contains gas concentration curve, disposal video, etc. Knowledge graph timeline shows that the actual concentration dropped to a safe value in 8min, 2min longer than predicted due to clogged ventilation system filter. This case is classified as "equipment failure affecting disposal", and the model is incrementally trained to optimize the instruction: "check the ventilation equipment status before leak disposal".

[0070] 8. Execution of other modules Multi-campus resource scheduling: Improved ant colony algorithm (pheromone evaporation 0.1, heuristic factor 2.0) is used to schedule East Campus resources, time decay factor half-life 30s, resource response time is shortened.

[0071] Federal learning: 3 campus local training model, encrypted transmission parameters (2048 bit homomorphic encryption), model recognition accuracy is improved after 15 rounds of aggregation.

[0072] Adversarial training: generate "ethanol leakage + power off" virtual samples, and the model's extreme scenario decision success rate is improved.

[0073] Running effect data representation Table 1: Comparison of primary school fire emergency disposal effect Indicators Traditional emergency solution Invention method Scenario recognition time 4.5s 1.2s Decision generation time 9.2s 3.8s Evacuation completion time 75s 42s Student instruction compliance rate 65% 92% Secondary risk occurrence rate 22% 3% Table 1 data comes from a 3-story fire simulation test in a primary school. The traditional scheme relies on manual reporting and identification, which takes 4.5s, and decision generation requires reviewing the plan, which takes up to 9.2s. The instructions use adult language, and the compliance rate of low-grade students is only 65%. The secondary risk rate due to route congestion during evacuation is 22%, and the evacuation time is 75s. The invention realizes second-level identification through multi-modal fusion, generates child-adapted instructions through specialized models, and the compliance rate is increased to 92%. Real-time feedback adjusts the route to avoid congestion, reducing the secondary risk to 3%, and shortening the evacuation time by 44%. The data verifies the practical value of cross-modal perception, low-age adaptive decision, and other modules, which adapt to the needs of primary school emergency scenarios.

[0074] Table 2: Comparison of college hazardous chemical leakage disposal effect Indicators Traditional emergency solution Invention method Leakage concentration control time 15 min 8 min Cross-campus resource response time 12 min 5 min Professional instruction accuracy rate 70% 98% Privacy data leakage risk High None Extreme scenario coping ability Weak Strong Table 2 data comes from a leakage test in a college chemical building. The traditional scheme lacks real-time concentration monitoring, and the control time is as long as 15 minutes. Cross-campus resource scheduling relies on telephone communication, and the response time is 12 minutes. The instruction has many professional term errors, with an accuracy rate of 70%, and there is a privacy risk in data sharing. It has no ability to cope with extreme scenarios such as "leakage + power off". The invention realizes real-time monitoring + ventilation control through gas sensors, reducing to a safe value in 8 minutes. Improved ant colony algorithm + federated learning realizes fast scheduling across campuses, responding in 5 minutes. The field fine-tuning model ensures an instruction accuracy rate of 98%, and the differential privacy technology eliminates the risk of leakage, and the adversarial training strengthens the ability to cope with extreme scenarios. The data verifies the effectiveness of resource scheduling, privacy protection, and other modules, which adapt to complex emergency scenarios in colleges.

[0075] Reference Figure 2 This figure directly shows the core value of the cross-modal attention fusion and the field enhanced model in claim 2. The traditional method relies on single modal data and general models, and has weak recognition ability for complex scenarios. The accuracy rate of superimposed scenarios is only 65%, far below the practical threshold. The invention dynamically weights and fuses multi-modal features such as temperature, sound, and image, and combines a large language model fine-tuned with campus domain corpus, with an accuracy rate of more than 90% for various scenarios, and 92% for superimposed scenarios. This verifies the ability of the cross-modal fusion mechanism to capture related features and the adaptability of the field fine-tuning to professional scenarios, solving the "recognition error" pain point of emergency decision-making from the source.

[0076] Reference Figure 3The figure clearly reflects the forward-looking value of the risk evolution model in claim 3. The traditional fixed plan uses the initial risk value for static decision-making, and always executes the 80-point high-risk instruction within 40s, lacking dynamic adaptation. The application updates the risk value in real time through the risk evolution differential equation predicts the trend after 30s, and adjusts the route based on the prediction at 10s, with a significantly faster risk value decline rate than the traditional scheme. This verifies the ability of the risk dynamic modeling to capture the spread of events, enabling the decision-making to shift from "passive response" to "active prediction" and avoiding potential risks in advance.

[0077] Referring to Figure 4 The figure highlights the practical effect of the age adaptation mechanism in claim 4. The traditional scheme uses unified instructions and routes, and the preschool group takes as long as 90s to evacuate due to complex instructions and poor speed adaptation, with a compliance rate of only 60%; the middle school group's efficiency is not optimal due to overly basic instructions. The application designs adaptive schemes for different age groups: it strengthens visual instructions and simplifies steps for the preschool group, optimizes speech clarity for the elementary school group, and reserves autonomous decision-making space for the middle school group, significantly shortening the evacuation time for each age group and breaking through 90% compliance rate. This solves the adaptability problem of traditional "one-size-fits-all" decision-making and improves the emergency execution effect of different groups.

[0078] Referring to Figure 5 The figure verifies the advantages of distributed resource collaborative scheduling in claim 5. The traditional cross-campus scheduling relies on manual communication and paper records, and the resource positioning and matching takes a long time, with medical personnel responding in 12 minutes, far exceeding the golden threshold. The application optimizes the scheduling path by improving the ant colony algorithm and implements encrypted sharing of multi-campus resource information through federated learning, with the response time of all types of resources compressed to within 8 minutes, and the response time of medical personnel only 5 minutes. This reflects the optimization capability of intelligent scheduling algorithms for paths and the balance of federated learning for data privacy and sharing, solving the "resource lag" problem of cross-campus emergencies.

[0079] Referring to Figure 6 The figure intuitively presents the value of closed-loop review and model evolution in claim 10. The initial model relies on basic training data, with a success rate of only 70% in extreme scenarios. With the accumulation of review cases, decision-making flaws are mined through comparative learning, and model parameters are optimized through incremental training. After adding adversarial virtual samples at 200 cases, all indicators are accelerated, and the success rate of extreme scenarios reaches 93% at 400 cases. This verifies the effectiveness of the "data-decision-review-optimization" closed-loop mechanism, enabling the model to continuously absorb emergency experience and continuously enhance its ability to cope with complex scenarios, achieving "self-evolution" of the emergency decision-making system.

[0080] The above merely describes preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes within the technical scope disclosed by the present application and according to the technical solutions and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for generating emergency decisions based on a large language model in an educational setting, characterized in that, Includes the following steps: Multimodal emergency information is perceived in real time and aligned in time and space. Heterogeneous devices are deployed in key areas of the campus to collect multi-source data, and spatial alignment is achieved through coordinate system transformation matrix. After signal interference removal and feature extraction, the data is compressed into a unified feature vector by Transformer encoder. Domain-enhanced large language model construction and fine-tuning, building a campus emergency case annotation corpus, and performing two-stage fine-tuning of the 20B parameter base model: LoRA technology freezes 90% of the parameters to learn domain knowledge; Design a scene recognition prompt template, input fused features and output classification results; for fuzzy scenes, initiate a knowledge distillation mechanism. Dynamic risk field modeling and level evolution: Based on the identification results, an event-specific risk assessment framework is invoked, and the entropy weight-TOPSIS algorithm is used to calculate the initial risk value; a risk evolution differential equation is constructed. Multi-objective emergency decision generation intelligently generates a decision knowledge base using a large language model, and uses a case and rule reasoning fusion strategy to generate an initial set of solutions; it constructs an optimization function with evacuation time, casualty probability, and resource consumption as objectives, and solves the optimal solution using an improved NSGA-III algorithm; Distributed decision-making command execution and feedback resolves the optimal solution into multi-terminal commands; millimeter-wave radar monitors pedestrian flow in real time, triggering adjustment mechanisms when encountering congestion or deviation; and a large language model generates local optimization commands. Three-dimensional situation visualization and interactive command: Construct a digital twin model of the campus, generate situation maps through edge and cloud collaborative rendering, and design an interactive interface to support touch viewing of details; The system performs full-cycle data review and model evolution, automatically records the entire emergency response data chain, and uses knowledge graphs to construct event timelines. It also mines decision failure modes through comparative learning, builds an error case library, and optimizes the model through incremental training and reinforcement learning.

2. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes a cross-modal attention fusion mechanism, which constructs a three-modal interactive attention matrix of temperature, sound, and image. The self-attention module mines the spatiotemporal correlation within the modality, and the cross-attention module calculates the intermodal dependency coefficient. The feature weights of each modality are dynamically adjusted. The weights of temperature features increase non-linearly with the degree of deviation from normal temperature. The weights of sound features have a quadratic function relationship with the confidence of abnormal sound. The weights of image features are adjusted in segments according to the gradient of personnel density. The fused feature vector is then dimensionality reduced by principal component analysis.

3. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes crowd flow dynamics prediction, and the construction of crowd evacuation models based on continuum mechanics. Where ρ(x,y,t) is the population density at coordinate (x,y) at time t, v(x,y,t) is the population flow velocity at that point, and ∇・ is the divergence operator. This equation describes the population conservation relationship. The acceleration field is calculated by combining the social force model, and the population distribution in the next 30 seconds is solved by the finite difference method. The large language model optimizes the evacuation route based on the prediction results to avoid potential congestion areas.

4. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes an adaptive decision-making mechanism, which builds a decision-making suitability assessment system to address the differences in cognitive and behavioral abilities among different age groups. In the early childhood scenario, it focuses on the proportion of visual instructions and the simplification of steps; in the primary school scenario, it increases the clarity of voice instructions and the ratio of guides; in the middle school scenario, it strengthens the space for autonomous decision-making; and in the university scenario, it focuses on the precision of professional instructions. It also designs a decision complexity adjustment algorithm to dynamically adjust the instruction density based on the average movement speed of the crowd monitored in real time.

5. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes distributed emergency resource collaborative scheduling, building a dynamic matching network of resources and needs, abstracting fire equipment, medical resources, and manpower support into resource nodes, and quantifying attribute values ​​according to capability dimensions; using an improved ant colony algorithm to solve the optimal scheduling path, while introducing a time decay factor, and activating a priority sorting mechanism when the total amount of resources is insufficient; and realizing multi-campus resource sharing through federated learning.

6. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes multimodal and multilingual command adaptation, building a full-modal command generation system for foreign teachers and students and special groups on campus. Text commands support real-time conversion of Chinese, English, Japanese and Korean, voice commands use emotional speech synthesis, visual commands include dynamic icons and sign language animations, and tactile commands output vibration codes through smart bracelets. Define an instruction validity evaluation function. Where E represents overall effectiveness. For modal weights, This represents the recognition error rate for this mode.

7. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes adversarial scenario enhancement training. For extreme events with insufficient training data coverage, a virtual scenario library is constructed using generative adversarial networks. The generator generates extreme samples based on the feature distribution of real events, and the discriminator distinguishes between real and virtual samples. The generated samples are mixed into the training set at a ratio of 1:3 to perform adversarial fine-tuning on the large language model.

8. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes an edge-cloud collaborative reasoning mechanism, which breaks down tasks such as scene recognition, risk assessment, and decision generation according to computational complexity. Edge nodes handle real-time tasks, while cloud servers handle complex tasks. Model compression technology is used to reduce the size of large language models to 30% of their original size, and 5G slicing technology is used to ensure data transmission. A dynamic task scheduling strategy is designed to automatically push some cloud tasks to the edge when network latency is greater than 100ms.

9. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes a digital twin assessment of emergency response capabilities, constructing an assessment indicator system for campus emergency response capabilities, setting 12 specific indicators from three dimensions: perception, decision-making, and execution; simulating 20 typical scenarios based on the digital twin model, collecting indicator data to calculate a comprehensive score, and using causal inference algorithms to locate weak links.

10. The emergency decision generation method based on a large language model in an educational setting according to claim 1, characterized in that, It also includes privacy-preserving data review, which uses a federated learning framework to achieve collaborative data review across multiple campuses. Each campus completes data cleaning and feature extraction locally, exchanges model parameters through homomorphic encryption technology, and the aggregation server uses secure multi-party computation to fuse parameters. Differential privacy protection is implemented for data involving personal information, and a federated knowledge graph is used to build a cross-campus event association network.

Images