A preplan digitalization method and system based on large model technology

Abstract

The application relates to the field of digital emergency management, and provides a preplan digitalization method and system based on large model technology, which comprises the following steps: creating a multi-level label system of a preconfigured industry; when a user instruction contains a label feature in the multi-level label system, generating a prompt word template guided by domain knowledge corresponding to the label feature; receiving the prompt word template through a preset prompt word information large model, performing structured analysis on unstructured preplans, and generating a first digitalization instruction set; obtaining error correction backflow data of the first digitalization instruction set, updating the prompt word large model and the corresponding model interface fine tuning through the error correction backflow data, and generating a second digitalization instruction set after model structure fine tuning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of digital emergency management, in particular to a preplan digital method and system based on large model technology. BACKGROUND

[0002] In the emergency management of enterprises, digital emergency management is usually adopted. Digital emergency management is a comprehensive emergency event digital management system formed by new generation information technologies such as big data, artificial intelligence, Internet of Things, blockchain and digital twin.

[0003] It is mainly applied to local service platforms and interfaces with cloud platforms. The cloud platform interfaces with a software platform. In the traditional digital preplan processing, the software platform manually defines a tag system and a structured rule of tag recognition. Through keyword matching and template analysis, the elements are extracted to position and analyze the emergency event. However, it can only handle limited and fixed types of emergencies. However, for complex scenarios, there are multiple different events or emergencies triggered by multiple tags in user instructions, and multiple tags corresponding to emergencies exist at one time. The ability is limited, and it cannot handle complex emergency events in complex scenarios.

[0004] The software platform uses rule engines and simple machine learning models for event and instruction analysis in terms of intelligence. It cannot deeply understand emergency events in complex scenarios, can only generate fixed preplans, and may trigger incorrect preplans due to the ambiguity of user instructions during event and instruction analysis, which cannot achieve dynamic preplan automatic evolution generation.

[0005] Existing simple machine learning models mostly lack closed-loop optimization capabilities, so there is no automatic error correction mechanism, the model iteration efficiency is low, and it is easily affected by human bias. It can only deal with fixed emergency events in emergency scenarios, and there is a data island problem. The resource scheduling in emergency command is disconnected from the actual demand.

[0006] Application Content

[0007] The present application provides a preplan digital method and system based on large model technology. By constructing a multi-level tag system and a cross-modal graph neural network, the intelligent deconstruction of unstructured preplan elements and the spatio-temporal correlation modeling are realized. Through the dynamic analysis engine driven by the large model, the recognition accuracy of nested conditional statements and user instructions is improved. The instruction tree branch activation mechanism driven by real-time meteorological data is introduced to dynamically match preplan clauses and environmental parameters, and to automatically adapt emergency plans in complex scenarios.

[0008] In view of the defects of the traditional model in the optimization level, through the double-channel error correction backflow and the synergy of federated learning and adversarial sample generation, the model iteration period is compressed to the hour level, while ensuring the knowledge consistency in the updating process. In view of the emergency event, individual adjustment can be realized, and model fine-tuning can be realized for different user instructions, so that the plan is more adaptive to the overall complex environment in complex environment.

[0009] In the first aspect, the application provides a plan digitization method based on large model technology, specifically including: creating a multi-level label system of pre-configured industries;

[0010] When the user instruction has a label feature in the multi-level label system, a prompt word template guided by the label feature corresponding domain knowledge is generated;

[0011] The prompt word information large model receives the prompt word template, performs structured analysis of unstructured plans, and generates a first digital instruction set;

[0012] Obtain the error correction backflow data of the first digital instruction set, and update the prompt word information large model and the corresponding model interface fine-tuning through the error correction backflow data, to generate a second digital instruction set after model structure fine-tuning.

[0013] The plan digitization method provided by the application is applied to the emergency event alarm process in the cloud, and the event label of the emergency event existing in the user instruction is analyzed through the received instruction, so that an automatic emergency event guiding mechanism can be realized, avoiding the need for understanding the context meaning and manual intervention in the traditional emergency event processing process. The existing label system cannot dynamically adapt to different scene plans, but only adapts to fixed scene plans. Then a prompt word template guided by the knowledge corresponding to the label feature is generated, and the prompt word template is associated with the label feature. The difference between the prompt word template and the traditional prompt word template is that the application can solve the problem of fixed prompt word template in the prior art, which also corresponds to fixed emergency events. The prompt word template of the application corresponds to the label feature and guides accordingly. After the prompt word information large model receives the prompt word template, the emergency event scene simulation is performed, the structured analysis of unstructured plans is performed, that is, the analysis of non-fixed emergency event plans is performed, and the digital instruction existing in the plan is determined, that is, the digital task to be executed. The error correction backflow data is an iterative error correction analysis mechanism realized by the prompt word information large model in the uninterrupted analysis process of the prompt word template, which can judge the task instructions that cannot be implemented in the first digital instruction set and the missing digital instructions, realize the continuous updating of the prompt word information large model, and generate a second digital instruction set, which is a new processing plan implementation instruction corresponding to the second digital instruction set after the processing plan of the emergency event is corrected.

[0014] In combination with the first aspect, the multi-level label system construction comprises:

[0015] According to the pre-configured industry, an industry classification code is determined;

[0016] The industry classification code is converted into a time sequence feature vector, and an industry-event spatio-temporal correlation matrix is constructed based on real-time meteorological data and holiday information;

[0017] A cross-modal graph including an event node, an environmental state node and a social impact node is established, wherein the cross-modal graph simulates an event evolution path based on a gated graph convolution network, and calculates the effect probability of different disposal measures through counterfactual intervention;

[0018] The industry-event spatio-temporal correlation matrix and the cross-modal graph are combined with an event feature embedding layer and a graph attention mechanism to output a multi-level label system with confidence weight, and each label is marked by SHAP value.

[0019] In this embodiment, the application determines the industry classification code in order to adjust the weight of the label according to meteorological data, holiday data and other data, which can realize spatio-temporal adjustment, i.e. the classification code corresponding to the label of the emergency event of different industries is different in different meteorological data, holiday data, etc. For example, tourist congestion caused by snowstorm is different in holiday and non-holiday, and the emergency degree is changing all the time. The application combines the industry-event spatio-temporal correlation matrix and the cross-modal graph with the event feature embedding layer and the graph attention mechanism, which can explain different labels by integral gradient, and the corresponding explanation information of the label is different in different situations of the current emergency event.

[0020] In combination with the first aspect, the cross-modal graph is further used for:

[0021] Based on the preset real-time emergency resource inventory data, the connection strength of the event node and the environmental state node is adjusted, and the semantic alignment degree of the event feature and the meteorological data is simultaneously optimized;

[0022] When the SHAP value detects that there is a label anomaly, a double-channel optimization of federated learning data completion and adversarial sample generation is triggered, wherein the adversarial sample is preferentially synthesized in a low confidence event scene in a historical case.

[0023] In this embodiment, the cross-modal graph can align the semantic of the event content of the emergency event and the environmental condition of the region where the emergency event occurs according to the real-time emergency event, so as to prevent the problems of scheme lag and unreasonable scheme of the traditional fixed pre-plan scheme. Through the double-channel optimization of federated learning data completion and adversarial sample generation, whether the label feature of the emergency event has a low confidence label can be found, and the data that is unreasonable or does not conform to the event status of the emergency event can be prevented from appearing in the event factors of the emergency event, so as to prevent the pre-plan of the emergency event from deviating and causing resource waste.

[0024] In combination with the first aspect, the prompt word template is further used for:

[0025] Real-time monitoring of different label trigger frequencies and error correction data in user historical operations;

[0026] Dynamic adjustment of different label weight distribution ratios according to industry risk levels;

[0027] Generating a hierarchical guided prompt word template containing priority identification, wherein high-risk labels are automatically displayed in the front.

[0028] In the embodiment of the application, the prompt word template is used to provide the prompt word information large model with deduction prompt words and determine the deduction target. Since the prompt word template is composed of label features, the label weight distribution can be performed based on the current industry corresponding to the emergency event by judging the error-triggered label, so as to increase the weight of necessary and key labels, and reduce the weight of labels that may have errors or low trigger frequency and low relevance, and then in the guidance, through the priority identification, i.e. under the condition of different weights, different labels are classified to realize hierarchical guidance and improve the processing accuracy of the emergency event in the large model.

[0029] In combination with the first aspect, the prompt word template is further used for:

[0030] Analyzing the abnormal event chain in the historical operation record, extracting semantic contradiction points, data missing segments and rule conflict domains, and constructing a multi-dimensional error feature vector library;

[0031] Based on the error feature vector library, enhanced training samples with composite interference characteristics are generated through semantic replacement, time sequence disturbance and context offset algorithms, wherein the sample complexity of the enhanced training samples is positively correlated with the error frequency;

[0032] The error type severity of the enhanced training sample is set as a sample training priority, and multi-round adversarial training is performed on high-frequency low-confidence events until the model output stability exceeds a preset threshold.

[0033] This application uses the chain of abnormal events in historical operation records to determine a multi-dimensional error feature vector library composed of multiple interconnected feature combinations, thereby achieving training enhancement through error feature-driven composite perturbation. Based on multi-round adversarial training, a stable model is obtained. When generating prompt word templates, the stable model will improve accuracy while reducing the correlation between tags.

[0034] In conjunction with the first aspect, the prompt word template is also used for:

[0035] Integrate a digital twin simulation interface into the prompt word template to generate a set of prompt words with simulation sandbox identifiers;

[0036] Based on the indicators in the simulation sand table, a virtual scenario simulation space for emergency events is generated.

[0037] This application generates a digital twin sand table capable of simulation and deduction based on prompt word templates, realizes closed-loop simulation, dynamically processes the label information of emergency events, generates prompt words, can verify the relevance of prompt words while excluding irrelevant labels and prompt words, and can also refine the prompt words.

[0038] In conjunction with the first aspect, the structured parsing includes:

[0039] Analyze textual plans, on-site images, and sensor data streams to generate emergency response elements with three-dimensional spatial coordinate annotations;

[0040] Based on emergency response elements, nested conditional statements are identified, and a hierarchical instruction tree with failure branch warnings is created; whereby failure branch warnings are used to characterize invalid instructions corresponding to false emergency response element annotations in nested conditional statements.

[0041] By reverse-engineering the hierarchical instruction tree from the historical case library, a multi-department collaborative task chain with weighted identifiers is generated.

[0042] During the structured analysis process, this application generates a three-dimensional space of the emergency event scene based on the prompt word template. Emergency response elements are labeled within this three-dimensional space. By identifying these elements, it determines whether nested conditional statements exist, i.e., whether there are related emergency event descriptions. This process then generates a hierarchical instruction book for handling the emergency event, complete with failure branch warnings. These warnings indicate events that, based on the prompt word template analysis and related emergency event descriptions, should exist during emergency response but may not, or may have already been handled. The hierarchical instruction book is then reverse-engineered using a historical case database to generate various emergency response tasks, forming a multi-departmental collaborative task chain.

[0043] In conjunction with the first aspect, the error correction backflow data update includes:

[0044] The three types of error correction sources, user feedback data, system execution logs and expert review marks, are collected, abnormal data is cleaned through a conflict detection algorithm, and a training sample set with confidence weight is generated;

[0045] According to the training sample set, the prompt word information large model is updated in layers, and the full parameter fine-tuning is used for layering;

[0046] According to the fine-tuned prompt word information large model, a second instruction set is generated, and when the verification accuracy of the second instruction set is lower than the preset threshold, the system automatically restores to the previous stable version and triggers the manual review process.

[0047] In the implementation of the present application, through different types of error correction sources, it is judged whether there is abnormal data through a conflict detection algorithm, a new training sample set is formed, and then the new training sample set is updated in layers in the prompt word information large model. The result of the layer data update will have the prompt word information large model updated, and a new instruction set will be generated. When the accuracy of the new instruction set is lower than the preset threshold, the prompt word information large model will restore to the state before updating.

[0048] In combination with the first aspect, the model interface fine-tuning further includes:

[0049] The edge device deploys a lightweight interface verification model to filter low-quality error correction data in real time and upload high-value samples to the prompt word information large model for model updating;

[0050] Key features are extracted from the updated prompt word information large model to generate a simplified interface adapter for low-power terminal calling;

[0051] A unified interface protocol conversion framework is constructed to convert the second digital instruction set into an input format compatible with the historical version model.

[0052] In the implementation of the present application, in the process of fine-tuning the model interface, a lightweight interface deployed by the edge device is also used to realize edge verification, and then the model is fine-tuned on the edge. After edge fine-tuning, lightweight adaptation is performed, so that in the system of the present application, in the case of no network connection and insufficient computing power of the edge device, the lightweight adaptation distillation method can also be applied to the model of the present application to realize the issuance of instructions for emergency events.

[0053] In a second aspect, a pre-plan digital system based on large model technology includes:

[0054] The label configuration module is used to create a multi-level label system for pre-configured industries;

[0055] The template setting module generates a prompt word template guided by domain knowledge when the user instruction has a label feature in the multi-level label system.

[0056] The first digitalization module is configured to receive a prompt word template through a preset prompt word information large model, perform structured analysis of an unstructured plan, and generate a first digitalization instruction set.

[0057] The second digitalization module is configured to obtain error correction backflow data of the first digitalization instruction set, update the prompt word large model and the corresponding model interface fine-tuning through the error correction backflow data, and generate a second digitalization instruction set after model structure fine-tuning.

[0058] The pre-plan digitalization method provided in the embodiments of the present application is applied to the emergency event alarm process in the cloud. By analyzing the event label of the emergency event existing in the received instruction, an automatic emergency event guiding mechanism can be realized, and the problems that context meaning needs to be understood and manual intervention is needed with a high probability in the traditional emergency event processing process can be avoided. The existing label system cannot dynamically adapt to different scene plans, and can only adapt to fixed and set scene plans. Then, a prompt word template corresponding to the knowledge guidance of the label feature is generated. The prompt word template is associated with the label feature. The difference between the prompt word template and the traditional prompt word template is that the present application can solve the problem that the prompt word template in the prior art is fixed and corresponds to a fixed emergency event. The prompt word template of the present application corresponds to the label feature for guidance. The prompt word template can be used to simulate the emergency event scene after the prompt word information large model receives the prompt word template, perform structured analysis of the unstructured plan, that is, perform analysis of the non-fixed emergency event plan, determine the digitalization instructions that need to exist in the plan, that is, the digitalization tasks that need to be executed. The error correction backflow data is an iterative error correction analysis mechanism realized by the prompt word information large model in the process of uninterrupted analysis of the prompt word template. The error correction backflow data can determine the task instructions that cannot be implemented in the first digitalization instruction set and the missing digitalization instructions, constantly update the prompt word information large model, and generate a second digitalization instruction set, which is a new processing plan implementation instruction corresponding to the second digitalization instruction of the emergency event processing plan.

[0059] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0060] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0061] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation on the present application. In the drawings:

[0062] Figure 1 A method flowchart of a preplan digitalization method based on large model technology in an embodiment of the present application;

[0063] Figure 2 An implementation hardware diagram of a preplan digitalization method based on large model technology in an embodiment of the present application;

[0064] Figure 3 A flowchart of a multi-level label system construction in an embodiment of the present application;

[0065] Figure 4 A cross-modal graph dynamic optimization mechanism schematic diagram in an embodiment of the present application;

[0066] Figure 5 A process schematic diagram of a risk self-adaptive prompt word template in an embodiment of the present application;

[0067] Figure 6 A process schematic diagram of an error-driven adversarial training in an embodiment of the present application;

[0068] Figure 7 An implementation process schematic diagram of a digital twin deduction integrated interface in an embodiment of the present application;

[0069] Figure 8 An implementation process schematic diagram of a multi-source structured analysis engine in an embodiment of the present application;

[0070] Figure 9 An implementation process schematic diagram of a hierarchical model update closed loop in an embodiment of the present application;

[0071] Figure 10 An implementation process schematic diagram of a cloud-edge-end collaborative fine-tuning in an embodiment of the present application;

[0072] Figure 11 A system composition diagram of a preplan digitalization system based on large model technology in an embodiment of the present application. DETAILED DESCRIPTION

[0073] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to explain and illustrate the present application, and are not used to limit the present application.

[0074] In order to solve the technical defects of emergency events in the process of digital preplan generation in the background art, an embodiment of the present application provides a preplan digitalization method based on large model technology.

[0075] Embodiment 1:

[0076] As shown in FIG. 1, the preplan digitalization method based on large model technology in an embodiment of the present application comprises the following steps: Figure 1As shown, the present application proposes a pre-plan digitalization method based on large model technology, which realizes the conversion and continuous optimization of unstructured pre-plan to structured instructions through the combination of multi-level label system and prompt word guidance.

[0077] The method of the present application specifically includes:

[0078] As Figure 1 As shown, the software platform will first trigger the pre-plan label based on the emergency management event, such as step S100: creating a multi-level label system for pre-configured industries;

[0079] First, create a multi-level label system for pre-configured industries, the core of which is to build a multi-dimensional classification framework of industry knowledge.

[0080] Through the hierarchical structure of pre-set industry characteristics, complex pre-plan content is decomposed into manageable knowledge units.

[0081] For example, in the field of emergency event handling: divide into three levels of labels:

[0082] The first level of labels can distinguish between natural disasters, accident disasters, etc.;

[0083] The second level of labels is refined to earthquake, fire, etc. scenarios;

[0084] The third level of labels may involve evacuation routes, material allocation, etc. specific elements;

[0085] It can also be divided into different levels of labels according to the importance of emergency event handling, user instructions; it can also be based on the same emergency event scenario elements; for example, under the fire scenario, fire area and ignition material, and the number of trapped people as the first level of labels, fire-fighting materials in the fire area and fire-fighting water connection area as the second level of labels, flammable and non-flammable areas in the fire scene as the third level of labels, etc.

[0086] Combined with the multi-level label system, through the form of intelligent question and answer, the user's input information can be improved, and according to the user's direct instruction, a fire occurs in a certain place, the multi-level label system automatically extracts the information of the corresponding area, generates label features automatically, so that the large model can quickly locate the handling measures of the emergency event, and avoid the handling deviation caused by information overload or insufficient information.

[0087] Step S101: When the user instruction contains label features in the multi-level label system, generate a prompt word template guided by the label feature corresponding domain knowledge;

[0088] In this application, when receiving a user instruction, the label feature is determined through the analysis of the user instruction, and the label feature is matched based on the multi-level label system. At this time, the industry corresponding to the user instruction is the preconfigured industry of the multi-level label system. Through the annotation and filling extension of the label feature by the prompt word template, the prompt word template based on the domain knowledge guidance is constituted. The domain knowledge guidance is the knowledge guidance related to the emergency event associated with the label feature in the preconfigured industry. Specifically, through semantic analysis, the user instruction is identified to handle the emergency event, and the key label based on the emergency event handling requirement is constituted. This process will combine the industry knowledge base of the preconfigured industry to constitute a structured query template.

[0089] When the input instruction contains the label feature word "chemical fire", the prompt word template will guide the domain knowledge of the chemical emergency event through the chemical fire information corresponding to the emergency event. The knowledge elements related to the chemical field such as hazardous chemical disposal procedures and fire fighting resource distribution are automatically associated to generate a prompt template containing specific parameter constraints. The specific parameter constraint is the representation parameter of the maximum impact degree and the minimum impact degree of the event element in the chemical fire scene after parameterization. It is used to represent the influence of the event element on the chemical fire. It solves the model output deviation problem caused by the ambiguous expression of the prompt word in the traditional method. By fusing the industry knowledge system, the user instruction is converted into an operable prompt engineering, and the simple user instruction is changed into a more professional and accurate emergency event handling instruction guide type prompt information.

[0090] Step S102: receiving the prompt word template through the preset prompt word information large model, performing structured analysis of the unstructured plan, and generating a first digital instruction set;

[0091] In this application, the prompt word information large model performs structured analysis of the unstructured plan. In this process, when processing the text information of the prompt word template, the prompt word information large model combines the constraint conditions of the prompt word template through semantic understanding ability and large model deduction ability, integrates the scattered plans of the emergency event, and extracts the key information of the emergency event handling related to the prompt word information in the scattered plans.

[0092] For example, the emergency process, the responsible subject, the execution standard, the emergency handling measure, the emergency dispatch information and the like are converted into standardized instruction nodes.

[0093] For example, "immediately start the chemical fire three-level response mechanism", after analysis, contains: trigger condition, response level, dispatch department, chemical area range, chemical fire emergency handling tool, emergency handling precautions, material allocation instruction and other fields of structured data. In the scattered plan, the plan handling item associated with the structured data is called;

[0094] Further, by processing complex semantic expressions, the degree of automation of the pre-plan digitization is improved, the manual annotation cost is reduced, and the integrity of the business logic is maintained.

[0095] S103: Obtain error correction backflow data of the first digitized instruction set, and update the prompt word information large model and the corresponding model interface fine-tuning through the error correction backflow data to generate a second digitized instruction set after model structure fine-tuning.

[0096] In the present application, based on the model updating and interface fine-tuning of the error correction backflow data, the error cases in actual application, such as missing instruction parameters and logical sequence errors, are used as training data to optimize the model weight parameters through incremental training, and the data verification rules of the API interface are adjusted.

[0097] For example, when the defect of "material allocation instruction not containing transportation tool type" is found, the mandatory verification of this field is added in the subsequent model reasoning. Adaptive evolution is realized to dynamically adapt to the update of industry standards and the expansion of business scenarios, preventing the knowledge solidification of traditional static models and further performance degradation.

[0098] In the present application, the multi-level label system provides a structured knowledge framework for prompt word generation, ensuring the accuracy of model input; the domain-guided prompt engineering effectively injects industry knowledge into the model reasoning process, ensuring the professionalism of the output content; the structured analysis module realizes the standardized conversion of unstructured data to form executable digital instructions; and the error correction backflow mechanism builds an iterative cycle for self-improvement of the system.

[0099] Figure 2 An implementation scenario diagram of an emergency event digitization processing method is shown in FIG. 1. Figure 2 As shown in FIG. 1, the implementation system of the pre-plan digitization method of the emergency event is deployed in a cloud center 1, an edge end 2, and a terminal access end 3. The edge end 2 is for users and service providers of the pre-plan digitization system, and is used to realize instruction distribution and data transmission when an emergency event occurs.

[0100] For example, the user instruction of the present application is received by the terminal access end 3. The user can receive the user instruction through terminal devices such as environmental sensors 31, intelligent cameras 32, wearable devices 33, tablets, computers, and mobile phones. The user instruction is transmitted to the edge computing node 21 through the industrial Ethernet or the Internet of the Internet through the field switch 22. The edge computing node 21 analyzes the instruction through the multi-level label system to determine the label features, and then generates a prompt word template under the guidance of domain knowledge through the label features.

[0101] The prompt word template is transmitted to the edge gateway in the 5G network. The edge network 14 is a gateway in the cloud center. The prompt word information large model inside the distributed computing cluster 13 performs unstructured plan analysis. The generated digital instruction set is not systematic, but only spliced scattered plans, which can solve emergency events, but there may be plans in the plan that are unrelated to emergency event elements, and there may be plans that are related to emergency event elements, but are not sufficient to solve the emergency event.

[0102] The cloud center dual-active data center 11 and the core router 12 are used to perform encryption functions and provide high-speed network SD-WAN for emergency event processing. The dual-active data center 11 also has the function of updating the prompt word information large model through error correction backflow data, thereby realizing model fine-tuning and generating a second digital instruction set.

[0103] The edge gateway 14 receives the second digital instruction set issued by the distributed computing cluster 13, controls the emergency command end 24 in the Bluetooth and wifi communication mode through the edge computing node 23, i.e. the server of the emergency processing personnel management end, based on the Mesh network, controls the emergency tablet 34 to display the emergency processing scheme to the processing personnel of the emergency event, and controls the rescue equipment 35 through Bluetooth to be used by the emergency event processing personnel.

[0104] The identity authentication server 15 takes the input analysis of the management user instruction through the Radius protocol, realizes traffic cleaning through the firewall cluster 16, lets the prompt word information large model process the scenic area, processes the update of the prompt word information large model through the hardware encryption module 17, and prevents the prompt word information large model from malfunctioning or being abnormal.

[0105] Embodiment 2:

[0106] In the process of creating a multi-level label system of pre-configured industries, i.e. building a multi-level label system corresponding to emergency events, the present application refers to Figure 3 .

[0107] Firstly, the present application determines the industry classification code of the industry corresponding to the emergency event through the emergency command end 24 of the edge end 2, establishes a benchmark coordinate system of labels associated with emergency events based on industry knowledge, and constructs a multi-level label system through the benchmark coordinate system. Different coordinate points of the benchmark coordinate system are associated with different labels corresponding to different emergency events, and the benchmark is the spatio-temporal correlation of industry-events. Specifically, the correlation between emergency events and industries, and the rules of emergency events in different meteorological data and different holiday information.

[0108] Exemplarily, the application converts the classification code into a time sequence feature vector, and incorporates real-time meteorological data and holiday information, and through construction of a time-space correlation matrix, enables the label system to capture the periodicity of industry events in the time dimension and the environmental dependence in the space dimension.

[0109] The application also constructs a cross-modal graph to determine the effect and probability of different measures in emergency events. It can be understood that the application correlates event nodes, environmental state nodes, and social influence nodes, simulates possible paths of event evolution over time using a gated graph convolution network, and evaluates the effect probability of different measures by means of counterfactual intervention technology, giving the label system the effect of predicting the evolution of emergency events.

[0110] Finally, the application fuses the time-space correlation matrix and the cross-modal graph analysis results through an event feature embedding layer and a graph attention mechanism, outputs multi-level labels with confidence weights, and marks the importance of each label of the multi-level labels through an explainability tool, forming label body information with high confidence and explainability.

[0111] According to the pre-configured industry, determine the industry classification code;

[0112] It can be understood that the application has significant differences in pre-plan structure, terminology system, and focus for emergency events in different industries. For example, the power industry focuses more on equipment operation and power grid stability, while the medical industry focuses on diagnosis and treatment processes and emergency response. The establishment of industry classification codes is like establishing a specialized language dictionary for different fields. Through the language dictionary, the industry classification code is converted into a time sequence feature vector, and based on real-time meteorological data and holiday information, an industry-event time-space correlation matrix is constructed, through dynamic variables in multiple dimensions such as time, weather, and environment.

[0113] The time sequence feature vector can capture the periodicity of industry events over time;

[0114] Exemplarily: the high incidence period of construction accidents in the construction industry may be related to the time distribution of the rainy season, and holiday information can reflect the impact of changes in social activities on industry events. For example, the surge in passenger flow during holidays in the transportation industry may trigger different emergency needs;

[0115] Through the construction of the time-space correlation matrix, the label system is no longer a static classification framework, but an intelligent model that can dynamically adjust with time and environmental changes, enhancing the adaptability of the label to actual scenarios.

[0116] The application establishes a cross-modal graph containing event nodes, environmental state nodes, and social influence nodes, wherein the cross-modal graph simulates the evolution path of events based on a gated graph convolution network, and calculates the effect probability of different measures through counterfactual intervention.

[0117] In this process, the cross-modal graph fuses multi-dimensional information such as the event itself, the scene environment corresponding to the emergency event, and the social influence, to form a stereoscopic knowledge network;

[0118] Exemplary: In the preplan of the chemical industry, the event node may be a pipeline leakage, and the environmental state node includes: temperature, wind speed, etc. The social influence node involves: the safety of surrounding residential areas, etc.

[0119] Then, through the gated graph convolution network, the dynamic interaction relationship between nodes is captured, and all possible scenarios from the occurrence to the development of the event are simulated.

[0120] The counterfactual intervention technology of the application is through the "hypothesis-deduction" way, and the consequences caused by different disposal measures in the event evolution process, i.e. disposal results, are sequentially judged.

[0121] Exemplary: The preplan is to close the valve in advance. The counterfactual intervention technology can judge whether it can reduce the degree of accident harm, so that the label system can not only describe the characteristics of the event.

[0122] The application combines the industry-event spatio-temporal correlation matrix and the cross-modal graph with the event feature embedding layer and the graph attention mechanism, outputs a multi-level label system with confidence weight, and marks each label through SHAP value.

[0123] The event feature embedding layer converts complex industry data into numerical vectors that can be processed by the prompt word information large model. Through the graph attention mechanism, the key nodes in the cross-modal graph are automatically focused, for example, the environmental factors that have a greater impact on the evolution of the event are preferentially processed.

[0124] The role of the confidence weight is to quantify the labels, and the quantified parameters are used to represent the reliability index. The SHAP value marking can be used for model interpretation, and is used to represent the contribution degree of each label in the decision.

[0125] Exemplary: In the medical emergency preplan, the golden emergency disposal time label is clearly displayed through the SHAP value.

[0126] Embodiment 3

[0127] The cross-modal graph of the application can also prevent the problems of scheme lag and unreasonable scheme of the traditional fixed preplan scheme, find out whether there is a label with low confidence in the label features of the emergency event, prevent unreasonable or data that does not conform to the event status of the emergency event from appearing in the event factors of the emergency event, cause deviation of the preplan of the emergency event, and cause the problem of waste of resources, see Figure 4 :

[0128] The application adjusts the connection strength of the event node and the environmental state node based on the preset real-time emergency resource inventory data, and synchronously optimizes the semantic alignment degree of the event characteristics and the meteorological data.

[0129] In the application, the real-time emergency resource inventory data is determined by the emergency command terminal 24 through the edge computing node 23 to determine the material storage condition of the emergency event processing; the connection strength is used to represent that the emergency event processing difficulty will increase or decrease due to the environmental condition of the emergency event processing area according to the instruction information corresponding to the emergency event, and the event characteristics and meteorological data disclosed in the user instruction corresponding to the emergency event are unified in the description of the same condition, so that the emergency resources can be distributed more quickly, and the traditional emergency event processing measures are relative.

[0130] Exemplary: in the flood disaster plan, if it is monitored in real time that the sandbag inventory of a certain area is insufficient, the prompt information of the event node is: dike reinforcement; the emergency prompt information of the environmental state node is: continuous rainfall; the connection strength of the two will be automatically enhanced to prevent the influence of the environment on the event evolution from being intensified due to resource shortage;

[0131] Synchronously optimizing the semantic alignment degree is to adjust the data mapping relationship, so that the event description, rainstorm warning and meteorological data such as rainfall and wind speed can be accurately matched in the semantic level, and the understanding deviation caused by the difference in terminology is avoided.

[0132] Exemplary: the extreme weather label in the industry plan is matched with the specific threshold in the meteorological data, and the rainfall exceeding a certain standard establishes a clear corresponding relationship, so as to prevent understanding errors and low rainfall in understanding.

[0133] When the SHAP value detects that there is a label anomaly, the application triggers the double-channel optimization of federated learning data completion and adversarial sample generation; wherein, the adversarial sample is preferentially synthesized in the low confidence event scene of the historical case.

[0134] The SHAP value of the application is an explainability tool, which can locate the labels with low confidence or abnormal contribution in the label system;

[0135] Exemplary: in the chemical leakage plan, the label is: surrounding personnel evacuation route, and the SHAP value of the label suddenly decreases; it indicates that the weight of the label in the current model is too low, and the optimization mechanism is triggered;

[0136] The federated learning data completion supplements the missing event scene data by securely aggregating multi-party data, such as emergency cases in different regions, without leaking privacy, to solve the problem of data sparseness;

[0137] The generation of adversarial samples aims to generate difficult samples close to the real distribution through algorithms for historical low-confidence scenarios and composite disaster scenarios, so that the model has feature differences in easily confused scenarios.

[0138] For example, generate adversarial samples of high-temperature drought and equipment failure occurring at the same time to improve the analysis accuracy of the model under complex working conditions.

[0139] Embodiment 4:

[0140] The application prompt word template is used to display the label, see Figure 5 :

[0141] In the process of transmitting the prompt word template to the prompt word information large model by the edge computing node 21 of the application, the different label trigger frequencies and error correction data in the user's historical operations will be monitored in real time.

[0142] The trigger frequency of different labels by the user during use is used to represent the focus in a specific industry scenario.

[0143] For example, in the power operation and maintenance plan processing, the label is: equipment fault diagnosis, and its trigger frequency is significantly higher than that of other labels, indicating that the user pays more attention to the fault handling link.

[0144] The error correction data in the application is used to correct the deficiencies of the current prompt word template and give reminders.

[0145] For example, the error correction display: the label analysis error of the emergency power supply switching step generates a prompt information, and the corresponding label domain knowledge guidance is strengthened in the edge computing node 21.

[0146] Through error correction and frequency supervision, it is judged whether the prompt word template needs to be optimized, and it is learned from user behavior without relying on fixed preset rules.

[0147] The application will dynamically adjust the weight distribution ratio of different labels according to the industry risk level.

[0148] Because the distribution characteristics of different industries are also different; for example, in the chemical industry: "dangerous goods leakage"; and the first aid response time label in the medical industry, which are all high-risk scenarios, through the industry risk assessment model, the risk level of the label can be determined.

[0149] When the application enters a high-risk period in an industry, for example: the summer high temperature leads to a surge in power load, the weight of the related high-risk label will automatically increase; then, when the prompt word template analyzes the plan, it will focus on the corresponding instruction information for the key steps that cause serious consequences, to prevent errors in digital instruction output due to unimportant information.

[0150] The application generates a hierarchical guided prompt word template containing priority identification, wherein the high-risk label is automatically displayed in the front.

[0151] The application addresses the traditional prompt words that list information in a straightforward manner, and the user needs to filter the key content by themselves. The hierarchical guided template divides the labels into different levels through priority identification, and the high-risk label is displayed in a prominent way. For example, in the fire emergency plan, high-risk labels such as "personnel safety evacuation route" and "firefighting equipment location" will appear at the beginning of the prompt word, guiding the user and the large model to generate the corresponding plan first. This way improves the information processing efficiency and speeds up the time to issue the plan by about four seconds, and the accuracy meets the emergency scene: the operation logic of prioritizing high-risk tasks.

[0152] Embodiment 5:

[0153] The prompt word template of the application is also used for self-optimization, see Figure 6 :

[0154] Analyze the abnormal event chain in the historical operation record, extract semantic contradictions, data missing segments, and rule conflict domains, and construct a multi-dimensional error feature vector library;

[0155] In the application, the abnormal event chain is used to represent the logical discontinuity that occurs in the plan analysis process, such as reversed order of disposal steps, missing key parameters of information gaps, or contradictory descriptions of the same operation by different clauses that conflict with rules;

[0156] In actual operation, different abnormalities are analyzed by link tracing to determine the defects of the prompt word template in semantic understanding, data processing, and rule matching for different label features.

[0157] For example, in a chemical accident plan, the generated plan is incorrect due to the failure to identify the association between "temperature threshold" and "pressure parameter", and the semantic contradiction is extracted and recorded as a "cross-parameter logical discontinuity" feature, which is included in the vector library and optimized. At the same time of summarizing errors, model the error-prone points for targeted training.

[0158] Based on the error feature vector library, the application generates enhanced training samples with compound interference features through semantic replacement, time sequence disturbance, and context offset algorithms, wherein the sample complexity of the enhanced training samples is positively correlated with the error frequency;

[0159] The semantic replacement in the application simulates the language diversity in real scenarios by replacing synonymous terms and adjusting sentence patterns. For example, "immediately cut off the power" is replaced with "disconnect the circuit connection at the first time", to test the consistency of the model's understanding of different expressions.

[0160] Timing disturbance is used to characterize the time sequence of the disturbance event or to extend the time interval of the key step, to test the logical reasoning ability of the model under non-standard timing;

[0161] The context shift algorithm adds or reduces environmental variables, for example, in a heavy rain contingency plan, additional conditions such as "excessive wind speed" and "traffic control" are added;

[0162] Further construct a complex scene coupled by multiple factors, for the event scene corresponding to the emergency event with higher error occurrence frequency, the generated sample complexity is higher;

[0163] For example: high-frequency "data missing" errors will be configured as complex interference samples containing semantic ambiguity and timing disorder, and reliable instructions can still be output under extreme conditions of incomplete information and unclear expression.

[0164] The application sets the error type severity of the enhanced training sample as the sample training priority, and implements multiple rounds of adversarial training for high-frequency low-confidence events until the model output stability exceeds the preset threshold.

[0165] For example: the severity assessment of error types needs to correspond to industry risk levels and actual impact consequences. In a medical emergency plan, "emergency measure sequence error" is a high-severity error, and its training priority is significantly higher than that of "non-standard term expression" and other low-impact errors.

[0166] For example: high-frequency low-confidence events represent scenarios that frequently occur in historical operations but have large fluctuations in prompt word template processing results. For example: "multiple device chain failure" disposal in the power system, as it involves multi-link linkage and has fewer historical cases, the model analysis is prone to sudden confidence drop.

[0167] The application implements multiple rounds of adversarial training, i.e., repeatedly iterating between the generated sample and the corrected output at the end of the prompt word, gradually strengthening the processing ability of complex difficult examples until the output result in the same scene, such as confidence fluctuation range and analysis accuracy, does not conform to reality.

[0168] Embodiment 6:

[0169] The prompt word template of the application is also used for function expansion, see Figure 7 :

[0170] The application integrates a digital twin simulation interface in the prompt word template to generate a prompt word set with a deduction sand table identifier;

[0171] For example: the digital twin simulation interface converts each prompt word into an operable node in the virtual scene by establishing a mapping relationship between industry contingency elements and virtual model parameters.

[0172] For example, in the fire emergency plan, the prompt word "start the fire sprinkler system" is associated with the sprinkler equipment parameters and coverage area in the digital twin model, generating a prompt word with a specific deduction identifier. This identifier not only contains the text instruction, but also carries the device location in the virtual space, linkage logic, and other multi-dimensional data.

[0173] The integrated prompt word method turns a single semantic guidance tool into a digital bridge connecting real plans and virtual deductions.

[0174] According to the deduction sand table identifier, the application generates a virtual scene deduction space for emergency events.

[0175] For example, in the flood disaster deduction, after inputting the prompt word "pile up sandbags to reinforce the dam", the deduction space will dynamically simulate the impact of sandbag piling position and quantity on the flood direction according to real-time meteorological data rainfall, water level changes, and geographic information dam material, surrounding terrain. It can even present the differences in disaster development paths under different disposal schemes.

[0176] Not limited to text-based logical analysis, the virtual space "rehearses" the actual effect of disposal measures, discovers potential problems in advance, such as disposal step omission, unreasonable resource allocation, etc.

[0177]

[0178] Embodiment 7:

[0179] The structured analysis of the application refers to Figure 8 :

[0180] In this application, the text plan, on-site image and sensor data stream are analyzed to generate emergency disposal elements with three-dimensional space coordinate labels;

[0181] Traditional plan analysis relies heavily on text information, but in real emergency scenarios, accident site layout, equipment operation parameters, and environmental indicators are more easily described by sensor data.

[0182] For example, in a building fire plan, the fire source position is identified through image recognition, combined with temperature data from sensors, to generate disposal elements such as "fire hydrant deployment points" and "safe evacuation routes" with three-dimensional coordinates. The instruction set not only contains textual descriptions, but also has spatial positioning, solving the "where" execution positioning problem in traditional technology.

[0183] ​The application identifies nested conditional statements, creates a hierarchical instruction tree with failure branch warning, and dynamically activates / freezes corresponding branches according to real-time meteorological data;

[0184] The emergency plan of the application often contains nested conditions such as "if A occurs, execute B steps, if B steps fail, start C scheme", and traditional analysis is easy to miss branches due to complex logic levels.

[0185] The application constructs a hierarchical instruction tree, and each conditional branch corresponds to a specific environmental trigger mechanism, such as wind speed and rainfall in real-time meteorological data.

[0186] For example, in the flood warning plan, when the real-time rainfall exceeds the preset threshold, the "start backup drainage pump" branch is automatically activated, and if it fails due to device failure, a warning is triggered and related instructions are frozen, ensuring that the analysis result can be dynamically adjusted with environmental changes, solving the "when" and "how" conditional response problems.

[0187] The application reversely deduces the instruction responsibility subject through the historical case library to generate a multi-department collaborative task chain with weight identification;

[0188] For example, the historical case library is used to analyze the coordination efficiency and responsibility fulfillment of each department in past emergency events, such as "a chemical leakage accident caused by delayed response of a certain department led to the expansion of consequences", and reversely deduce the responsibility subject of each disposal step in the current plan, such as the environmental protection department, the fire department, or the production unit.

[0189] According to the department responsibilities and historical performance, weight identification is assigned. For example, when generating the "pollutant detection and isolation" task, the environmental protection department is clearly identified as the main responsible unit, and the production unit is identified as the assisting unit. The weight identification determines the priority of task allocation and the order of resource allocation, solving the "who" responsibility attribution problem and avoiding unclear rights and responsibilities in execution.

[0190] In the automatic analysis process, an expert collaboration annotation window is embedded to convert manual correction traces into analysis model training samples in real time.

[0191] For example:

[0192] Although automatic analysis is efficient, in extremely complex scenarios, such as new disasters and multiple accident superposition, expert experience is important. The application embeds an annotation window, and experts real-time correct the deviations in the analysis results to generate emergency responses.

[0193] For example, in the earthquake secondary disaster plan, experts manually adjust the execution order of "gas pipeline shutdown" and "power cut-off", and the correction traces are converted into training samples to optimize the model and solve the "how to improve" continuous optimization problem.

[0194] In actual implementation, the emergency response countdown device is also connected and time-sensitive instructions are prioritized. If the instructions are not executed within the time limit, the backup scheme injection interface is automatically triggered.

[0195] In emergency treatment, "golden rescue time" is crucial to the result. The countdown device monitors the task execution progress in real time, and the time-sensitive instructions such as "cardiopulmonary resuscitation operation" and "dangerous area personnel evacuation" are marked with priority. The analysis system prioritizes processing such instructions and generates an execution plan with a timestamp. If a command (such as "start emergency power supply") is not executed due to equipment failure, the system automatically triggers a backup scheme (such as switching to battery power), avoiding global failure due to single link delay and solving the "when must do" time efficiency problem.

[0196] Example 8:

[0197] The error correction backflow data updating process of the present application refers to Figure 9 :

[0198] The present application collects three types of error correction sources: user feedback data, system execution logs, and expert audit marks. Through conflict detection algorithm, abnormal data is cleaned, and a training sample set with confidence weight is generated;

[0199] For example, user feedback data can represent operational deviations in actual applications, such as manually corrected instruction errors.

[0200] Execution logs record the internal defects of the model during runtime, including parsing time-consuming exceptions and output logic breaks.

[0201] Expert audit marks incorporate authoritative judgments from domain knowledge, but professional term interpretation errors are corrected. The present application cross- validates the three types of data sources through conflict detection algorithm. For example, compare the consistency of user feedback and expert marks to eliminate abnormal data caused by errors. Finally, according to the data source, the training sample set is ensured to be comprehensive and pure through confidence weight.

[0202] The prompt word large model is updated in layers. The industry-specific layer uses full-parameter fine-tuning, and the general semantic layer uses low-value adapter lightweight update.

[0203] The prompt word large model includes: general semantic layer, used for processing basic syntax and logical relationships; industry-specific layer, used for storing domain knowledge and professional rules.

[0204] The application fine-tunes the implementation parameters of the industry-specific layer to deeply optimize and avoid problems in error correction data; for example, for the "hazardous goods disposal process" in the chemical industry, the prompt word information large model improves the understanding accuracy of specific domain terminology and logic;

[0205] The general semantic layer uses low-rank adapter technology to add small-scale trainable parameters for lightweight adjustment to prevent over-adjustment; for example, when optimizing the "first aid procedure analysis" in the medical industry, only the parameters related to the diagnosis and treatment process in the special layer are fine-tuned.

[0206] When the application is actually implemented, the API interface parameter mapping relationship is automatically reconstructed according to the structural changes of the fine-tuned model to maintain the data format compatibility of the front and back ends;

[0207] Model parameter updates may cause changes in the field definition and format specification of the output data, such as adding a "responsible subject" field or adjusting the time format. If not adapted in time, it may cause front-end display abnormalities or back-end system integration failures.

[0208] The automatic reconstruction mechanism dynamically updates the parameter mapping rules of the API interface by analyzing the field types and enumeration values of the model output metadata; it maps the "disposal priority" value output by the model to the "red / yellow / green" labels displayed on the front end;

[0209] The internal structure of the prompt word information large model of the application remains consistent and compatible in terms of data format for the interfaces provided externally. For example, after the model adds an "environmental risk level" output field, the interface automatically maps this field to the corresponding table item in the back-end database, avoiding system errors caused by missing fields.

[0210] When the verification accuracy of the second instruction set is lower than the preset threshold, automatically revert to the previous stable version and trigger the manual review process. Verification accuracy is a core indicator of model performance. When the updated model performs poorly in critical scenarios and high-risk industry contingency analysis, automatically reverting to the previous stable version can prevent actual losses caused by incorrect outputs, such as in financial risk control plans, if the accuracy of the updated model for "abnormal transaction identification" suddenly decreases, immediately revert to the previous stable version to prevent the risk of misjudgment from expanding.

[0211] The simultaneously triggered manual review process is conducted by domain experts to deeply investigate the model update process, check whether the error correction data contains outdated cases, and whether the parameter adjustment deviates from industry rules, to ensure that the root cause is accurately identified and corrected, avoiding self-reinforcing errors caused by data bias or algorithm limitations in the automated process.

[0212] Embodiment 9:

[0213] The model interface fine-tuning of the present application is in a lightweight processing process, refer to Figure 10 :

[0214] The edge device of the present application deploys a lightweight interface verification model to filter low-quality error correction data in real time, and upload high-value samples to a cloud large model;

[0215] In the edge device, field sensors and mobile terminals generate a large amount of error correction data in real time in emergency scenarios, including repeated, false or low-value information, and misjudgment records caused by temporary network fluctuations.

[0216] The lightweight verification model of the present application performs data integrity checking and outlier detection on the local data through a pre-set rule engine, and uploads samples with clear error correction value, expert annotated key step errors, and high frequency analysis deviation to the cloud.

[0217] Key features are extracted from the updated large model to generate a simplified interface adapter for low-power terminals; the parameter scale of the traditional large model is large, and it is difficult to be directly deployed on mobile phones, portable terminals and other low-power devices.

[0218] Through feature extraction technology, the most relevant key features to interface call are stripped from the cloud large model to generate a simplified adapter with small volume and high running efficiency.

[0219] For example, in a fire emergency scenario, a low-power handheld terminal can quickly analyze "fire extinguisher type selection" and "escape route planning" basic instructions by calling a simplified adapter, without loading a complete pre-plan analysis model, which not only ensures the response speed of the terminal device, but also reduces the hardware configuration requirements, and expands the equipment coverage of the technology application.

[0220] In actual implementation, a unified interface protocol conversion framework is also constructed to convert the second digital instruction set into an input format compatible with the historical version model; in actual application, users may use different versions of models or old systems at the same time, and if the newly generated second instruction set (such as adding "responsibility subject" and "time priority") cannot be recognized by the historical version, it will cause system interfacing failure. The unified protocol conversion framework establishes a mapping relationship between new and old fields, for example, it converts "multi-department collaborative task chain" in the new version into "task allocation list" that can be recognized by the old version, and performs real-time conversion on the data format at the interface layer; for example, it simplifies the complex instruction set in JSON format into a CSV table, so that the new instruction set can be seamlessly connected to the historical system.

[0221] For example, in the process of upgrading the power system, the conversion framework automatically converts the disposal elements containing three-dimensional spatial coordinates in the new version into two-dimensional position information that can be received by the old monitoring system, avoiding the repeated development cost caused by system incompatibility.

[0222] The interface fine-tuning frequency is dynamically adjusted according to the CPU / memory load of the terminal device. The hardware performance of different terminal devices differs significantly, and a high-performance server can support high-frequency model fine-tuning to quickly adapt to new scenarios, while frequent fine-tuning of a low-performance embedded device can cause running lag. The dynamic adjustment strategy adjusts the interface fine-tuning frequency in real time by monitoring indicators such as CPU usage and memory occupancy of the device: when the device load is below a threshold, for example, a high-performance server in the monitoring center, increase the fine-tuning frequency to quickly absorb the latest error correction data; when the device load is too high, for example, a battery-powered on-site monitor, reduce the fine-tuning frequency or suspend unnecessary updates to ensure stable operation of core functions. For example, in a field emergency scenario, a portable device automatically enters "energy-saving mode" when the battery level is below 20%, reducing the interface fine-tuning frequency from once every minute to once every ten minutes, while ensuring key instruction analysis and extending device battery life.

[0223] Embodiment 10:

[0224] The present application proposes a pre-plan digital system based on large model technology, referring to Figure 11 , comprising:

[0225] The label configuration module is used to create a multi-level label system for pre-configured industries; the template setting module is used to generate a prompt word template guided by domain knowledge when the user instruction contains a label feature in the multi-level label system;

[0226] The first digitalization module is used to receive the prompt word template through the preset prompt word information large model, perform structured analysis of unstructured preplans, and generate a first digitalization instruction set;

[0227] The second digitalization module is used to obtain error correction backflow data of the first digitalization instruction set, and update the prompt word large model and the corresponding model interface fine-tuning through the error correction backflow data, to generate a second digitalization instruction set after model structure fine-tuning.

[0228] Those skilled in the art will appreciate that various aspects, or implementations of various aspects, of the present application can be embodied as a system, method or computer program product. Accordingly, various aspects, or implementations of various aspects, of the present application can take the form of an entirely hardware embodiment, an entirely software embodiment.

[0229] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A method for digitizing contingency plans based on large-scale modeling technology, characterized in that, include: Create a pre-configured multi-level tagging system for the industry; When user commands contain tag features in a multi-level tag system, generate prompt word templates that guide users with domain knowledge corresponding to the tag features; The system receives prompt word templates through a pre-set prompt word information big model, performs structured parsing of unstructured plans, and generates the first digital instruction set. Obtain the error correction feedback data of the first digital instruction set, and update the prompt word information big model and the corresponding model interface fine-tuning through the error correction feedback data to generate the second digital instruction set after model structure fine-tuning; The construction of the multi-level tagging system includes: Determine the industry classification code based on the pre-configured industries; The industry classification codes are converted into time-series feature vectors, and an industry-event spatiotemporal correlation matrix is ​​constructed based on real-time meteorological data and holiday information. A cross-modal graph containing event nodes, environmental state nodes, and social impact nodes is established. The cross-modal graph simulates the event evolution path based on a gated graph convolutional network, and calculates the probability of the effects of different intervention measures through counterfactual intervention. By combining the industry-event spatiotemporal correlation matrix and cross-modal graph with the event feature embedding layer and graph attention mechanism, a multi-level label system with confidence weights is output, and each label is marked by SHAP value; The structured parsing includes: Analyze textual plans, on-site images, and sensor data streams to generate emergency response elements with three-dimensional spatial coordinate annotations; Based on emergency response elements, nested conditional statements are identified, and a hierarchical instruction tree with failure branch warnings is created; whereby failure branch warnings are used to characterize invalid instructions corresponding to false emergency response element annotations in nested conditional statements. By reverse-engineering the hierarchical instruction tree from the historical case library, a multi-department collaborative task chain with weighted identifiers is generated.

2. The method for digitizing contingency plans based on large-scale modeling technology as described in claim 1, characterized in that, The cross-modal graph is also used for: Based on preset real-time emergency resource inventory data, adjust the connection strength between event nodes and environmental status nodes, and simultaneously optimize the semantic alignment between event features and meteorological data; When the SHAP value detects label anomalies, a dual-channel optimization of federated learning data completion and adversarial example generation is triggered; among them, adversarial examples are synthesized first from low-confidence event scenarios in historical cases.

3. The method for digitizing contingency plans based on large-scale modeling technology as described in claim 1, characterized in that, The prompt word template is also used for: Real-time monitoring of the trigger frequency and error correction data of different tags in users' historical operations; The weight allocation ratio of different tags is dynamically adjusted according to the industry risk level, and a hierarchical guiding prompt template containing priority indicators is generated, in which high-risk tags are automatically displayed first.

4. The method for digitizing contingency plans based on large-scale modeling technology as described in claim 1, characterized in that, The prompt word template is also used for: Analyze the abnormal event chain in historical operation records, extract semantic contradictions, missing data segments and rule conflict domains, and construct a multi-dimensional error feature vector library; Based on the error feature vector library, enhanced training samples with composite interference features are generated through semantic substitution, temporal perturbation and context shift algorithms. The sample complexity of the enhanced training samples is positively correlated with the error occurrence frequency. The severity of error types in the enhanced training samples is used to prioritize sample training. Multiple rounds of adversarial training are conducted on frequently occurring low-confidence events until the model output stability exceeds a preset threshold.

5. The method for digitizing a contingency plan based on large-scale modeling technology as described in claim 1, characterized in that, The prompt word template is also used for: Integrate a digital twin simulation interface into the prompt word template to generate a set of prompt words with simulation sandbox identifiers; Based on the indicators in the simulation sand table, a virtual scenario simulation space for emergency events is generated.

6. The method for digitizing contingency plans based on large-scale modeling technology as described in claim 1, characterized in that, The error correction backflow data update includes: The system collects three types of error correction sources: user feedback data, system execution logs, and expert review and marking. Abnormal data is cleaned using a conflict detection algorithm to generate a training sample set with confidence weights. Based on the training sample set, the hierarchical parameters of the prompt word information model are updated; the hierarchical process uses full parameter fine-tuning. Based on the fine-tuned prompt word information big model, a second digital instruction set is generated. When the verification accuracy of the second instruction set is lower than the preset threshold, it automatically reverts to the previous stable version and triggers the manual review process.

7. The method for digitizing contingency plans based on large-scale modeling technology as described in claim 1, characterized in that, The model interface fine-tuning also includes: Lightweight interface verification models are deployed on edge devices to filter low-quality error correction data in real time and upload high-value samples to the prompt word information big model; Key features are extracted from the updated prompt word information model to generate a simplified interface adapter for low-computing-power terminals to use. A unified interface protocol conversion framework is constructed to convert the second digital instruction set into an input format compatible with historical version models.

8. A contingency planning digitization system based on large-scale modeling technology, characterized in that, include: Tag configuration module: Used to create pre-configured multi-level tag systems for various industries; Template setting module: When user commands contain tag features in a multi-level tag system, generate prompt word templates that guide users with domain knowledge corresponding to the tag features; The first digital module is used to receive prompt word templates through a preset prompt word information big model, perform structured parsing of unstructured plans, and generate the first digital instruction set. The second digitization module is used to acquire the error correction feedback data of the first digitization instruction set, and update the prompt word model and the corresponding model interface fine-tuning through the error correction feedback data to generate the second digitization instruction set after model structure fine-tuning. The construction of the multi-level tagging system includes: Determine the industry classification code based on the pre-configured industries; The industry classification codes are converted into time-series feature vectors, and an industry-event spatiotemporal correlation matrix is ​​constructed based on real-time meteorological data and holiday information. A cross-modal graph containing event nodes, environmental state nodes, and social impact nodes is established. The cross-modal graph simulates the event evolution path based on a gated graph convolutional network, and calculates the probability of the effects of different intervention measures through counterfactual intervention. By combining the industry-event spatiotemporal correlation matrix and cross-modal graph with the event feature embedding layer and graph attention mechanism, a multi-level label system with confidence weights is output, and each label is marked by SHAP value; The structured parsing includes: Analyze textual plans, on-site images, and sensor data streams to generate emergency response elements with three-dimensional spatial coordinate annotations; Based on emergency response elements, nested conditional statements are identified, and a hierarchical instruction tree with failure branch warnings is created; whereby failure branch warnings are used to characterize invalid instructions corresponding to false emergency response element annotations in nested conditional statements. By reverse-engineering the hierarchical instruction tree from the historical case library, a multi-department collaborative task chain with weighted identifiers is generated.

Images