A semantic recognition and automatic execution method, system, device and medium for audit rule configuration

Abstract

The application relates to the technical field of power system automation operation and maintenance, and discloses a semantic recognition and automatic execution method, system, equipment and medium for audit rule configuration, which comprises the following steps: receiving an input natural language rule, performing semantic analysis through a knowledge graph and a template library, and generating a structured intermediate representation; then, the structured intermediate representation is converted into a rule intelligent agent, logical conflict detection and elimination are performed, simulation evaluation is performed in a digital twin sandbox, and the rule intelligent agent is deployed to a production environment through a post-deployment, parameters are optimized by using reinforcement learning, and finally, incremental learning and continuous optimization of the knowledge graph and the analysis model are realized through monitoring execution feedback and manual intervention. The method realizes intelligent closed-loop management of the whole life cycle of the rule, and greatly improves the reliability, self-adaptability and execution efficiency of rule processing.

Description

Technical Field

[0001] This invention relates to the field of power system automation operation and maintenance technology, and in particular to a semantic recognition and automatic execution method, system, device and medium for audit rule configuration. Background Technology

[0002] With the deepening of smart grid construction, the operation and maintenance audit rules of power systems are becoming increasingly complex and sophisticated. Traditionally, these rules are mostly described in natural language by domain experts and rely on manually written scripts or configuration logic to achieve automated execution. In recent years, Natural Language Processing (NLP) technology, especially semantic parsing methods based on pre-trained models (such as BERT and GPT), has made significant progress in rule generation in general domains, capable of converting simple natural language instructions into preliminary structured queries or code snippets. Simultaneously, research on rule conflict detection, parameter optimization, and testing verification has also developed, leading to auxiliary technologies such as static conflict checking based on logical reasoning, parameter tuning based on A / B testing, and simulation-based verification. These technological advancements have laid a preliminary foundation for building automated rule management systems.

[0003] However, when existing technologies are applied to specialized fields with high reliability requirements, such as the power industry, a series of limitations still need to be addressed. First, general-purpose NLP models lack prior knowledge in the power sector and have limited understanding of technical terms such as "three-phase imbalance rate" and "distribution transformer overload," as well as vague expressions like "close observation required." This results in low accuracy and high ambiguity in the analysis results, heavily relying on manual correction. Second, existing conflict detection largely depends on static logical rule bases or manual checks, making it difficult to dynamically adapt to the complex spatiotemporal conditions and contextual scenarios in power grid operation (such as the instantaneous impact of thunderstorms on alarm rules), and failing to efficiently discover and resolve deep logical mutual exclusions between rules. Third, many fuzzy parameters in the rules (such as "significant fluctuations") rely on subjective settings based on expert experience, lacking automated and refined optimization mechanisms based on historical feedback data, leading to low rule execution efficiency and insufficient adaptability. Finally, new rules lack security verification in near-real-world environments before deployment. Traditional testing methods cannot simulate large-scale, diverse user electricity consumption behaviors, potentially causing system avalanches or numerous false alarms after direct rule deployment, resulting in extremely high operational risks. These shortcomings together constitute the core bottleneck in the implementation of automated management of power audit rules. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a semantic recognition and automatic execution method for audit rule configuration, which can solve the comprehensive problem that existing technologies cannot accurately, automatically, and securely convert complex natural language audit rules in the power field into executable code and ensure their continuous, efficient, and reliable operation.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a semantic recognition and automatic execution method for audit rule configuration, comprising: receiving input natural language rules; performing semantic parsing using a pre-built domain knowledge graph and rule template library to generate a structured intermediate rule representation; converting the structured intermediate rule representation into a rule agent, and performing logical conflict detection and resolution with existing rule agents in the system based on a predefined conflict knowledge base to obtain conflict-free rules; simulating the conflict-free rules in a digital twin sandbox environment and generating a simulation evaluation report; deploying the simulated rules to a production environment, and optimizing their execution parameters based on a reinforcement learning model and historical feedback data of rule execution; monitoring rule execution feedback and manual correction behavior in the production environment, and updating the domain knowledge graph and semantic parsing model based on an incremental learning algorithm.

[0007] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the semantic parsing includes identifying domain entities and business relationships from natural language rules and linking them with a domain knowledge graph.

[0008] Match the linked entities and relationships with templates in the rule template library;

[0009] The successfully matched entities and relationships are populated into the corresponding rule templates to generate the structured intermediate representation of the rules.

[0010] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the step of performing logical conflict detection and resolution includes creating a new rule agent for the structured intermediate representation of rules;

[0011] Enables the new rule agent to communicate and negotiate with existing rule agents that represent deployed rules in the system;

[0012] During the negotiation process, the execution of rules under different scenarios is simulated based on a conflict knowledge base to detect logical conflicts.

[0013] When a conflict is detected, add conflict exclusion conditions to the rule or prompt manual intervention according to the predefined resolution strategy.

[0014] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the simulation operation in the digital twin sandbox environment includes constructing virtual users and corresponding electricity consumption data streams using historical data and generation technology.

[0015] The conflict-free rules are applied to the virtual user data stream to simulate execution;

[0016] Record all alarm events triggered during rule execution, and generate a simulation evaluation report containing the estimated number of alarms and false alarm rate based on the alarm events and the preset abnormal states of virtual users.

[0017] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the step of optimizing the execution parameters based on the reinforcement learning model includes defining configurable execution parameters for the rule as the action space of the reinforcement learning model.

[0018] The current parameter configuration and execution environment state of the rule are used as the state of the reinforcement learning model;

[0019] The reward signal of the reinforcement learning model is calculated based on whether the result of rule execution captures a real anomaly or generates a false alarm.

[0020] Reinforcement learning models learn and output optimized execution parameter configurations based on historical states, actions, and reward signals.

[0021] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the step of updating the domain knowledge graph and semantic parsing model based on incremental learning algorithm includes collecting correction records of rule parsing results and rule execution feedback in the production environment to form a sample set to be learned;

[0022] The elastic weight consolidation algorithm is adopted to incrementally train the semantic parsing model using the sample set to be learned, while retaining the memory of existing knowledge.

[0023] Based on the semantic parsing model trained incrementally and the sample set to be learned, the entities, relations and templates in the rule template library of the domain knowledge graph are expanded and updated.

[0024] As a preferred embodiment of the semantic recognition and automatic execution method for audit rule configuration described in this invention, the generation of the simulation evaluation report includes starting a rule refinement loop when the simulation evaluation report fails verification.

[0025] The threshold for passing the simulation evaluation report is defined as follows: the estimated false alarm rate is lower than the set false alarm rate threshold, and the expected total number of alarms per unit time is lower than the system processing capacity threshold.

[0026] If any one of the conditions is not met, the test is deemed unsuccessful.

[0027] The system automatically analyzes the combination of conditions or parameter settings that cause alarms or false alarms, and attaches the analysis results and corresponding virtual sample data as structured feedback information to the original rules, and returns them to the semantic parsing.

[0028] This invention provides a semantic recognition and automatic execution system for audit rule configuration.

[0029] As a preferred embodiment of the semantic recognition and automatic execution system for audit rule configuration described in this invention, the system comprises: a rule intelligent parsing module, a conflict collaborative resolution module, a sandbox simulation verification module, a parameter self-optimization module, and a continuous evolutionary learning module.

[0030] The rule intelligent parsing module is responsible for receiving natural language rules input by the user and converting them into a precise structured representation;

[0031] The conflict resolution module is responsible for ensuring the consistency between the new rules and the existing rule set. Based on the multi-agent system (MAS) architecture, it simulates the interaction between the new rules and the existing rules.

[0032] The sandbox simulation verification module simulates the execution process of rules on a massive number of virtual users in a constructed high-fidelity digital twin environment.

[0033] The parameter self-optimization module operates on rules already deployed in the production environment and is responsible for fine-tuning their execution parameters.

[0034] The continuous evolution learning module collects rule execution feedback and manual correction records in the production environment in real time, and automatically updates the domain knowledge graph and rule template library.

[0035] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a semantic recognition and automatic execution method oriented to audit rule configuration.

[0036] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a semantic recognition and automatic execution method for audit rule configuration.

[0037] The beneficial effects of this invention are as follows: By integrating domain knowledge graphs and rule templates, ambiguity in natural language is effectively eliminated, achieving accurate understanding and structured transformation of technical terms and complex logic. Utilizing a multi-agent negotiation mechanism, automated and intelligent dynamic detection and resolution of logical conflicts between rules are achieved, ensuring the overall consistency of the rule set. By introducing a reinforcement learning mechanism, fuzzy parameters are automatically optimized based on historical execution feedback, significantly improving the execution efficiency and accuracy of rules. A high-fidelity simulation environment is constructed using digital twin technology, enabling safety verification and effect evaluation before new rules are deployed, greatly reducing the risk of accidents in the production environment. Through an incremental learning mechanism, the system can continuously learn new knowledge from human feedback and business changes, preventing model rust and adapting to the dynamic development of the power industry. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of a semantic recognition and automatic execution method for audit rule configuration provided in one embodiment of the present invention. Detailed Implementation

[0040] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0041] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a semantic recognition and automatic execution method for audit rule configuration, including:

[0042] S1: Receives input natural language rules, performs semantic parsing using pre-built domain knowledge graphs and rule template libraries, and generates a structured intermediate representation of the rules.

[0043] S2: Transform the structured intermediate representation of rules into a rule agent, and perform logical conflict detection and resolution with the existing rule agents in the system based on a predefined conflict knowledge base to obtain conflict-free rules.

[0044] S3: Simulate and run conflict-free rules in a digital twin sandbox environment and generate a simulation evaluation report.

[0045] S4: Deploy the simulated rules to the production environment and optimize their execution parameters based on the historical feedback data of rule execution using a reinforcement learning model.

[0046] S5: Monitors rule execution feedback and manual correction behavior in the production environment, and updates the domain knowledge graph and semantic parsing model based on incremental learning algorithms.

[0047] Example 2 is an embodiment of the present invention. Based on the above embodiment, a semantic recognition and automatic execution method for audit rule configuration is provided.

[0048] Furthermore, in this embodiment, step S1 receives the input natural language rules, performs semantic parsing using a pre-built domain knowledge graph and rule template library, and generates a structured intermediate representation of the rules. Specific steps include S11-S13:

[0049] S11: Use a Transformer-based dual encoder model for semantic understanding;

[0050] The first encoder is a domain entity encoder. Its input is natural language rule text, and its output layer is linked to the entity set in the power domain knowledge graph to identify professional term entities in the text, such as "three-phase imbalance rate" and "distribution transformer overload".

[0051] The second encoder is a logical relation encoder, which takes as input the text vector representation of linked entities and outputs a relation vector for similarity matching with a predefined rule template library.

[0052] S12: The rule template library is abstracted from historical expert rules. The template form is a logical framework of "IF[Condition]THEN[Action]", in which the conditions and actions include type constraints.

[0053] S13: The identified entities are constrained by type and filled into the condition and action slots of the best matching logical relation template, thereby generating an unambiguous and executable AST.

[0054] Furthermore, in this embodiment, step S2 transforms the structured intermediate rule representation into a rule agent, and performs logical conflict detection and resolution with existing rule agents in the system based on a predefined conflict knowledge base to obtain conflict-free rules. Specific steps include:

[0055] S21: The rule-based agent contains a set of logical predicates for its rule AST. The negotiation process is the exchange of logical predicates between the agents.

[0056] S22: Conflict detection is based on first-order logic reasoning to check whether there are predicate overlaps but mutually exclusive conclusions under different environmental states. The conflict knowledge base defines common mutually exclusive action pairs, such as "immediate alarm" and "delayed alarm".

[0057] When a conflict is detected, the resolution strategy automatically adds an exclusive clause to the condition node of the AST of the newly effective rule, with the logical form: [New Rule Condition] AND NOT([Conflicting Rule Condition]).

[0058] Furthermore, in this embodiment of the application, step S3 simulates the conflict-free rules in a digital twin sandbox environment and generates a simulation evaluation report. Specific steps include S31-S34:

[0059] S31: Constructed using an improved Conditional Time Series Generative Adversarial Network (CT-GAN), which contains a generator G and a discriminator D;

[0060] The generator G employs an encoder-decoder structure, where the encoder is a bidirectional gated recurrent unit network used to capture the temporal dependencies of electricity consumption data; the decoder is an attention-based gated recurrent unit network used to focus on key time steps when generating data; the generator's input consists of a random noise vector z, a condition vector c (including user type, region, and season information), and a randomly initialized initial state vector h0, and the output is simulated multidimensional electricity consumption time series data:

[0061] X fake =x1,x2,x3,…,x T

[0062] Discriminator D employs a structure combining a one-dimensional convolutional neural network and fully connected layers, with its input being real electricity consumption data X. real Or generate data X fake And with X real or X fake The corresponding condition vector c has a multi-task discrimination: first, to determine the authenticity of the data; second, to reconstruct the input condition vector c', and to ensure a strong correlation between the generated data and the conditions by comparing the differences between c and c'.

[0063] S32: The loss function of CT-GAN is a weighted sum of the improved Wasserstein distance and the reconstruction loss, and its objective function is as follows:

[0064]

[0065] in, The total loss of the discriminator, The total loss of the generator. As the expected value, It is a random interpolation of real data and generated data, λ gp These are the weight coefficients of the gradient penalty term. λrec is the reconstruction loss function of the conditional vector, and λrec is the weight coefficient of the reconstruction loss.

[0066] After stabilization through adversarial training, the generator can produce high-fidelity, diverse power consumption data that meets the specified user conditions, which can be used for simulation testing of the rules under high concurrency and extreme scenarios.

[0067] S33: Define the passing threshold for the simulation evaluation report as follows: the estimated false alarm rate FPRest is lower than the threshold δfpr, and the expected total number of alarms N per unit time is... alarm Below the system processing capacity threshold δ cap If FPRest > δfpr or N alarm >δ cap If so, it is judged as failing;

[0068] The system automatically analyzes which combinations of conditions or parameter settings led to a large number of alarms or false alarms, and attaches the analysis results (e.g., "the rule is prone to triggering a large number of alarms under the conditions of [summer][industrial and commercial users][midday hours]") and the corresponding virtual sample data as structured feedback information to the original rule, and returns them to step S1.

[0069] S34: Engineers can refer to this feedback information to modify the original natural language rule description (such as adding restrictions: "During the midday peak electricity consumption of industrial and commercial users in summer, planned maintenance situations need to be excluded"). The system can also use the returned virtual sample data to pre-annotate the fuzzy terms in the rules before parsing, assisting the semantic parsing model to extract entities and relationships more accurately, thus starting a new round of "parsing-conflict detection-simulation" cycle until the rules are verified through simulation.

[0070] Furthermore, in this embodiment, step S4 deploys the simulated rules to the production environment and optimizes their execution parameters based on the reinforcement learning model and historical feedback data of rule execution. Specific steps include:

[0071] For each rule, a proximal policy optimization (PPO) reinforcement learning agent is defined. The model contains an Actor network and a Critic network, which share the underlying feature extraction layer.

[0072] state space st : A composite vector, including the parameters currently configured for the rule (threshold θ) t Sampling interval τ t Statistical characteristics of electricity consumption data within the recent time window T (mean μ) t Standard deviation σ t skewness ξ t kurtosis κ t ), and environmental context features (weather conditions) t Holiday signs h t );

[0073] Action space a t Defined as a continuous adjustment amount to the rule parameters, such as Δθ. t ,Δτ t The action output is parameterized by the Actor network and follows a Gaussian distribution. Where the mean μ(s) t The standard deviation σ is a learnable parameter, output by the network.

[0074] Reward function R(s) t ,a t The calculation formula is as follows:

[0075]

[0076] Where TP, FP, TN, and FN are the true positive, false positive, true negative, and false negative counts of the rule within the evaluation time window T, respectively, obtained by comparing manually confirmed work orders with rule alarm records; α, β, γ, and η are weighting coefficients used to balance recall, precision, parameter variation, and computational resource consumption; Δθ t , Δθ t-1 Δτ represents the continuous adjustment of the threshold at the current and previous time points. t , Δτ t-1 The sampling interval between the current and previous times is continuously adjusted; CPU cost The rule uses new parameters (θ) t ,τ t The CPU time consumed during execution. max It is the maximum CPU time limit allocated by the system for this rule.

[0077] The reward function not only evaluates the effectiveness of the rule (recall and precision), but also optimizes its stability (penalty parameter drastic changes) and resource efficiency; the PPO algorithm updates the policy by maximizing an alternative objective function with a pruning mechanism.

[0078] Furthermore, in this embodiment, step S5 monitors rule execution feedback and manual correction behavior in the production environment, and updates the domain knowledge graph and semantic parsing model based on incremental learning algorithms. Specific steps include S51-S52:

[0079] Semantic parsing models are variants of pre-trained language models based on Transformer;

[0080] S51: The incremental learning process uses the Online Elastic Weight Consolidation (Online-EWC) algorithm, which restricts the direction of change of important parameters to the quadratic approximation range of the optimal solution of the old task through regularization terms;

[0081] After obtaining a new batch of data At that time, loss function Defined as the sum of the new task loss and all old task constraints:

[0082]

[0083] in, The model is on new data The standard loss function on, λ is the value of the k-th parameter after training for task i. ewc It is the strength hyperparameter of the regularization term. Let the k-th diagonal element of the Fisher information matrix for task i be used to calculate its approximation online.

[0084]

[0085] This value measures the parameter θ k Importance of task i; N i Let x be the total number of training samples for task i. n Let y be the nth input sample for task i. n Corresponding to x n The target value, θ *(i) The optimal parameters that the model finally determines after completing the training of task i are fixed values.

[0086] S52: By minimizing this loss function, when learning new tasks (such as identifying new devices like "photovoltaic inverters" and their related rules), the model's key parameters are constrained to near the optimal solutions for older tasks (such as identifying rules for "transformers" and "circuit breakers"), thus effectively mitigating catastrophic forgetting. The fine-tuned model output is used to expand and update entities and relationships in the power sector knowledge graph.

[0087] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that:

[0088] This embodiment also provides a semantic recognition and automatic execution system for audit rule configuration, including:

[0089] The module includes a rule intelligent parsing module, a conflict collaborative resolution module, a sandbox simulation verification module, a parameter self-optimization module, and a continuous evolutionary learning module.

[0090] The rule intelligent parsing module is responsible for receiving natural language rules input by the user and converting them into a precise structured representation;

[0091] The conflict resolution module is responsible for ensuring the consistency between the new rules and the existing rule set. Based on the multi-agent system (MAS) architecture, it simulates the interaction between the new rules and the existing rules.

[0092] The sandbox simulation verification module simulates the execution process of rules on a massive number of virtual users within a constructed high-fidelity digital twin environment.

[0093] The parameter self-optimization module operates on rules already deployed in the production environment, and is responsible for fine-tuning their execution parameters.

[0094] The continuous evolution learning module collects rule execution feedback and manual correction records in the production environment in real time, and automatically updates the domain knowledge graph and rule template library.

[0095] This embodiment also provides an electronic device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize a semantic recognition and automatic execution method for audit rule configuration as proposed in the above embodiment.

[0096] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a semantic recognition and automatic execution method for audit rule configuration as proposed in the above embodiments.

[0097] The storage medium proposed in this embodiment and the semantic recognition and automatic execution method for implementing audit rule configuration proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0098] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for semantic recognition and automatic execution of audit rule configuration, characterized in that: include, It receives input natural language rules, performs semantic parsing using a pre-built domain knowledge graph and rule template library, and generates a structured intermediate representation of the rules. The structured intermediate representation of rules is transformed into a rule agent, and logical conflicts are detected and resolved with the existing rule agents in the system based on a predefined conflict knowledge base to obtain conflict-free rules. The conflict-free rules are simulated and run in a digital twin sandbox environment, and a simulation evaluation report is generated. The simulation operation in the digital twin sandbox environment includes constructing virtual users and corresponding electricity consumption data streams using historical data and generation technology; The conflict-free rules are applied to the virtual user data stream to simulate execution; Record all alarm events triggered during rule execution, and generate a simulation evaluation report containing the estimated number of alarms and false alarm rate based on the alarm events and the preset abnormal states of virtual users; The improved Conditional Time Series Generative Adversarial Network (CT-GAN) is constructed, which contains a generator G and a discriminator D. After stabilization through adversarial training, the generator can produce high-fidelity and diverse power consumption data that meets the specified user conditions, which can be used to conduct simulation tests on the rules under high concurrency and extreme scenarios. The passing threshold of the simulation evaluation report is defined as: estimated false alarm rate Below the threshold , and the total number of alarms per unit time Below the system processing capacity threshold ; if Or , it is determined as not passing; The system automatically analyzes the combination of conditions or parameter settings that cause a large number of alarms or false alarms, and attaches the analysis results and the corresponding virtual sample data as structured feedback information to the original rules, and returns them to the semantic parsing. Engineers can modify the original natural language rule description based on the feedback information. The system can use the returned virtual sample data to pre-annotate the fuzzy terms in the rule before parsing, assist the semantic parsing model in extracting entities and relations, and start a new cycle until the rule is verified by simulation. The rules, which are simulated, are deployed to the production environment, and their execution parameters are optimized based on the historical feedback data of the rule execution, using a reinforcement learning model. The optimization of the execution parameters based on the reinforcement learning model includes defining configurable execution parameters for the rules as the action space of the reinforcement learning model; The current parameter configuration and execution environment state of the rule are used as the state of the reinforcement learning model; The reward signal of the reinforcement learning model is calculated based on whether the result of rule execution captures a real anomaly or generates a false alarm. Reinforcement learning models learn and output optimized execution parameter configurations based on historical states, actions, and reward signals; Monitor rule execution feedback and manual correction behavior in the production environment, and update the domain knowledge graph and semantic parsing model based on incremental learning algorithms; The method of updating the domain knowledge graph and semantic parsing model based on incremental learning algorithm includes collecting correction records of rule parsing results and rule execution feedback in the production environment to form a sample set to be learned; The elastic weight consolidation algorithm is adopted to incrementally train the semantic parsing model using the sample set to be learned, while retaining the memory of existing knowledge. Based on the semantic parsing model trained incrementally and the sample set to be learned, the entities, relations and templates in the rule template library of the domain knowledge graph are expanded and updated.

2. The method of claim 1, wherein the method is configured for audit rules. The semantic parsing includes identifying domain entities and business relationships from natural language rules and linking them with the domain knowledge graph; Match the linked entities and relationships with templates in the rule template library; The successfully matched entities and relationships are populated into the corresponding rule templates to generate the structured intermediate representation of the rules.

3. The semantic recognition and automatic execution method for audit rule configuration as described in claim 2, characterized in that: The logical conflict detection and resolution includes creating a new rule agent for the structured rule intermediate representation; Enables the new rule agent to communicate and negotiate with existing rule agents that represent deployed rules in the system; During the negotiation process, the execution of rules under different scenarios is simulated based on a conflict knowledge base to detect logical conflicts. When a conflict is detected, add conflict exclusion conditions to the rule or prompt manual intervention according to the predefined resolution strategy.

4. The semantic recognition and automatic execution method for audit rule configuration as described in claim 3, characterized in that: The generation of the simulation evaluation report includes initiating a rule refinement loop if the simulation evaluation report fails verification. The threshold for passing the simulation evaluation report is defined as follows: the estimated false alarm rate is lower than the set false alarm rate threshold, and the expected total number of alarms per unit time is lower than the system processing capacity threshold. If any one of the conditions is not met, the test is deemed unsuccessful. The system automatically analyzes the combination of conditions or parameter settings that cause alarms or false alarms, and attaches the analysis results and corresponding virtual sample data as structured feedback information to the original rules, and returns them to the semantic parsing.

5. A semantic recognition and automatic execution system for audit rule configuration, employing the semantic recognition and automatic execution method for audit rule configuration as described in any one of claims 1 to 4, characterized in that, include: The module includes a rule intelligent parsing module, a conflict collaborative resolution module, a sandbox simulation verification module, a parameter self-optimization module, and a continuous evolutionary learning module. The rule intelligent parsing module is responsible for receiving natural language rules input by the user and converting them into a precise structured representation; The conflict resolution module is responsible for ensuring the consistency between the new rules and the existing rule set. Based on the multi-agent system (MAS) architecture, it simulates the interaction between the new rules and the existing rules. The sandbox simulation verification module simulates the execution process of rules on a massive number of virtual users in a constructed high-fidelity digital twin environment. The parameter self-optimization module operates on rules already deployed in the production environment and is responsible for fine-tuning their execution parameters. The continuous evolution learning module collects rule execution feedback and manual correction records in the production environment in real time, and automatically updates the domain knowledge graph and rule template library.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the semantic recognition and automatic execution method for audit rule configuration as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the semantic recognition and automatic execution method for audit rule configuration as described in any one of claims 1 to 4.

Images