One-network unified management event intelligent processing method, device and equipment and storage medium

Abstract

The invention provides a one-network unified management event intelligent processing method, device and equipment and a storage medium, and relates to the technical field of computers. The method comprises the steps of obtaining historical event information and categories thereof, performing vectorization operation on the event information, and extracting feature vectors of the event information; constructing a classification model by taking the feature vector of the event information as the input of the model and the category of the event as the output of the model; selecting an optimal hyper-parameter of the classification model by using conditional probability distribution, substituting the optimal hyper-parameter into the classification model, and training the classification model; classifying the target event by using the trained classification model to obtain rule information of a category to which the target event belongs; in combination with the target event information and the rule information, constructing a new prompt, inputting the new prompt into the enhanced Agent, and giving a treatment plan corresponding to the target event; and completing the disposal process of the target event based on the disposal plan. Compared with a traditional IT event processing mode, the method is higher in processing efficiency.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, equipment and storage medium for intelligently handling events managed by one network. Background Art

[0002] In the current one-network unified management system, the number of urban governance events reported to the one-network unified management system every day is very large. Taking a first-tier city as an example, according to statistics, on average, more than thousands of urban governance events are reported to the one-network unified management system every day. For government service personnel, on the one hand, although the functions of the one-network unified management event handling system have been relatively complete, and the process from event collection to event settlement has been visualized and transparent, the one-network unified management event handling system based on the traditional IT architecture still faces some problems. For example, the system has a low level of intelligence, and event handling requires manual reporting, handling, and settlement. It is undoubtedly difficult to deal with a large number of different types of events every day and give opinions on the handling of the events; on the other hand, there may be multiple links in the process from event reporting to transfer and settlement, such as the handover between the organizer and the co-organizer, and the feedback of the superiors and subordinates on the handling opinions. Affected by individual subjective factors, if unreasonable handling opinions are provided, the event cannot be handled correctly, resulting in the return and re-handling of the event, which greatly affects the efficiency of event handling. It can be seen that the current traditional mode of manually giving handling opinions will cause high labor costs, as well as the problem of easy errors and slow event handling efficiency. Therefore, designing an artificial intelligence-based one-network event handling method and device, aiming to intelligently handle events in one-network management, has become a key issue that needs to be solved urgently.

[0003] In order to solve the drawbacks of the manual processing opinion mode and improve the accuracy of event handling opinions, some methods have begun to automatically give processing opinions for the one-network unified event handling through artificial intelligence (AI) technology. As an example, a possible method for automatically giving processing opinions for the one-network unified event handling using AI technology is given below.

[0004] At present, the industry generally adopts the combination of Retrieval Augmented Generation (RAG) and large language models to realize actual business scenarios. According to the event information reported by various groups of people such as parties, managers, and witnesses, the most similar one or more pieces of information are recalled from the knowledge base, and a new prompt is rebuilt based on the original question and the recalled data. The large model generates event handling opinions for the event based on the new prompt, and automatically gives a one-network unified event handling opinion method.

[0005] The applicant has found that the method described above has at least the following defects:

[0006] (I) First, since the corpus input into the large model is the most similar one or more retrieved from the knowledge base, the number of event data returned by the retrieval needs to be manually adjusted. If the target event scenario needs to strictly comply with laws and regulations, then the number of event data returned by the retrieval should be as small as possible to limit the scope of the model's answer; if the target event scenario requires some extensible event handling methods, then the number of event data returned by the retrieval should be as large as possible to expand the model's answer as much as possible. In addition, when a large model is used in a specific scenario, its parameters usually need to be fine-tuned, and fine-tuning often consumes more computing power resources. If it is not fine-tuned, even if it is combined with an external knowledge base, the answer effect of the large model will not be very good. Finally, the event handling opinions given by the large model to the target event are random, so there is a great risk of handling the target event solely relying on the event handling opinions given by the model, resulting in the failure of the target event handling and delaying the progress of the event handling.

[0007] (II) In the handling of target events, the big model, as the brain of the agent, gives the opinions on the handling of target events based on historical knowledge. If it encounters some sudden target events that the model has not encountered before, it will fabricate the opinions on the handling of events based on historical knowledge, and such opinions are often non-compliant. Therefore, the adoption of the agent collaboration method can increase the confidence of artificial intelligence in event handling. However, the traditional agent collaboration method cannot guarantee that the agents can work and interact in a consistent and harmonious manner every time. During this period, deadlocks often occur between agents, that is, multiple agents cannot perform their next actions; or livelocks occur between agents, that is, multiple agents continue to work but make no progress. Summary of the invention

[0008] Purpose of the invention: To propose a method, device, equipment and storage medium for intelligent handling of events managed by one network to solve the above-mentioned problems existing in the prior art.

[0009] A first aspect of the present invention provides a method for intelligently handling events managed by one network, comprising the following steps:

[0010] Acquire historical event information and its category, perform vectorization operation on the event information, and extract feature vectors of the event information;

[0011] The feature vector of event information is used as the input of the model, and the category to which the event belongs is used as the output of the model to build a classification model;

[0012] Constructing a conditional probability distribution, selecting an optimal hyperparameter of the classification model using the conditional probability distribution, substituting the optimal hyperparameter into the classification model, and training the classification model;

[0013] Using the trained classification model to classify the target event, so as to obtain the rule information of the category to which the target event belongs;

[0014] Combine the target event information and rule information to construct a new prompt, input the new prompt into the enhanced agent, and the enhanced agent gives a disposal plan corresponding to the target event;

[0015] Complete the handling process of the target event based on the handling plan.

[0016] In a further embodiment of the first aspect, the process of building a classification model specifically includes:

[0017] Based on LightGBM, an initial classification model is constructed for k historical event information f(x 1:k )=[f(x1),f(x2),…,f(x k )], the prior distribution of LightGBM is as follows:

[0018] f(x 1:k )~Normal(μ0(x 1:k ),Σ0(x 1:k ,x 1:k ))

[0019] Where μ0(*) is the mean function, Σ0(*) is the covariance function, and they satisfy:

[0020] μ0(x 1:k )=[μ0(x1),μ0(x2),…,μ0(x k )]

[0021] Σ0(x 1:k ,x 1:k )=[Σ0(x1,x1),…,Σ0(x1,x k ),…,Σ0(x k ,x k )]

[0022] In the formula, μ0(x 1:k ) represents the mean function of k events; μ0(x k ) represents the mean of the kth event; Σ0(x 1:k ,x 1:k ) represents the covariance function of k events; Σ0(x k ,x k ) represents the covariance of the kth event.

[0023] In a further embodiment of the first aspect, the conditional probability distribution of LightGBM is as follows:

[0024]

[0025] In the formula, μ n (x) = E(f(x)|f(x) 1:n )), indicating that in the existing data f(x 1:n ), the expected value of the next target f(x); Indicates that in the existing data f(x 1:n ) under the condition of measuring the next target f(x) and the existing data f(x 1:n )’s variance.

[0026] In a further embodiment of the first aspect, the process of selecting the optimal hyperparameters of the classification model using conditional probability distribution specifically includes:

[0027] Based on the existing optimal target value Find the next optimal target point

[0028]

[0029] In the formula, max x′ Indicates maximizing the current value; μ n+1 (x′) represents the mean function of the next target point.

[0030] The KG method of maximizing the mean function is used to maximize the observed value as much as possible, and the output observed value is the optimal hyperparameter:

[0031]

[0032] x n+1 = argmax n KG n (x)

[0033] In the formula, E n (*) represents the mean function of the nth observation.

[0034] In a further embodiment of the first aspect, the process of training the classification model specifically includes:

[0035] The feature vectors of the collected event information are shuffled and then randomly divided into five parts. Four of the data are used as training sets and the remaining one is used as a validation set to test the classification accuracy of the event classification model.

[0036] The average of the five accuracies obtained on the validation set is taken to evaluate the classification performance of the classification model on the event classification problem.

[0037] The second aspect of the present invention proposes a device for intelligently handling events managed by a single network, which is used to execute the method for intelligently handling events managed by a single network as described in the first aspect. The device includes: an event acquisition unit; a model construction unit; a model training unit; a model output unit; and an enhanced Agent.

[0038] The event acquisition unit is used to acquire historical event information and its category, perform vectorization operation on the event information, and extract feature vectors of the event information;

[0039] The model building unit is used to take the feature vector of the event information as the input of the model and the category to which the event belongs as the output of the model to build a classification model;

[0040] The model training unit is used to construct a conditional probability distribution, select the optimal hyperparameter of the classification model using the conditional probability distribution, substitute the optimal hyperparameter into the classification model, and train the classification model;

[0041] The model output unit classifies the target event using the trained classification model to obtain the rule information of the category to which the target event belongs;

[0042] Combine the target event information and rule information to build a new prompt, input the new prompt into the enhanced agent, and the enhanced agent gives a disposal plan corresponding to the target event; based on the disposal plan, the disposal process of the target event is completed.

[0043] The enhanced Agent includes rule module, memory module, LLM module, tool module and planning module; the rule module stores social rule data set; the memory module is used to store historical behavior trajectories; the LLM module is used to present the reasoning and planning process; the tool module is used to call image generation and API interface; the planning module is used to formulate the disposal plan corresponding to the target event.

[0044] Beneficial effects: The embodiments of the present invention provide a method, device, equipment and storage medium for intelligent handling of events managed by one network. The target event is classified using the Bayesian optimized LightGBM classification model to obtain the rule constraints of the category to which the target event belongs. The rule constraints are combined with the data chain of the event to construct a new prompt input into the enhanced Agent to generate a handling plan for the target event. In the main-co-organized cooperation mode, the enhanced Agent works together as the main and co-organized Agents to formulate and improve the handling plan and plan of the event, and implement the plan. Government service personnel conduct quantitative evaluations on the event handling opinions of each scenario for reference by the enhanced Agent when handling similar events in the future. In summary, the present invention proposes a method and device for handling events managed by one network based on artificial intelligence. By constructing a new Agent architecture-enhanced Agent, a combination of large and small models is adopted, and a multi-objective classification model is built in to improve the accuracy of the model's answer. Compared with traditional IT event handling methods, this method has higher processing efficiency, higher intelligence and better processing effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The figure is a flow chart of handling a network management event in an embodiment of the present invention.

[0046] Figure 2 Schematic diagram of the architecture of a Bayesian optimized LightGBM algorithm in an embodiment of the present invention.

[0047] Figure 3 The figure is a schematic diagram of the architecture of an enhanced Agent in an embodiment of the present invention.

[0048] Figure 4 The present invention is a flowchart of an artificial intelligence-based one-network unified event handling process in an embodiment of the present invention.

[0049] Figure 5 It is a detailed structural diagram of an event handling device for one-network management based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described.

[0051] This embodiment proposes a method and device for intelligently handling events under one network. The data used in this embodiment can be event handling information of projects belonging to urban governance. A Bayesian optimized LightGBM classification model is used for each event handling information to obtain the rule constraints of the category to which the event handling information belongs. The specific steps are as follows:

[0052] S1, receive event information, process the event information, use the LightGBM algorithm to build a classification model, and use the Bayesian optimization algorithm to select the optimal hyperparameters of the LightGBM learner;

[0053] In an optional embodiment of the present invention, step S1 includes the following steps:

[0054] S11. Obtain historical event information and the category to which the event belongs. The historical event information may be text, picture, or a combination of the two, so the event information is vectorized to extract the feature vector of the target event.

[0055] In a possible implementation, a classification model is constructed by taking a feature vector of event information as input of a model and taking the category to which the event belongs as output of the model.

[0056] S12, use Gaussian process regression to model the objective function. If there are k events f(x 1:k )=[f(x1),f(x2),…,f(x k )], construct a prior distribution

[0057] f(x 1:k )~Normal(μ0(x 1:k ),Σ0(x 1:k ,x 1:k )).

[0058] Where μ0 is the mean function, Σ0 is the covariance function, and satisfies

[0059] μ0(x 1:k )=[μ0(x1),μ0(x2),…,μ0(x k )],

[0060] Σ0(x 1:k ,x 1:k )=[Σ0(x1,x1),…,Σ0(x1,x k ),…,Σ0(x k ,x k )].

[0061] Here the mean and variance functions are

[0062]

[0063] Σ0(x,x′)=α0exp(-||xx′|| 2 ).

[0064] Gaussian process regression is a non-parametric model that can automatically adjust its complexity according to the existing data and can capture the uncertainty of the function well. According to Bayes' theorem, Gaussian process regression is selected as the prior estimate of the target function of the event f(x).

[0065] S13. Construct conditional probability distribution:

[0066]

[0067] μ n (x) = Σ0(x,x 1:n )Σ0(x 1:n ,x 1:n ) -1 (f(x 1:n )-μ0(x 1:n ))+μ0(x),

[0068]

[0069] According to Bayes' theorem, the previous observation data is combined with the prior distribution to update the Gaussian process regression. At this time, the conditional probability distribution is a more accurate description of the target function, including the prediction of the function value and uncertainty.

[0070] S14, based on the existing optimal target value Find the next optimal target point

[0071]

[0072] Since there must be noise in event data, the method of maximizing the mean function (Knowledge Gradient, KG) is used to maximize the observed value as much as possible.

[0073]

[0074] x n+1 = argmax n KG n (x).

[0075] But because KG n (x) does not have a complete mathematical expression, so the Monte Carlo method is used to simulate the KG method.

[0076]

[0077] S15. Repeat the above steps and use the conditional probability distribution of the proxy model to select the optimal hyperparameters to obtain the optimal solution.

[0078] S2. Substitute the selected optimal hyperparameters into the LightGBM algorithm training model;

[0079] S3, perform 5-fold cross validation on the training set and evaluate the model prediction performance on the test set;

[0080] First, the collected original data is shuffled and then randomly divided into five parts. Four of the data are used to train the model in turn, and the remaining data is used as a validation set to detect the classification accuracy of the event classification model. Finally, the average of the five accuracies obtained on the validation set is taken to evaluate the classification performance of the model on the event classification problem.

[0081] The enhanced Agent includes a rule module 401 , a memory module 402 , an LLM module 403 , a tool module 404 , an inference module 405 , a planning module 406 , a verification module 407 , and an execution module 408 .

[0082] The rule module 401 stores a set of relevant social rule data such as laws, regulations, and industry rules. In the process of handling the target event, factors such as the scene, nature, and urgency of the event will affect the changes in the processing rules. The rule module 401 is not only a module for storing rules, but also has a built-in Bayesian optimized LightBGM classification model. The rule constraints corresponding to the category to which the target event belongs will restrict the behavior of the reinforcement agent and improve the accuracy of the reinforcement agent's opinions on event handling.

[0083] The memory module 402 plays a vital role in the collaborative cooperation of reinforcement agents. The memory module 402 consists of three components: a rule base, a knowledge base, and a historical behavior trajectory. The rule base is actually a rule stored in the memory module in the form of memory under a specific environment. Unlike the rule module, the relevant rules of the event handling will be stored in the rule base only when the reinforcement agent handles the relevant event. The knowledge base acts as a persistent repository to retain a comprehensive record of the task, including its event information, event handling process, event handling opinions, etc. The historical behavior trajectory, as a temporary area with a fixed length, follows the first-in, first-out method. It stores necessary information such as the latest language-based status, analysis, intentions to teammates, and plans used, while discarding the most outdated and useless data. The historical behavior trajectory component can retrieve only specific parts of the memory module, which promotes the implementation of effective decision-making and planning of reinforcement agents. In general, the memory module significantly improves the ability of reinforcement agents to access relevant information and effectively collaborate with teammates. Learning from past experience and historical data, the memory module enables reinforcement agents to make wise decisions in cooperative tasks.

[0084] LLM module 403 and tool module 404 are similar to traditional agents. LLM module 403 plays the role of strengthening the "brain" of the agent. It not only stores knowledge and memory, but also undertakes functions such as information processing and decision-making, and can present the process of reasoning and planning, and can cope with unknown tasks well. Tool module 404, although LLM has excellent knowledge reserves and professional capabilities, it may also face a series of challenges such as robustness problems and hallucinations when facing specific problems. As an extension of the user's capabilities, tools can provide help in terms of professionalism, factuality, and explainability. For example, you can use a calculator to calculate mathematical problems, use a search engine to search for real-time information, call image generation, existing API interfaces, etc.

[0085] The inference module 405 mainly helps the reinforcement agent to work better with other agents. For the target event, the reinforcement agent infers the intention of the collaborative teammates and gives the event plan of the event. Based on the inference correction mechanism, the reinforcement agent's inferred plan is compared with the actual plan of the collaborative agent, and the reinforcement agent's event plan for the event is further corrected.

[0086] The planning module 406 mainly strengthens the Agent's powerful reasoning ability based on LLM to formulate an event handling plan for the event. Finally, the event handling plan of the event will also be affected by the event planning of the event in the inference module 305.

[0087] The verification module 407 includes a component for plan failure analysis and another component for converting the plan into agent behavior. In the case of plan failure, the plan failure analysis component plays a key role in evaluating any unreasonable or flawed plans generated by the large language model. The plan failure analysis component can analyze the root causes of these deficiencies and provide valuable insights and suggestions for improvement. This can provide a detailed explanation of the problems encountered by the agent, allowing the reinforcement agent to return to the planning stage and re-formulate the plan based on the memorized historical behavior data. In the case of a successful plan, the verification module 407 sends the plan that the reinforcement agent can execute to the next module.

[0088] The execution module 408 includes two components: a controller and an action. The controller is intended to convert the executable plan of the reinforcement agent into low-level operations that can be executed in the environment. The controller will decompose the feasible plan into several execution behaviors, and through the action component, let the reinforcement agent perform specific actions in the cooperative environment.

[0089] The rule module 401, the inference module 405, the planning module 406, the verification module 407 and the execution module 408 are all stored in the memory module as the historical behavior data of the reinforcement agent. These historical behavior data can be used to correct the event handling opinions of the reinforcement agent on the event, thereby reducing the probability of event handling failure.

[0090] This application adopts a large and small model combination method, combining the classification model with the enhanced agent to further improve the accuracy of the model in handling events managed by one network, which specifically includes the following steps:

[0091] S1. Use the Bayesian optimized LightGBM classification model to classify the target event to obtain the rule information of the category to which the target event belongs;

[0092] In one possible implementation, an event may correspond to multiple categories at the same time, but the classification model can only tell us which category it belongs to. In this case, the problem is transformed into multi-label classification instead of single category classification. In this way, the model can predict multiple categories at the same time and extract all the rule information corresponding to the category to which the event belongs.

[0093] S2, combining the target event information and the rule information, constructing a new prompt, inputting it into the enhanced agent, and the enhanced agent gives a disposal plan for the event according to the target event;

[0094] In a possible implementation, the data link of the target event may be text data, image data, or both text data and image data. Before the data link is input into the enhanced Agent, it needs to be vectorized.

[0095] When the data link of the event is text data, it needs to be converted into a feature vector corresponding to the text data. When the data link of the event is image data, it needs to be converted into a feature vector corresponding to the image data. The data link of the event may also include both text data and image data. For example, when a grid worker reports information about a city event, he uploads both text content and pictures. In this case, the feature vector may include both the feature vector corresponding to the text data and the feature vector corresponding to the image data. In addition, the feature vector may also be a feature vector obtained by weighting or splicing the feature vector corresponding to the text data and the feature vector corresponding to the image data.

[0096] S3. In a cooperative environment, adopt the host-co-organizer cooperation model, strengthen the cooperation between the Agent as the host and the co-organizer, give a preliminary handling plan for the incident and a preliminary handling plan for the incident, improve the handling plan for the incident, further improve the handling plan for the incident based on the handling plan for the incident, implement the handling plan for the incident, and give handling opinions for the incident;

[0097] Furthermore, in a possible implementation, the reinforcement agent modifies the collaborative agent's disposal plan based on its disposal plan to obtain a comprehensive disposal plan. As the organizer of the disposal of the event, the reinforcement agent first gives the disposal plan of the event, and gives supplementary opinions and evaluations through the co-sponsor agent (cooperative teammate), and gradually corrects its own disposal plan for the event; then, in the planning module, the reinforcement agent formulates the final disposal plan for the event based on the preliminary disposal plan of the event and the disposal plan of the event; the final event disposal plan is verified by the verification mechanism, and the verifier repeatedly evaluates and adjusts the plan until the most reasonable plan is determined. If the plan is feasible, the controller decomposes the plan into several executable behaviors and executes them in the cooperative environment; if the plan is not feasible, the co-sponsor agent further modifies the reinforcement agent's plan, provides detailed explanations to help it enter the re-planning cycle, and starts the recalibration process; finally, the controller mechanism selects an optimal plan, and the controller decomposes it into low-level actions, allowing the reinforcement agent to effectively interact with the task or environment. The controller may be a rule-based method, a hierarchical method based on RL, or a machine learning algorithm, such as a path planning algorithm, a simulated annealing algorithm, etc.;

[0098] In a possible implementation, given an event handling of a city under unified management by one network, both the strengthening agent and the co-sponsoring agent need to provide an event handling plan for the event, wherein the strengthening agent not only needs to independently formulate a plan, but also needs to adjust and optimize its own plan according to the plan of the co-sponsoring agent.

[0099] For example, suppose that we need to solve the problem of handling event B in city A under the unified management of one network. First, the data of event B in city A is vectorized and used as the input of the enhanced agent. Secondly, the enhanced agent generates a preliminary event B handling plan C1 based on these data. At the same time, the co-sponsoring agent will also provide a handling plan C2 for event B. Then, the enhanced agent modifies the handling plan C1 according to the handling plan C2 to obtain a comprehensive version of the handling plan C. Then, the verifier verifies the handling plan C to ensure that it meets the relevant handling specifications. If the verification passes, the controller decomposes plan C into several specific execution steps S1, S2,…, S n Finally, the enhanced agent gives a complete event handling opinion for event B.

[0100] S4. The government service personnel conduct a quantitative evaluation of the handling opinions of the incident given by the reinforcement agent. The results of the quantitative evaluation are used in the reinforcement learning stage of the large language model (LLM). The results are recorded in the memory module of the reinforcement agent as historical behavior trajectory data for reference by the reinforcement agent when handling similar incidents in the future.

[0101] As the whole process is continuously iterated, the effect of strengthening the collaborative cooperation of agents is gradually optimized. This implementation method combines the actual application scenarios of one-network management, fully utilizes the powerful logical reasoning and planning capabilities of the large model, and the automatic planning method. The whole process does not require additional model fine-tuning. It only requires manual assessment of whether the event handling opinions meet the expected requirements and scoring of the event handling opinions, so that a good handling effect can be achieved in the event handling of one-network management.

[0102] In addition, this embodiment also discloses an electronic device, which includes a processor, a memory, a communication interface and a communication bus. The processor, the memory and the communication interface communicate with each other through the communication bus. The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute all the process steps of the intelligent handling method for a network management event disclosed in the above embodiment, which will not be repeated here.

[0103] The electronic device may also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may communicate with one or more devices that enable a user to interact with the electronic device, and / or may communicate with any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication may be performed through an input / output (I / O) interface. In addition, the electronic device may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through a network adapter. The network adapter communicates with other modules of the electronic device through a bus. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0104] It should be noted that the above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0105] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0106] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for intelligently handling events managed by one network, characterized in that: The steps include: Acquire historical event information and its category, perform vectorization operation on the event information, and extract feature vectors of the event information; The feature vector of event information is used as the input of the model, and the category to which the event belongs is used as the output of the model to build a classification model; Constructing a conditional probability distribution, selecting an optimal hyperparameter of the classification model using the conditional probability distribution, substituting the optimal hyperparameter into the classification model, and training the classification model; Using the trained classification model to classify the target event, so as to obtain the rule information of the category to which the target event belongs; Combine the target event information and rule information to construct a new prompt, input the new prompt into the enhanced agent, and the enhanced agent gives a disposal plan corresponding to the target event; Complete the handling process of the target event based on the handling plan.

2. The one-network-managed event intelligent handling method according to claim 1 is characterized in that: The construction of the classification model specifically includes: Based on LightGBM, the initial classification model is constructed. For k historical event information f(X1:k)=[f(x1), f(x2), ..., f(xk)], the prior distribution of LightGBM is as follows: f(x1:k)~Normal(μ0(x1:k),∑0(x1:k,x1:k)) Where μ0(*) is the mean function, ∑0(*) is the covariance function, and they satisfy: μ0(x1:k)=[μ0(x1),μ0(x2),…,μ0(x k )] ∑0(x1: k， x1:k)=[∑0(x1,x1),…,∑0(x1,x k ),…,∑0(x k ,x k )] Where μ0(X1:k) represents the mean function of k events; μ0( xk ) represents the mean of the k-th event; ∑0(x1:k, x1:k) represents the covariance function of the k-th event; ∑0(xk, xk) represents the covariance of the k-th event.

3. The one-network-managed event intelligent handling method according to claim 2 is characterized in that: The conditional probability distribution of LightGBM is as follows: In the formula, μ n (x) = E(f(x)|f(X 1：n )), indicating that in the existing data f(x 1：n ), the expected value of the next target f(x); Indicates that in the existing data f(x 1：n ) under the condition of measuring the next target f(x) and the existing data f(x 1：n )’s variance.

4. The one-network-managed event intelligent handling method according to claim 2 is characterized in that: The optimal hyperparameters of the classification model are selected using conditional probability distribution, including: Based on the existing optimal target value Find the next optimal target point In the formula, max x ′ represents maximizing the current value; μ n+ 1(x′) represents the mean function of the next target point; The KG method of maximizing the mean function is used to maximize the observed value as much as possible, and the output observed value is the optimal hyperparameter: x n+1 =argmax n KG n (x) In the formula, E n (*) represents the mean function of the nth observation.

5. The one-network-managed event intelligent handling method according to claim 4 is characterized in that: The training classification model specifically includes: The feature vectors of the collected event information are shuffled and then randomly divided into five parts. Four of the data are used as training sets and the remaining one is used as a validation set to test the classification accuracy of the event classification model. The average of the five accuracies obtained on the validation set is taken to evaluate the classification performance of the classification model on the event classification problem.

6. A one-network-wide-management-event intelligent handling device, used to execute the one-network-wide-management-event intelligent handling method as claimed in any one of claims 1 to 5; characterized in that: The device includes: An event acquisition unit; the event acquisition unit is used to acquire historical event information and its category, perform vectorization operation on the event information, and extract a feature vector of the event information; A model building unit; the model building unit is used to take the feature vector of the event information as the input of the model and the category to which the event belongs as the output of the model to build a classification model; Model training unit; the model training unit is used to construct a conditional probability distribution, select the optimal hyperparameters of the classification model using the conditional probability distribution, substitute the optimal hyperparameters into the classification model, and train the classification model; Model output unit; the model output unit classifies the target event using the trained classification model to obtain rule information of the category to which the target event belongs; Strengthen the Agent; combine the target event information and rule information to build a new prompt, input the new prompt into the strengthened Agent, and the strengthened Agent gives a disposal plan corresponding to the target event; complete the disposal process of the target event based on the disposal plan.

7. The one-network-managed event intelligent handling device according to claim 6, characterized in that: The enhanced Agent includes a rule module, a memory module, an LLM module, a tool module, and a planning module; The rule module stores a social rule data set; The memory module is used to store historical behavior trajectories; The LLM module is used to present the process of reasoning and planning; The tool module is used to call image generation and API interface; The planning module is used to formulate a treatment plan corresponding to the target event.

8. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for intelligent handling of events managed by one network as described in any one of claims 1 to 5 is implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores at least one executable instruction. When the executable instruction is executed on the electronic device, the electronic device executes the one-network-managed event intelligent handling method according to any one of claims 1 to 5.

Images