Preplan digitization method and system based on large model technology

Abstract

The invention relates to the field of digital emergency management, and provides a plan digitization method and system based on a large model technology, and the method comprises the steps: creating a multi-stage label system of a pre-configuration industry; when the user instruction has the tag features in the multi-level tag system, generating a cue word template guided by the domain knowledge corresponding to the tag features; receiving a cue word template through a preset cue word information large model, executing structured analysis of the unstructured plan, and generating a first digital instruction set; and obtaining error correction backflow data of the first digital instruction set, updating the cue word large model and the corresponding model interface fine tuning through the error correction backflow data, and generating a second digital instruction set after model structure fine tuning.

Description

Technical Field

[0001] The present application relates to the field of digital emergency management, and in particular to a method and system for digitalizing emergency plans based on large model technology. Background Art

[0002] In the internal emergency management of enterprises, digital emergency management is usually adopted. Digital emergency management is a comprehensive digital management system for emergency events through new generation information technologies such as big data, artificial intelligence, Internet of Things, blockchain, digital twins, etc.

[0003] It is mainly used in local service platforms and connected to cloud platforms. The cloud platforms are connected to software platforms. In traditional digital emergency plan processing, the software platforms manually define the label system and the structured rules of label recognition, and extract elements through keyword matching and template parsing to locate and analyze emergency events. However, it can only handle limited, fixed types of emergencies. However, for complex scenarios, there are multiple different events or emergencies are triggered by multiple tags in user instructions, and there are emergencies corresponding to multiple tags at one time. The capabilities are limited and cannot handle emergency events in complex scenarios.

[0004] In terms of the intelligence level of software platforms, they mostly use rule engines and simple machine learning models to analyze events and instructions. They cannot deeply understand emergency events in complex scenarios and can only generate fixed plans. Moreover, when parsing events and instructions, the ambiguity of user instructions may trigger incorrect plans, making it impossible to achieve dynamic automatic evolution and generation of plans.

[0005] Most existing simple machine learning models lack closed-loop optimization capabilities, so there is no automated error correction mechanism, the model iteration efficiency is low, and they are easily affected by human bias. They can only target fixed emergency events in emergency scenarios, and there are data silos. Resource scheduling in emergency command is out of touch with actual needs.

[0006] Application Contents

[0007] This application proposes a method and system for digitizing emergency plans based on large-scale model technology. By constructing a multi-level labeling system and a cross-modal graph neural network, it integrates the intelligent deconstruction of unstructured emergency plan elements with spatiotemporal correlation modeling. A dynamic parsing engine driven by large models improves the recognition accuracy of nested conditional statements and user instructions. The introduction of a command tree branch activation mechanism driven by real-time meteorological data enables dynamic matching of plan terms with environmental parameters, enabling automatic adaptation of emergency plans in complex scenarios.

[0008] To address the optimization shortcomings of traditional models, dual-channel error correction reflow, coupled with the synergy of federated learning and adversarial example generation, compresses model iteration cycles to hours while ensuring knowledge consistency during updates. For emergencies, personalized adjustments can be made, fine-tuning the model based on specific user instructions, making emergency plans more adaptable to complex environments.

[0009] In the first aspect, this application proposes a method for digitizing emergency plans based on large model technology, specifically including: creating a multi-level labeling system for pre-configured industries;

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

[0011] Receiving the prompt word template through a preset prompt word information macromodel, performing structured analysis of the unstructured plan, and generating a first digital instruction set;

[0012] The error correction return flow data of the first digital instruction set is obtained, and the prompt word information large model and the corresponding model interface are fine-tuned using the error correction return flow data to generate a second digital instruction set with fine-tuned model structure.

[0013] The method for digitizing the emergency plan provided by the embodiment of the present application is applied to the emergency event alarm process in the cloud. By receiving the instructions and analyzing the event tags of the emergency events in the user instructions, an automatically triggered emergency event guidance mechanism can be realized, avoiding the problem of needing to understand the contextual meaning and the high probability of manual intervention in the traditional emergency event handling process. The existing label system is also unable to dynamically adapt to the plans of different scenarios and can only adapt to fixed and set scenario plans. Then a knowledge-guided prompt word template corresponding to the label feature is generated. The prompt word template is associated with the label feature. The difference from 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 is guided accordingly based on the label feature. Through the prompt word template, after the prompt word information model receives the prompt word template, the emergency event scene is simulated and the structured analysis of the unstructured plan is performed, that is, the analysis of the non-fixed emergency event plan is performed to determine the digital instructions that need to exist in the plan, that is, the digital tasks that need to be performed. The error correction reflux data is an iterative error correction analysis mechanism implemented by the prompt word information large model during the uninterrupted analysis of the prompt word template. It can determine the task instructions that cannot be implemented and the missing digital instructions in the first digital instruction set, and realize the continuous updating of the prompt word information large model. The generated second digital instruction set is a new processing plan implementation instruction corresponding to the second digital instruction after the emergency event processing plan is revised.

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

[0015] Determine the industry classification code based on the pre-configured industry;

[0016] Convert industry classification codes into time series feature vectors, and construct an industry-event spatiotemporal correlation matrix based on real-time meteorological data and holiday information;

[0017] Establish a cross-modal graph containing event nodes, environmental state nodes, and social impact nodes. The cross-modal graph simulates the event evolution path based on a gated graph convolutional network and calculates the effectiveness probability of different treatment measures through counterfactual interventions.

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

[0019] In this embodiment, the purpose of determining the industry classification code in this application is to adjust the weight of the label according to meteorological data, holiday data and other data to achieve spatiotemporal adjustment. That is, the labels corresponding to the classification codes of emergency events in different industries are different in different meteorological data and holiday data. For example, the urgency of tourists stranded due to blizzards changes all the time during holidays and non-holidays, as well as in subsequent weather conditions. By combining the industry-event spatiotemporal association matrix and the cross-modal graph with the event feature embedding layer and the graph attention mechanism, this application can interpret different labels by means of integral gradients. Correspondingly, based on the same label, the interpretation information corresponding to the label is different in different current emergency events.

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

[0021] Based on preset real-time emergency resource inventory data, the connection strength between event nodes and environmental status nodes is adjusted, and the semantic alignment between event features and meteorological data is simultaneously optimized;

[0022] When the SHAP value detects the presence of label anomalies, it triggers dual-channel optimization of federated learning data completion and adversarial sample generation, where adversarial samples prioritize synthesizing low-confidence event scenarios in historical cases.

[0023] In this embodiment, the cross-modal graph can align the event content of the emergency event and the environmental conditions of the area where the emergency event is located, and the semantics related to the event conditions and the meteorological environment based on real-time emergency events, thereby preventing the problems of plan lag and unreasonable plans in traditional fixed plans. Through the dual-channel optimization of federated learning data completion and adversarial sample generation, it can be discovered whether there are low-confidence labels in the label features of the emergency event, and prevent the occurrence of unreasonable data or data that does not conform to the current status of the emergency event in the event factors of the emergency event, which will lead to deviations in the emergency plan and waste of resources.

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

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

[0026] Dynamically adjust the weight distribution ratio of different tags according to the industry risk level;

[0027] Generate a hierarchical guidance prompt word template containing priority identifiers, in which high-risk labels are automatically displayed in the front.

[0028] In the embodiment of the present application, the prompt word template always provides the prompt word information big model with derivation prompt words to determine the derivation target. Because the prompt word template is composed of label features, it is possible to determine the labels that are triggered by errors and distribute label weights based on the industry corresponding to the current emergency event, thereby increasing the weights of necessary and critical labels, and reducing the weights of labels that may have errors or have low triggering frequencies and low relevance. Then, during guidance, through priority identification, that is, different weights, different labels are graded to achieve hierarchical guidance, thereby improving the processing accuracy of emergency events in the big model.

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

[0030] Analyze abnormal event chains in historical operation records, extract semantic contradictions, data missing segments, and rule conflict domains, and build a multi-dimensional error feature vector library;

[0031] Based on the error feature vector library, enhanced training samples with composite interference features are generated through semantic replacement, temporal perturbation, and context shift algorithms. The sample complexity of the enhanced training samples is positively correlated with the frequency of error occurrence.

[0032] The sample training priority is set according to the severity of the error type of the enhanced training samples, and multiple rounds of adversarial training are performed on high-incidence low-confidence events until the model output stability exceeds the preset threshold.

[0033] This application will determine the combination of multiple features in series to form a multi-dimensional error feature vector library that is correlated through the chain of abnormal events in historical operation records, realize the training enhancement of compound perturbations driven by error features, and obtain a stable model based on multiple rounds of adversarial training. The stable model will improve the accuracy while reducing the correlation between labels when generating prompt word templates.

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

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

[0036] Based on the simulation sand table logo, a virtual scene simulation space for emergency events is generated.

[0037] Based on the prompt word template, this application generates a digital twin sandbox capable of simulation and deduction, realizes closed-loop deduction, dynamically processes the label information of emergency events, generates prompt words, verifies the relevance of prompt words, excludes irrelevant labels and prompt words, and can polish the prompt words.

[0038] In combination with the first aspect, the structured analysis includes:

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

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

[0041] The hierarchical instruction tree is reversely deduced through the historical case library to generate a multi-department collaborative task chain with weight identification.

[0042] During the process of structured analysis, this application will generate a three-dimensional space of the emergency event scene based on the prompt word template, mark the elements of emergency response in the three-dimensional space, and through the identification of emergency response elements, determine whether there are nested conditional statements, that is, there are related emergency event descriptions, and then generate a hierarchical instruction book for handling emergency events, with a failure branch warning. The failure branch warning is for emergency events that should exist when handling them based on the analysis of the prompt word template and the related emergency event descriptions, but may not exist when the emergency event is actually handled, or may have already been handled; then the hierarchical instruction book is reversely deduced through the historical case library to generate a variety of emergency handling tasks, forming a multi-department collaborative task chain.

[0043] In combination with the first aspect, the error correction reflow data update includes:

[0044] Collect three types of error correction sources: user feedback data, system execution logs, and expert review marks. Use conflict detection algorithms to clean abnormal data and generate training sample sets with confidence weights.

[0045] Based on the training sample set, the layered parameters of the prompt word information model are updated; among them, the layered parameters are fine-tuned;

[0046] Based on the fine-tuned prompt word information model, a second instruction set is generated. When the verification accuracy of the second instruction set is lower than the preset threshold, it is automatically restored to the previous stable version and the manual review process is triggered.

[0047] During implementation, this application uses different error correction sources and a conflict detection algorithm to determine whether there are abnormal data, forming a new training sample set. This new training sample set then undergoes a hierarchical data update within the prompt word information model. The updated hierarchical data results in a new instruction set being generated after the prompt word information model is updated. If the accuracy of the new instruction set falls below a preset threshold, the prompt word information model will be restored to its pre-update state.

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

[0049] Edge devices deploy lightweight interface verification models to filter low-quality error-correction data in real time and upload high-value samples to the large prompt word information model for model updates.

[0050] Extract key features from the updated prompt word information model and generate a simplified interface adapter for low-computing-power terminals to call;

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

[0052] In the embodiment of the present application, in the process of fine-tuning the model interface, edge verification is also implemented through the lightweight interface deployed by the edge device, and then the edge fine-tuning of the model is performed. After the edge fine-tuning, lightweight adaptation is performed. Therefore, in the system of the present application, when the system is not connected to the Internet and the computing power of the edge device is insufficient, the lightweight adaptation distillation method can also be applied to the model of the present application to implement the issuance of instructions for emergency events.

[0053] The second aspect is a digital emergency plan system based on large-scale model technology, including:

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

[0055] Template setting module: When the user instruction contains tag features in the multi-level tag system, a prompt word template guided by the corresponding domain knowledge of the tag features is generated;

[0056] A first digitization module is configured to receive a prompt word template through a preset prompt word information macromodel, perform structured analysis of an unstructured plan, and generate a first digitized instruction set;

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

[0058] The method for digitizing the emergency plan provided by the embodiment of the present application is applied to the emergency event alarm process in the cloud. By receiving the instructions and analyzing the event tags of the emergency events in the user instructions, an automatically triggered emergency event guidance mechanism can be realized, avoiding the problem of needing to understand the contextual meaning and the high probability of manual intervention in the traditional emergency event handling process. The existing label system is also unable to dynamically adapt to the plans of different scenarios and can only adapt to fixed and set scenario plans. Then a knowledge-guided prompt word template corresponding to the label feature is generated. The prompt word template is associated with the label feature. The difference from 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 is guided accordingly based on the label feature. Through the prompt word template, after the prompt word information model receives the prompt word template, the emergency event scene is simulated and the structured analysis of the unstructured plan is performed, that is, the analysis of the non-fixed emergency event plan is performed to determine the digital instructions that need to exist in the plan, that is, the digital tasks that need to be performed. The error correction reflux data is an iterative error correction analysis mechanism implemented by the prompt word information large model during the uninterrupted analysis of the prompt word template. It can determine the task instructions that cannot be implemented and the missing digital instructions in the first digital instruction set, and realize the continuous updating of the prompt word information large model. The generated second digital instruction set is a new processing plan implementation instruction corresponding to the second digital instruction after the emergency event processing plan is revised.

[0059] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained through the written description and the structures particularly pointed out in the accompanying drawings.

[0060] The technical solution of the present application is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0062] Figure 1 This is a method flow chart of a method for digitizing a plan based on large model technology in an embodiment of the present application;

[0063] Figure 2 This is a hardware diagram for implementing a plan digitization method based on large model technology in an embodiment of the present application;

[0064] Figure 3 A flowchart of the multi-level labeling system constructed in the embodiment of the present application;

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

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

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

[0068] Figure 7 This is a schematic diagram of the implementation process of the digital twin deduction integration interface of the embodiment of the present application;

[0069] Figure 8 Schematic diagram of the implementation process of the multi-source structured parsing engine in the embodiment of the present application;

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

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

[0072] Figure 11 This is a system composition diagram of a plan digitization 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. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application and are not used to limit the present application.

[0074] In order to solve the technical defects in the process of generating digital emergency plans in the background technology, the embodiment of the present application provides a plan digitization method based on large model technology.

[0075] Example 1:

[0076] As attached Figure 1As shown, this application proposes a plan digitization method based on large model technology, which realizes the transformation and continuous optimization of unstructured plans into structured instructions through the combination of a multi-level label system and prompt word guidance.

[0077] The method of this application specifically includes:

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

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

[0080] By presetting a hierarchical structure based on industry characteristics, complex plan contents can be broken down into manageable knowledge units.

[0081] For example, in the field of emergency incident handling, three levels of labels are divided:

[0082] The first-level label can distinguish categories such as natural disasters, accidents and disasters;

[0083] Secondary labels are refined to scenarios such as earthquakes and fires;

[0084] Level 3 tags may involve specific elements such as evacuation routes and material allocation;

[0085] It can also be divided into different levels of labels according to the urgency of emergency events, user instructions; it can also be based on the scene elements of the same emergency event; for example, in a fire scene, the fire area and igniting materials, as well as the number of trapped people, are used as first-level labels, firefighting supplies and fire water supply areas in the fire area are used as second-level labels, and flammable and non-flammable areas of the fire scene are used as third-level labels.

[0086] Combined with a multi-level labeling system, the system improves user input information through intelligent question-answering. It can also automatically extract information about the corresponding area and generate label features based on the user's direct instructions. This enables the large model to quickly locate the handling measures for wire and pipe emergency events, avoiding processing deviations caused by information overload or insufficient information.

[0087] Step S101: When the user instruction has a tag feature in the multi-level tag system, a prompt word template guided by the domain knowledge corresponding to the tag feature is generated;

[0088] In this application, when a user instruction is received, the tag features are determined by parsing the user instruction, and the tag features are matched based on the multi-level tag system. At this time, the industry corresponding to the user instruction is the pre-configured industry of the multi-level tag system. The annotation filling of the tag features by the prompt word template is extended to form a prompt word template based on domain knowledge guidance. The domain knowledge guidance is the knowledge guidance related to emergency events associated with the tag features in the pre-configured industry. Specifically, through semantic analysis, the key tags for handling emergency events and based on the requirements for handling emergency events in user instructions are identified. This process will be combined with the industry knowledge base of the preset industry to form a structured query template.

[0089] For example, when the input instruction contains a label feature word such as "chemical fire", the prompt word template will be a chemical emergency event, and the domain knowledge of the corresponding emergency event will be guided through the chemical fire information. It will automatically associate knowledge elements related to the chemical field such as hazardous chemicals disposal procedures and fire protection resource distribution to generate a prompt template containing specific parameter constraints. Specific parameter constraints are characterization parameters used to characterize the maximum and minimum impact of event elements on chemical fires in the corresponding chemical fire scenario after parameterization of event elements that may exist in chemical fires. It is used to solve the problem of model output deviation caused by vague prompt word expressions in traditional methods. By integrating the industry knowledge system, it converts user instructions into actionable prompt projects, allowing simple user instructions to become more professional and accurate instruction-guided prompt information for emergency event handling.

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

[0091] This application performs structured analysis of unstructured plans through a large model of prompt word information. In this process, when processing the text information of the prompt word template, the large model of prompt word information integrates the scattered emergency plans through its semantic understanding ability and large model deduction ability, combined with the constraints of the prompt word template, and extracts the key information on emergency event handling in the scattered plans that is related to the prompt word information.

[0092] For example, emergency procedures, responsible parties, implementation standards, emergency response measures, emergency dispatch information, etc. are converted into standardized instruction nodes.

[0093] For example, "Immediately initiate the three-level chemical fire response mechanism" will contain structured data such as trigger conditions, response level, dispatch department, chemical area, chemical fire emergency response tools, emergency response precautions, and material allocation instructions after parsing. In the scattered emergency plan, the emergency plan processing items associated with the structured data are called;

[0094] Furthermore, by processing complex semantic expressions, the degree of automation of digital plans can be improved, the cost of manual labeling can be reduced, and the integrity of business logic can be maintained.

[0095] S103: Obtain error correction return flow 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 return flow data to generate a second digital instruction set after the model structure is fine-tuned.

[0096] This application updates the model and fine-tunes the interface based on error correction return data. It uses error cases in actual applications, such as missing instruction parameters, incorrect logical sequence, and other data, as training data. It optimizes the model weight parameters through incremental training and adjusts the data verification rules of the API interface.

[0097] For example, when a defect is discovered, such as "the material allocation instruction does not include the transport vehicle type," mandatory verification of this field is added in subsequent model reasoning. This enables adaptive evolution, dynamically adapting to industry standard updates and business scenario expansion, and preventing traditional static models from becoming rigid and subject to performance degradation.

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

[0099] Figure 2 This is an implementation scenario diagram of a digital emergency handling method, such as Figure 2 As shown, the implementation system of the digital method of emergency plan is deployed in the cloud center 1, edge terminal 2 and terminal access terminal 3 respectively. Among them, the edge terminal 2 is the user-oriented and service provider-oriented digital system of the emergency plan, which is used to realize command distribution and data transmission when an emergency event occurs.

[0100] Exemplarily, the user instructions of the present application are received by the terminal access end 3. The user can receive the user instructions through terminal devices such as environmental sensors 31, smart cameras 32, wearable devices 33, tablets, computers, mobile phones, etc., and transmit them to the edge computing node 21 through the industrial Ethernet or the Internet through the field switch 22. The function of the edge computing node 21 is to analyze the instructions and determine the label features through a multi-level label system, and then generate a prompt word template based on the label features under the guidance of domain knowledge.

[0101] The prompt word template is transmitted to the edge gateway via the 5G network. The edge network 14 is the gateway of the cloud center. The unstructured plan is parsed through the prompt word information model within the distributed computing cluster 13. The digital instruction set generated by the unstructured plan marking is not systematic. It is just a splicing of scattered plans. It can solve emergency events, but there may be plans in this plan that are unrelated to the elements of the emergency event, and there may also be plans that are related to the elements of the emergency event, but it is not enough to solve the inadequacy of the plan for the emergency event.

[0102] The cloud center's active-active data center 11 and core router 12 perform encryption and provide a high-speed SD-WAN network for emergency response. The active-active data center 11 also updates the large prompt word information model by correcting the return data, thereby fine-tuning the model and generating a second digital instruction set.

[0103] After receiving the second digital instruction set issued by the distributed computing cluster 13, the edge gateway 14 controls the emergency command terminal 24 through the edge computing node 23, that is, the server of the emergency handling personnel management end, based on the Mesh network, using communication methods such as Bluetooth and WiFi, and controls the emergency tablet 34 to display the emergency handling plan to the emergency event handling personnel, and controls the rescue equipment 35 through Bluetooth so that the rescue equipment 35 can be used by the emergency event handling personnel.

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

[0105] Example 2:

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

[0107] First, this application determines the industry classification code of the industry corresponding to the emergency event through the emergency command terminal 24 of the edge terminal 2, establishes a reference coordinate system for labels that associate emergency events with industry knowledge, and constructs a multi-level label system through the reference coordinate system. Different coordinate points of the reference coordinate system are associated labels corresponding to different emergency events, and the temporal and spatial correlation between industries and events is used as the benchmark, specifically the correlation between emergency events and industries, as well as the regularity of emergency events in the occurrence of different meteorological data and different holiday information.

[0108] For example, this application converts classification codes into time series feature vectors and incorporates real-time meteorological data and holiday information. By constructing a spatiotemporal correlation matrix, the labeling system can capture the periodic laws of industry events in the time dimension and the environmental dependence in the spatial dimension.

[0109] This application will also construct a cross-modal graph to determine the effects and probabilities of different response measures in emergency events. It can be understood that this application associates event nodes, environmental state nodes, and social impact nodes, uses a gated graph convolutional network to simulate the possible paths of event evolution over time, and uses counterfactual intervention technology to evaluate the effectiveness probabilities of different response measures, giving the labeling system a predictive effect on the evolution of emergency events;

[0110] Finally, this application fuses the spatiotemporal correlation matrix with the cross-modal graph analysis results through the event feature embedding layer and the graph attention mechanism, outputs multi-level labels with confidence weights, and uses interpretability tools to mark the importance of each label in the multi-level labels to form label body information with high confidence and can be explained.

[0111] Determine the industry classification code based on the pre-configured industry;

[0112] It is understandable that this application targets emergency events in different industries, and there are significant differences in the plan structure, terminology system and focus. For example, the power industry is more concerned with equipment operation and maintenance and grid stability, while the medical industry focuses on diagnosis and treatment processes and emergency response. The setting of industry classification codes is like establishing exclusive language dictionaries for different fields. Through the language dictionary, the industry classification codes are converted into time series feature vectors, and based on real-time meteorological data and holiday information, an industry-event spatiotemporal correlation matrix is ​​constructed, through dynamic variables in multiple dimensions such as time, weather, and environment.

[0113] Time series feature vectors can capture the periodic patterns of industry events over time;

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

[0115] By constructing a spatiotemporal correlation matrix, this application makes the labeling system no longer a static classification framework, but an intelligent model that can be dynamically adjusted over time and with environmental changes, thereby enhancing the adaptability of labels to actual scenarios.

[0116] This application establishes a cross-modal graph containing event nodes, environmental state nodes, and social impact nodes. The cross-modal graph simulates the event evolution path based on a gated graph convolutional network and calculates the effect probability of different disposal measures through counterfactual interventions.

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

[0118] For example, in a chemical industry emergency plan, the event node may be a pipeline leak, the environmental status node includes information such as temperature and wind speed, and the social impact node involves information such as the safety of surrounding residential areas.

[0119] Then, through the gated graph convolutional network, the dynamic interaction between nodes is captured to simulate all possible scenarios from the occurrence to the development of the event;

[0120] The counterfactual intervention technology of this application is to use the "hypothesis-deduction" method to sequentially target the possible consequences of different disposal measures during the evolution of the event, that is, the disposal results;

[0121] For example, the contingency plan is to close the valve in advance. The counterfactual intervention technology can determine whether the degree of harm caused by the accident can be reduced, so that the label system can not only describe the characteristics of the event.

[0122] This application combines the industry-event spatiotemporal correlation matrix and cross-modal graph with the event feature embedding layer and graph attention mechanism to output a multi-level label system with confidence weights, and marks each label with a SHAP value.

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

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

[0125] For example: In a medical emergency plan, the golden emergency response time label is clearly displayed through the SHAP value.

[0126] Example 3:

[0127] The cross-modal graph of this application can also prevent the problems of plan lag and unreasonable plan in traditional fixed plans, find out whether there are low-confidence labels in the label features of emergency events, and prevent the occurrence of unreasonable or inconsistent data in the event factors of emergency events, which will lead to deviations in emergency plans and waste of resources. Figure 4 :

[0128] This application adjusts the connection strength between event nodes and environmental status nodes based on preset real-time emergency resource inventory data, and simultaneously optimizes the semantic alignment between event features and meteorological data;

[0129] In this application, real-time emergency resource inventory data, the edge computing node 23 will determine the storage status of materials for emergency event handling through the emergency command terminal 24; the connection strength is used to characterize the trend of increasing or decreasing the difficulty of emergency event handling due to the command information corresponding to the emergency event and the environmental conditions in the emergency event handling area, and unify the event characteristics revealed in the user instructions corresponding to the emergency event and the meteorological data in the description of the same situation, so as to allocate emergency resources more quickly compared with traditional emergency event handling measures.

[0130] For example, in a flood disaster response plan, if real-time monitoring indicates insufficient sandbag inventory in a certain area, the event node prompts: "Dyke reinforcement"; the environmental status node prompts: "Continuous rainfall." The connection between the two nodes is automatically strengthened to prevent resource shortages from exacerbating the impact of environmental factors on the evolution of the event.

[0131] Synchronous optimization of semantic alignment is achieved by adjusting the data mapping relationship so that event descriptions, rainstorm warnings, and meteorological data such as rainfall and wind speed can be accurately matched at the semantic level, avoiding misunderstandings caused by terminology differences.

[0132] For example: establish a clear correspondence between extreme weather labels in industry plans and specific thresholds in meteorological data, where rainfall exceeds a certain standard, to prevent misunderstandings such as low rainfall.

[0133] This application triggers dual-channel optimization of federated learning data completion and adversarial sample generation when the SHAP value detects label anomalies; among them, adversarial samples prioritize synthesizing low-confidence event scenarios in historical cases.

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

[0135] For example, in a chemical spill emergency plan, the label is: evacuation route for nearby personnel. The SHAP value of this label suddenly drops, indicating that: in the current model, the weight of the label is too low, triggering the optimization mechanism;

[0136] Federated learning data completion securely aggregates data from multiple sources, such as emergency cases in different regions, to supplement missing event scenario data without compromising privacy, thus solving the data sparsity problem.

[0137] Adversarial sample generation targets historically low-confidence scenarios and complex disaster scenarios. It uses algorithms to synthesize difficult samples close to the real distribution, allowing the model to have characteristic differences that are easily confused by the scenarios.

[0138] For example, we can generate adversarial examples where high temperature and drought occur simultaneously with equipment failures to improve the model’s accuracy in complex working conditions.

[0139] Example 4:

[0140] This application prompt word template is used to display labels, see Figure 5 :

[0141] In the process of transmitting the prompt word template to the prompt word information large model, the edge computing node 21 of the present application monitors the triggering frequency and error correction data of different tags in the user's historical operations in real time;

[0142] The frequency with which users trigger different tags during use is used to characterize the focus of attention in specific industry scenarios;

[0143] For example, in the power operation and maintenance plan processing, the label "equipment fault troubleshooting" is triggered significantly more frequently than other labels, indicating that users are more concerned about the fault handling process;

[0144] The error correction data in this application is used to correct the shortcomings of the current prompt word template;

[0145] Exemplary: Error correction display: label parsing error of the emergency power supply switching step, generating prompt information, and strengthening the domain knowledge guidance of the corresponding label in the edge computing node 21.

[0146] Through error correction and frequency supervision, it determines whether the prompt word template needs to be optimized, learns from user behavior, and does not rely on fixed preset rules.

[0147] This application will dynamically adjust the weight distribution ratio of different tags according to the industry risk level;

[0148] Because different industries have different risk distribution characteristics, for example, labels such as "hazardous material leakage" in the chemical industry and emergency response time in the medical industry are high-risk scenarios. The industry risk assessment model can be used to determine the risk level of the label.

[0149] This application is used when an industry enters a high-risk period. For example, high temperatures in summer cause a surge in power load, and the weight of the relevant high-risk labels will automatically increase. Then, when parsing the plan, the prompt word template will focus on analyzing the corresponding instruction information for the key steps that may cause serious consequences, to prevent digital instruction output errors due to unimportant information.

[0150] This application will generate a hierarchical guided prompt word template containing priority identification, in which high-risk labels are automatically displayed in front.

[0151] This application addresses traditional prompts by presenting information in a straightforward manner, requiring users to filter for key content. A layered, guided template prioritizes labels and prioritizes high-risk labels. For example, in a fire emergency plan, high-risk labels such as "Safe Evacuation Routes" and "Location of Firefighting Equipment" appear at the beginning of the prompt, guiding users and the large model to generate the corresponding plan first. This approach improves information processing efficiency, speeding up the time it takes to generate a plan by approximately four seconds, and its accuracy aligns with the operational logic of prioritizing high-risk tasks in emergency scenarios.

[0152] Example 5:

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

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

[0155] In this application, the abnormal event chain is used to represent the logical faults that occur during the plan analysis process, such as: the order of handling steps is reversed, key parameters of information gaps are missing, or different clauses of the rules conflict with each other in describing the same operation;

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

[0157] For example, in a chemical accident emergency plan, the generated plan was incorrect because the relationship between "temperature threshold" and "pressure parameter" was not recognized. The semantic contradiction was extracted and recorded as a "cross-parameter logical break" feature, incorporated into the vector library, and optimized and adjusted. While summarizing the errors, modeling was also carried out on the error-prone points, and targeted training was carried out.

[0158] Based on the error feature vector library, this application generates enhanced training samples with composite interference features through semantic replacement, temporal perturbation and contextual offset algorithms. The sample complexity of the enhanced training samples is positively correlated with the frequency of error occurrence.

[0159] The semantic substitution in this application simulates the language diversity in real scenarios by replacing synonymous terms and adjusting the expression sentence structure. For example, "immediately cut off the power" is replaced with "disconnect the circuit immediately" to test the model's consistent understanding of different expressions.

[0160] Temporal perturbations are used to disrupt the temporal order of events or extend the time intervals between key steps, thereby testing the model's logical reasoning ability under non-standard temporal sequences.

[0161] The scenario shift algorithm increases or decreases environmental variables. For example, it can add additional conditions such as "wind speed exceeding the standard" and "traffic control" to the rainstorm plan.

[0162] Then, a complex scenario with multiple factors coupled is constructed. For emergency events with higher error frequency, the corresponding labels of the event scenarios will have higher sample complexity.

[0163] For example: "Data missing" errors occur frequently, and will be configured as composite interference samples that contain both semantic ambiguity and time sequence disorder, so that reliable instructions can still be output under extreme conditions of incomplete information and unclear expressions.

[0164] This application sets the sample training priority based on the severity of the error type of the enhanced training samples, and implements multiple rounds of adversarial training on high-incidence 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 the industry risk level and actual impact consequences. In a medical emergency plan, "wrong sequence of first aid measures" is a high-severity error, and its training priority is significantly higher than low-impact errors such as "incorrect terminology";

[0166] For example: high-incidence low-confidence events represent scenarios that frequently appear in historical operations, but the processing results of prompt word templates fluctuate greatly; for example: the handling of "multiple equipment cascading failures" in the power system, because it involves multi-link linkage and there are few historical cases, it is easy for the confidence to drop sharply during model analysis.

[0167] This application uses multiple rounds of adversarial training, that is, the prompt word "last class" repeatedly iterates between generating samples and correcting outputs, gradually strengthening the ability to handle complex and difficult examples until the output results are stable in similar scenarios, such as: the confidence fluctuation range, parsing accuracy, etc. are not in line with reality.

[0168] Example 6:

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

[0170] This application integrates the digital twin simulation interface into the prompt word template to generate a prompt word set with a simulation sandbox logo;

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

[0172] For example: In a fire emergency plan, the prompt word "start the fire sprinkler system" will be associated with the sprinkler equipment parameters, coverage area and other information in the digital twin model to generate a prompt word with a specific deduction logo. This logo not only contains text instructions, but also carries multi-dimensional data such as the equipment location and linkage logic in the virtual space.

[0173] The integrated prompt word approach transforms a single semantic guidance tool into a digital bridge connecting real-life plans and virtual simulations.

[0174] This application generates a virtual scene simulation space for emergency events based on the simulation sand table identification.

[0175] For example: the virtual scene simulation space is not a static three-dimensional model, but a dynamic simulation environment built based on real-time data and historical cases, which can render the event evolution process in real time according to the handling instructions in the prompt words.

[0176] For example: in a flood disaster simulation, after entering the prompt "stack sandbags to reinforce the embankment", the simulation space will dynamically simulate the impact of the location and quantity of sandbags on the direction of the flood based on the rainfall and water level changes in real-time meteorological data and the embankment material and surrounding terrain in geographic information. It can even show the differences in disaster development paths under different disposal plans.

[0177] It is not limited to logical analysis at the text level, but can "preview" the actual effects of disposal measures in a virtual space to discover potential problems in advance, such as omissions in disposal steps and unreasonable resource allocation.

[0178] Example 7:

[0179] For a structural analysis of this application, see Figure 8 :

[0180] In this application, text plans, on-site images, and sensor data streams are parsed to generate emergency response elements annotated with three-dimensional spatial coordinates;

[0181] Traditional plan analysis relies more on text information, but in real emergency scenarios, sensor data such as accident site layout, equipment operating parameters, and environmental indicators are more likely to describe emergency event scenarios.

[0182] For example: In a building fire emergency plan, the location of the fire source is identified through images and combined with the temperature data of the sensor to generate disposal elements such as "fire hydrant deployment points" and "safe evacuation passages" with three-dimensional coordinates. The instruction set not only contains text descriptions but also has spatial positioning, which solves the "where" execution positioning problem in traditional technologies.

[0183] This application identifies nested conditional statements, creates a hierarchical instruction tree with failure branch warnings, and dynamically activates / freezes corresponding branches based on real-time meteorological data;

[0184] The emergency plan of this application often contains nested conditions such as "If situation A occurs, execute step B; if step B fails, start plan C". Traditional analysis is prone to missing branches due to the complex logical hierarchy.

[0185] This application constructs a hierarchical instruction tree, where 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 a flood warning plan, when the real-time rainfall exceeds the preset threshold, the "start backup drainage pump" branch is automatically activated. If it fails due to equipment failure, an early warning is triggered and related instructions are frozen, ensuring that the analysis results can be dynamically adjusted with environmental changes, solving the conditional response problems of "when" and "how".

[0187] This application reversely deduces the responsible parties of instructions through the historical case library and generates 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 various departments in past emergency events. By analyzing cases such as "a chemical leak accident caused by a delayed response by a certain department, resulting in amplified consequences", the responsible parties for each step in the current response plan can be reversely deduced, for example, whether the responsible party is the environmental protection department, the fire department, or the production unit;

[0189] Departments are assigned weights based on their responsibilities and historical performance. For example, when generating a "pollutant detection and isolation" task, the environmental protection department is clearly designated as the primary responsible unit, with the production unit as the supporting unit. This weighting determines the priority of task assignments and the order of resource allocation, resolving the "who" (or "whose") responsibility is, and avoiding unclear responsibilities during execution.

[0190] Embed an expert collaborative annotation window in the automated parsing process to convert manual correction traces into parsing model training samples in real time;

[0191] Exemplary:

[0192] While automated analysis is highly efficient, expert experience is crucial in extremely complex scenarios, often involving new disasters and multiple incidents. This application embeds a annotation window, allowing experts to correct deviations in analysis results in real time and generate emergency responses.

[0193] For example, in an earthquake secondary disaster plan, experts manually adjusted the execution order of “gas pipeline shutdown” and “power cutoff.” The correction traces were converted into training samples, and the analytical model was reversely optimized to solve the continuous optimization problem of “how to improve.”

[0194] During the actual implementation process, it will also connect to the emergency response countdown device and prioritize parsing time-sensitive instructions. If the instructions are not executed within the time limit, the backup plan injection interface will be automatically triggered.

[0195] During emergency response, the "golden rescue time" is crucial to the outcome. A countdown device monitors task execution progress in real time, prioritizing time-sensitive instructions such as "Cardiopulmonary Resuscitation" and "Evacuate hazardous areas." The parsing system prioritizes these instructions and generates a timestamped execution plan. If a command (such as "Start emergency power") times out due to equipment failure, the system automatically triggers a backup plan (such as switching to battery power), preventing a single delay from causing a global failure and addressing the "when must do" issue.

[0196] Example 8:

[0197] For the error correction and reflux data update process of this application, please refer to Figure 9 :

[0198] This application collects three types of error correction sources: user feedback data, system execution logs, and expert review marks, cleans abnormal data through a conflict detection algorithm, and generates a training sample set with confidence weights;

[0199] For example, user feedback data can represent operational deviations in actual applications, such as manual correction of instruction errors;

[0200] Inherent defects in the execution logging model, including abnormal parsing time and broken output logic;

[0201] Expert review marks incorporate authoritative judgments of domain knowledge, but professional terminology is often misinterpreted. This application cross-validates three types of data sources through a conflict detection algorithm; for example: comparing the consistency of user feedback and expert marks, eliminating abnormal data and erroneous feedback caused by misoperation; finally, based on the data source, confidence weights are used to ensure that the training sample set is both comprehensive and pure.

[0202] Perform layered parameter updates on the large prompt word model; the industry-specific layer uses full parameter fine-tuning, while the general semantic layer uses lightweight updates using low-value adapters;

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

[0204] This application fine-tunes the parameters of the industry-specific layer to deeply optimize and avoid problems in error correction data. For example, in the chemical industry's "hazardous materials disposal process", the large model of prompt word information improves the accuracy of understanding specific field terminology and logic.

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

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

[0207] Model parameter updates may cause changes in the field definitions and format specifications of the output data. For example, adding a "responsible party" field or adjusting the time format may cause front-end display anomalies or back-end system connection failures if these changes are not made in a timely manner.

[0208] The automatic reconstruction mechanism dynamically updates the parameter mapping rules of the API interface by parsing 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] While the internal structure of the prompt word information model in this application has changed, the external interface still maintains data format consistency and compatibility. For example, after adding an "Environmental Risk Level" output field to the model, the interface will automatically map this field to the corresponding table in the backend database to avoid system errors caused by missing fields.

[0210] When the verification accuracy of the second instruction set falls below a preset threshold, the system automatically reverts to the previous stable version and triggers a manual review process. Verification accuracy is a core metric for measuring model performance. When an updated model performs poorly in critical scenarios or when analyzing emergency plans for high-risk industries, automatically reverting to the previous stable version can avoid actual losses caused by erroneous outputs. For example, in a financial risk control plan, if the updated model's accuracy in identifying "abnormal transactions" suddenly drops, the system immediately reverts to the previous stable version to prevent the risk of misjudgment from escalating.

[0211] At the same time, a manual review process is triggered, during which domain experts conduct an in-depth investigation of the model update process to check whether the correction data contains outdated cases and whether the parameter adjustments deviate from industry rules, ensuring that the root cause of the problem is accurately identified and corrected, and avoiding self-reinforcing errors in the automated process due to data bias or algorithm limitations.

[0212] Example 9:

[0213] The model interface of this application is fine-tuned during the lightweight processing process, see Figure 10 :

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

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

[0216] The lightweight verification model of this application uses a preset rule engine to perform data integrity checks and outlier detection to conduct preliminary screening of local data, and upload samples with clear error correction value, key step errors marked by experts, and frequently occurring analytical deviations to the cloud.

[0217] Extract key features from the updated large model and generate a simplified interface adapter for low-computing power terminals to call; traditional large models have huge parameter scales and are difficult to deploy directly on low-computing power devices such as mobile phones and portable terminals.

[0218] Through feature extraction technology, the key features most relevant to interface calls are extracted from the large cloud model to generate a streamlined adapter with small size and high operating efficiency.

[0219] For example, in fire emergency scenarios, low-computing-power handheld terminals can quickly parse basic instructions such as "fire extinguisher type selection" and "escape route planning" by calling streamlined adapters without loading a complete plan parsing model. This not only ensures the response speed of the terminal device, but also reduces hardware configuration requirements and expands the equipment coverage of technology applications.

[0220] During actual implementation, a unified interface protocol conversion framework will also be constructed to convert the second digital instruction set into an input format compatible with the historical version model. In actual applications, users may use different versions of models or old systems at the same time. If the newly generated second instruction set (such as adding fields such as "responsible party" and "time priority") cannot be recognized by the historical version, it will cause the system to fail to connect. The unified protocol conversion framework establishes a mapping relationship between new and old fields. For example, it converts the "multi-department collaborative task chain" in the new version into a "task assignment list" that can be recognized by the old version, and converts the data format in real time 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, during the power system upgrade process, the conversion framework automatically converts the processing elements containing three-dimensional spatial coordinates in the new version into two-dimensional location information that can be received by the old monitoring system, avoiding repeated development costs caused by system incompatibility.

[0222] Dynamically adjust the interface fine-tuning frequency based on the CPU and memory load of the end device. The hardware performance of different end devices varies significantly. High-computing servers can support frequent model fine-tuning to quickly adapt to new scenarios, while frequent fine-tuning on low-computing embedded devices may cause performance lag. The dynamic adjustment strategy monitors device metrics such as CPU usage and memory utilization in real time to adaptively adjust the interface fine-tuning frequency. When the device load is below a threshold, such as a high-performance server in a monitoring center, the fine-tuning frequency is increased to quickly absorb the latest error correction data. When the device load is excessively high, such as a battery-powered field monitor, the fine-tuning frequency is reduced or non-essential updates are suspended to ensure stable operation of core functions. For example, in field emergency scenarios, portable devices automatically enter "energy-saving mode" when the battery level drops below 20%, reducing the interface fine-tuning frequency from once per minute to once every ten minutes. This ensures critical command parsing while extending device battery life.

[0223] Example 10:

[0224] This application proposes a plan digitization system based on large model technology, see Figure 11 ,include:

[0225] Label configuration module: used to create a multi-level label system for pre-configured industries; Template setting module: when the user's instruction contains label features in the multi-level label system, it generates a prompt word template guided by the corresponding domain knowledge of the label features;

[0226] A first digitization module is configured to receive a prompt word template through a preset prompt word information macro model, perform structured analysis of an unstructured plan, and generate a first digitized instruction set;

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

[0228] Those skilled in the art will appreciate that various aspects of the present application, or possible implementations of various aspects, may be embodied as systems, methods, or computer program products. Therefore, various aspects of the present application, or possible implementations of various aspects, may be implemented entirely in hardware or entirely in software.

[0229] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for digitalizing emergency plans based on large model technology, characterized in that: include: Create a multi-level labeling system with pre-configured industries; When the user instruction contains label features in the multi-level label system, a prompt word template guided by the domain knowledge corresponding to the label features is generated; Receiving the prompt word template through a preset prompt word information macromodel, performing structured analysis of the unstructured plan, and generating a first digital instruction set; The error correction return flow data of the first digital instruction set is obtained, and the prompt word information large model and the corresponding model interface are fine-tuned using the error correction return flow data to generate a second digital instruction set with fine-tuned model structure.

2. A method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The multi-level labeling system construction includes: Determine the industry classification code based on the pre-configured industry; Convert industry classification codes into time series feature vectors, and construct an industry-event spatiotemporal correlation matrix based on real-time meteorological data and holiday information; Establish a cross-modal graph containing event nodes, environmental state nodes, and social impact nodes. The cross-modal graph simulates the event evolution path based on a gated graph convolutional network and calculates the effectiveness probability of different treatment measures through counterfactual interventions. The industry-event spatiotemporal correlation matrix and cross-modal graph are combined with the event feature embedding layer and graph attention mechanism to output a multi-level label system with confidence weights, and each label is marked with a SHAP value.

3. A method for digitizing a plan based on large model technology as claimed in claim 2, characterized in that: The cross-modal graph is also used to: Based on preset real-time emergency resource inventory data, the connection strength between event nodes and environmental status nodes is adjusted, and the semantic alignment between event features and meteorological data is simultaneously optimized; When the SHAP value detects label anomalies, it triggers dual-channel optimization of federated learning data completion and adversarial sample generation; among them, adversarial samples prioritize synthesizing low-confidence event scenarios in historical cases.

4. The method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The prompt word template is also used for: Real-time monitoring of different tag triggering frequencies and error correction data in user historical operations; The weight distribution ratio of different labels is dynamically adjusted according to the industry risk level, and a hierarchical guided prompt word template containing priority identification is generated, among which high-risk labels are automatically displayed in front.

5. The method for digitalizing a plan based on large model technology as claimed in claim 1, characterized in that: The prompt word template is also used for: Analyze abnormal event chains in historical operation records, extract semantic contradictions, data missing segments, and rule conflict domains, and build 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 replacement, temporal perturbation, and context shift algorithms. The sample complexity of the enhanced training samples is positively correlated with the frequency of error occurrence. The sample training priority is set according to the severity of the error type of the enhanced training samples, and multiple rounds of adversarial training are performed on high-incidence low-confidence events until the model output stability exceeds the preset threshold.

6. The method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The prompt word template is also used for: Integrate the digital twin simulation interface into the prompt word template to generate a prompt word set with a simulation sandbox logo; Based on the simulation sand table logo, a virtual scene simulation space for emergency events is generated.

7. The method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The structured analysis includes: Parse text plans, on-site images, and sensor data streams to generate emergency response elements with three-dimensional spatial coordinate annotations; Based on the emergency response elements, nested conditional statements are identified and a hierarchical instruction tree with failure branch warnings is created; wherein the failure branch warnings are used to represent invalid instructions corresponding to false emergency response element annotations in nested conditional statements; The hierarchical instruction tree is reversely deduced through the historical case library to generate a multi-department collaborative task chain with weight identification.

8. The method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The error correction reflux data update includes: Collect three types of error correction sources: user feedback data, system execution logs, and expert review marks. Use conflict detection algorithms to clean abnormal data and generate training sample sets with confidence weights. Based on the training sample set, the layered parameters of the prompt word information model are updated; among them, the layered parameters are fine-tuned; Based on the fine-tuned prompt word information model, a second digital instruction set is generated. When the verification accuracy of the second instruction set is lower than the preset threshold, it is automatically restored to the previous stable version and the manual review process is triggered.

9. The method for digitizing a plan based on large model technology as claimed in claim 1, characterized in that: The model interface fine-tuning also includes: Edge devices deploy lightweight interface verification models to filter low-quality error correction data in real time and upload high-value samples to the large prompt word information model; Extract key features from the updated prompt word information model and generate a simplified interface adapter for low-computing-power terminals to call; Build a unified interface protocol conversion framework to convert the second digital instruction set into an input format compatible with the historical version model.

10. A plan digitization system based on large model technology, characterized by: include: Label configuration module: used to create a multi-level label system for pre-configured industries; Template setting module: When the user instruction contains tag features in the multi-level tag system, a prompt word template guided by the domain knowledge corresponding to the tag features is generated; A first digitization module is configured to receive a prompt word template through a preset prompt word information macro model, perform structured analysis of an unstructured plan, and generate a first digitized instruction set; The second digitization module is used to obtain the error correction return data of the first digitization instruction set, and update the prompt word large model and the corresponding model interface fine-tuning through the error correction return data to generate the second digitization instruction set after the model structure is fine-tuned.

Images