Deep learning model-based knowledge learning agent training method

Abstract

The invention discloses a deep learning model-based knowledge learning agent training method, and relates to the technical field of auditing, and the method comprises the steps: carrying out the optimal disassembly task classification through a multi-model fusion knowledge classification method, and obtaining defined auditing knowledge and undefined auditing knowledge; optimal disassembly task classification is carried out by using a multi-model fusion knowledge classification method, and undefined audit knowledge is analyzed through an adaptive improved iteration mechanism, a fusion NLP algorithm and a deep learning audit agent to generate a question answer; the data security in the use process of the audit knowledge learning agent is controlled through a dynamic authority control and question and answer tracing mechanism; when facing large-scale and multi-level auditing tasks, the system can quickly position the core of a problem, redundant operation is reduced, and the response speed is increased.

Description

Technical Field

[0001] The present invention relates to the field of auditing technologies, and particularly to a training method for a knowledge learning agent based on a deep learning model. Background Art

[0002] With the rapid development of artificial intelligence technologies, intelligent auditing systems have gradually become the focus of attention in the power industry. Although existing intelligent auditing tools have made certain progress in data processing and automation, they still face many challenges and drawbacks in practical applications; First of all, the knowledge coverage rate of intelligent auditing systems is relatively low. Existing auditing knowledge bases mainly rely on historical data and rule bases and cannot effectively cover emerging auditing scenarios and complex technical problems. Secondly, the knowledge reserve of existing intelligent auditing systems is insufficient. Although the systems can store and process a large amount of auditing data, when facing complex cross-domain problems, existing knowledge bases often lack sufficient depth and breadth and are difficult to provide comprehensive support. In addition, there are also obvious deficiencies in the knowledge reuse of intelligent auditing systems. Existing knowledge management mechanisms are mostly static storage and cannot be dynamically adjusted and optimized according to new auditing requirements, resulting in a large amount of auditing experience being idle after completing one task and unable to provide continuous reference and reference for subsequent work; Therefore, it is of great significance to design an efficient knowledge learning agent. Summary of the Invention

[0003] The purpose of the present invention is to provide a training method for a knowledge learning agent based on a deep learning model to solve the deficiencies in the background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A training method for a knowledge learning agent based on a deep learning model, including: Breaking down an auditing problem through an intelligent strategy to obtain an optimal breakdown task, and classifying the optimal breakdown task through a knowledge classification method of multi-model fusion to obtain defined auditing knowledge and undefined auditing knowledge; Analyzing the undefined auditing knowledge through an adaptive improvement iteration mechanism, an auditing agent integrating NLP algorithms and deep learning to generate problem answers; Using a dynamic permission control and question-and-answer traceability mechanism to control data security during the use of the auditing knowledge learning agent.

[0005] In a preferred embodiment, the step of breaking down an auditing problem through an intelligent strategy to obtain an optimal breakdown task is as follows: Defining the level of the auditing problem as M; Using a one-to-one splitting strategy to divide the auditing problem into M parts, with each part being a sub-task; Each subtask is regarded as a disassembly path, the pheromone of each subtask is defined, and the optimal disassembly task is determined.

[0006] In a preferred embodiment, the steps of defining the pheromone of each subtask and determining the optimal disassembly task are as follows: Each subtask corresponds to a pheromone, which is represented based on task characteristics as: Wherein, represents the pheromone, represents the domain weight, represents the historical success rate, is an adjustable parameter; Evaluate the score of the path by integrating the pheromone concentration, task urgency, and resource cost: Wherein, represents the score, represents the task urgency, represents the resource cost; Sort the tasks according to the score, and preferentially allocate disassemblers to the paths with high scores to obtain the optimal disassembly task; At the same time, in the long conversation scenario, the update of the pheromone based on the context conversation includes positive enhancement and negative attenuation; Define a preset confidence level. When the subtask is executed and exceeds the preset confidence level, the pheromone is positively enhanced, and the increment is , where is an adjustable parameter, represents the incentive function; When the subtask has not been mentioned for more than the preset time, it is negatively attenuated: , where is an adjustable parameter, represents the subtask The pheromone negatively attenuated at time t; And set up a shared pool to store the disassembled subtasks for processing audit problems exceeding the similarity threshold, and the similarity threshold is based on the coincidence rate of the subtasks.

[0007] In a preferred embodiment, the steps of classifying the disassembled subtasks by the knowledge classification method of multi-model fusion are as follows: Input the optimal disassembly task into the word segmentation system to convert it into text data and divide it into multiple single-topic vocabulary; Compare multiple single-topic vocabulary with the audit domain knowledge graph to judge whether the single-topic vocabulary exists; If not, classify it through the KG-BERT classification model and output the classification result; where the KG-BERT classification model is a classification model that integrates the knowledge graph of the audit field and the large-scale pre-trained language model; If it exists, determine whether the single-topic vocabulary is unique. If it is unique, output the classification result; If the single-topic vocabulary maps to multiple topics, classify it through the topic weight determination model and output the classification result; where the topic weight determination model calculates the weights of multiple topics using the TF-IDF algorithm and selects the topics exceeding the preset weight as the classification result for output; where the classification result includes the topic category, the topic entity, and the topic association; Match the classification result with the audit knowledge base to obtain the matching results of defined audit knowledge and undefined audit knowledge.

[0008] In a preferred embodiment, the steps for processing the defined audit knowledge are as follows: Design a knowledge reuse mechanism for the defined audit knowledge; Adopt a meta-learning framework, extract historical audit meta-knowledge from historical audit tasks, and use the historical audit meta-knowledge to train the meta-learning neural network; Use the meta-learning neural network to analyze the defined audit knowledge to give answers to questions and output them to the user.

[0009] In a preferred embodiment, the steps for the audit agent to analyze and generate answers to questions for the undefined audit knowledge through the adaptive improvement iteration mechanism, integrating the NLP algorithm and deep learning are as follows: Design an adaptive improvement iteration mechanism including web crawling technology and a question-answer knowledge association tracing mechanism; Use web crawling technology to obtain network audit knowledge related to the undefined audit knowledge, and screen and integrate the network audit knowledge through the question-answer knowledge association tracing mechanism to obtain updated audit knowledge; At the same time, use the NLP algorithm to analyze the updated audit knowledge and the undefined audit knowledge to perform multi-level semantic analysis to obtain the structured information of the undefined audit knowledge; Construct an audit knowledge learning model based on the Transformer structure, construct an attention layer, and define a query set , a set of keys , a set of values , where each query vector , each key vector , each value vector ; The dot product calculation result is expressed as: where, is the dot product calculation result, ATT-Mask is used to limit the attention of each query vector to key-value pairs. If the attention of the query vector to the key vector is reduced, then otherwise ; Meanwhile, a generative adversarial network is introduced: where D represents the discriminative network and G represents the generative network, represents the distribution of the input data and represents the input noise variable; The structured information is input into the audit knowledge learning model for learning to obtain the question answer and output it to the user. Meanwhile, the updated audit knowledge is stored in the audit knowledge base.

[0010] In a preferred embodiment, the steps of the dynamic permission control are as follows: Define that the dynamic permission control includes a role environment function, role methods, a machine learning prediction algorithm, and a dynamic permission database; where the role environment function includes environment parameters, user roles, and task types; Define the initial role according to the user's function, allocate basic permissions, and preset a static permission set for each role; When the user logs in to the audit agent, the audit agent calls the role environment function and calls the corresponding role method according to the output of the environment function to generate real-time permissions; Meanwhile, during the user's usage process, the user's real-time permissions are updated in real time according to the user's historical behavior data and the current task context using the machine learning prediction algorithm; Write the adjusted real-time permissions into the dynamic permission database in real time.

[0011] In a preferred embodiment, the steps of the question and answer traceability mechanism are as follows: Construct a six-layer blockchain architecture including a data layer, a network layer, a consensus layer, an incentive layer, a contract layer, and an application layer; When the user initiates an audit question, a unique question and answer ID is generated based on the question and answer traceability mechanism, and the initial data including the time stamp, user ID, and question content is recorded as the block data fingerprint; The network layer broadcasts the audit question to all nodes, triggering the consensus layer to verify the legitimacy of the request; The nodes verify the legitimacy of the question through the consensus algorithm, package the verified audit question and the subsequent generated question answer into a block, and add it to the blockchain; Synchronize the block to all nodes, update the distributed ledger, and at the same time perform intelligent retrieval of the block through the semantic matching algorithm to trace audit problems.

[0012] In the above technical solution, the technical effects and advantages provided by the present invention are as follows: 1. The present invention disassembles complex audit problems into multiple subtasks through an intelligent strategy and combines a knowledge classification method of multi-model fusion for refined management. First, the one-to-one splitting strategy is used to divide the problem into multiple subtasks, and the optimal splitting path is evaluated through indicators such as pheromone concentration, urgency, and resource cost. This splitting method can not only effectively reduce the complexity of the problem but also optimize the task processing order through the dynamic update mechanism of pheromone. In addition, through the classification method of multi-model fusion, the system can quickly classify the subtasks into defined and undefined audit knowledge. This precise classification mechanism significantly improves the processing efficiency of audit problems. Especially when facing large-scale and multi-level audit tasks, the system can quickly locate the core of the problem, reduce redundant operations, and improve the response speed. 2. The present invention can dynamically obtain and integrate undefined audit knowledge by designing an adaptive improvement iteration mechanism, combining web crawling technology and a question-and-answer knowledge association tracing mechanism. Through multi-level semantic analysis of undefined audit knowledge by the NLP algorithm, the system can generate structured information and perform deep learning using an audit knowledge learning model based on Transformer. The attention mechanism of the Transformer model can capture complex semantic relationships, and the introduction of the generative adversarial network further enhances the model's generative ability and robustness. This adaptive iteration mechanism can not only handle existing undefined problems but also continuously update the system's knowledge base through continuous network knowledge updates to ensure its efficient processing ability when facing emerging audit problems. 3. The present invention ensures the data security and compliance of the audit knowledge learning agent during use through a dynamic permission control mechanism and blockchain tracing technology. The dynamic permission control adjusts the user's access permissions in real time according to the user's role, task type, and historical behavior, effectively preventing unauthorized data access and operations. At the same time, the permission prediction algorithm based on machine learning can automatically update the user's permissions according to the context environment to ensure the accuracy and security of permission allocation. In addition, through the six-layer architecture constructed by blockchain technology, the system can record the operation history of each audit problem, generate a unique question-and-answer ID and timestamp, ensuring the immutability and traceability of the data. This blockchain tracing mechanism not only enhances the transparency and compliance of the system but also can quickly locate the responsible party in case of problems, significantly improving the security and trust of the audit work. Description of the Drawings

[0013] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0014] Figure 1 It is a flowchart of the method of the present invention. Specific embodiments

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0016] Example 1. Please refer to Figure 1 As shown, the training method of the knowledge learning intelligent agent based on the deep learning model in this embodiment includes: S1. Disassemble the audit problem through an intelligent strategy to obtain the optimal disassembly task, and classify the optimal disassembly task through a knowledge classification method of multi-model fusion to obtain defined audit knowledge and undefined audit knowledge; S2. Analyze the undefined audit knowledge through an adaptive improvement iteration mechanism, an audit intelligent agent that integrates the NLP algorithm and deep learning to generate problem answers; S3. Use the dynamic permission control and question-and-answer traceability mechanism to control the data security during the use of the audit knowledge learning intelligent agent; As described in the above steps S1 - S3, with the rapid development of artificial intelligence technology, the intelligent audit system has gradually become the focus of attention in the power industry. Although the existing intelligent audit tools have made certain progress in data processing and automation, they still face many challenges and disadvantages in practical applications; First, the knowledge coverage rate of the intelligent audit system is relatively low. The existing audit knowledge base mainly relies on historical data and rule bases, and cannot effectively cover emerging audit scenarios and complex technical issues. Secondly, the existing intelligent audit system lacks sufficient knowledge reserves. Although the system can store and process a large amount of audit data, when faced with complex cross-domain problems, the existing knowledge base often lacks sufficient depth and breadth and is difficult to provide comprehensive support. In addition, there are also obvious deficiencies in the knowledge reuse of the intelligent audit system. The existing knowledge management mechanisms are mostly static storage and cannot be dynamically adjusted and optimized according to new audit requirements, resulting in a large amount of audit experience being idle after completing a task and unable to provide continuous reference for subsequent work; The present invention disassembles complex audit problems into multiple subtasks through an intelligent strategy and conducts refined management in combination with a knowledge classification method of multi-model fusion. First, the one-to-one splitting strategy is used to divide the problem into multiple subtasks, and the optimal splitting path is evaluated through indicators such as pheromone concentration, urgency, and resource cost. This splitting method can not only effectively reduce the complexity of the problem but also optimize the task processing order through the dynamic update mechanism of pheromones. In addition, through the classification method of multi-model fusion, the system can quickly classify the subtasks into defined and undefined audit knowledge. This precise classification mechanism significantly improves the processing efficiency of audit problems. Especially when faced with large-scale and multi-level audit tasks, the system can quickly locate the core of the problem, reduce redundant operations, and improve the response speed; By designing an adaptive improvement and iteration mechanism, combining web crawling technology and a question-and-answer knowledge association and tracing mechanism, undefined audit knowledge can be dynamically obtained and integrated. Through multi-level semantic analysis of the undefined audit knowledge using NLP algorithms, the system can generate structured information and perform deep learning using an audit knowledge learning model based on Transformer. The attention mechanism of the Transformer model can capture complex semantic relationships, and the introduction of a generative adversarial network further enhances the model's generative ability and robustness. This adaptive iteration mechanism can not only handle existing undefined problems but also continuously update the system's knowledge base through continuous network knowledge updates to ensure its efficient processing ability when faced with emerging audit problems; Through the dynamic permission control mechanism and blockchain traceability technology, the data security and compliance in the use of the audit knowledge learning agent are ensured. The dynamic permission control adjusts the user's access permissions in real time according to the user's role, task type, and historical behavior, effectively preventing unauthorized data access and operations. At the same time, the permission prediction algorithm based on machine learning can automatically update the user's permissions according to the context environment, ensuring the accuracy and security of permission allocation. In addition, through the six-layer architecture built by blockchain technology, the system can record the operation history of each audit question, generate a unique Q&A ID and timestamp, ensuring the immutability and traceability of the data. This blockchain traceability mechanism not only enhances the transparency and compliance of the system but also can quickly locate the responsible party in case of problems, significantly improving the security and trust of the audit work.

[0017] In one embodiment, step S1 of disassembling the audit question into the optimal disassembly task through the intelligent strategy includes: S11. Define the level of the audit question as M; S12. Use the one-to-one splitting strategy to divide the audit question into M parts, and each part is used as a subtask; S13. Regard each subtask as a disassembly path, define the pheromone of each subtask, and determine the optimal disassembly task; As described in the above steps S11 - S13, in the face of complex audit questions, the audit questions are divided into multiple levels, defined as M levels. Each level represents a different level of abstraction, containing subtasks from the overall audit question to specific details. Then, using the one-to-one splitting strategy, the audit question is divided into M subtasks. Each subtask focuses on solving a specific aspect of the audit question. At the same time, each subtask is regarded as a disassembly path, that is, each subtask can be used as the starting point and node of the disassembly path, and pheromone is defined for each subtask to select the optimal disassembly task. By defining the level M, adopting the one-to-one splitting strategy, and establishing the pheromone mechanism, complex audit questions can be flexibly and efficiently transformed into executable and optimizable specific tasks, ultimately improving the audit efficiency and accuracy.

[0018] In one embodiment, step S13 of defining the pheromone of each subtask and determining the optimal disassembly task includes: S131. Each subtask corresponds to a pheromone, which is represented based on the task characteristics as: S132. Among them, represents the pheromone, represents the domain weight, represents the historical success rate, is an adjustable parameter; S133. Scoring of the comprehensive pheromone concentration, task urgency, and resource cost assessment path: S134. Among them, represents the score, represents the task urgency, represents the resource cost; S135. Sort the tasks by score, and preferentially allocate disassemblers to the paths with high scores to obtain the optimal disassembly tasks; S136. At the same time, in the long conversation scenario, the update of pheromone based on the context conversation includes positive enhancement and negative attenuation; S137. Define the preset confidence level. When the subtask is executed and exceeds the preset confidence level, the pheromone is positively enhanced, and the increment is , where is an adjustable parameter, represents the incentive function; S138. When the subtask has not been mentioned for more than the preset time, it is negatively attenuated: , where is an adjustable parameter, represents the subtask the pheromone that is negatively attenuated at time t; S139. And set up a shared pool to store the disassembled subtasks for handling audit problems that exceed the similarity threshold. The similarity threshold is based on the coincidence rate of the subtasks; As described in the above steps S131 - S139, a pheromone is defined for each subtask. The composition of the pheromone consists of a domain weight and a historical success rate. The domain weight is used to quantify the inherent importance or priority of a subtask within the audit domain. For example, if the audit question raised by the user includes the security issue of a certain device, then the subtasks related to device security or failure rate have a heavier domain weight. The historical success rate is a core indicator for measuring task reliability and execution efficiency, indicating the proportion of a certain subtask that has been successfully completed in the past executions in the audit system. For new tasks or tasks without historical data, an initial success rate can be assigned, which can be selected as 0.7 to represent a neutral level of trust. At the same time, in the actual implementation plan, expert knowledge injection can be adopted. Through expert experience, a higher initial success rate can be manually set for key tasks. The decomposition path starts from the task without a previous task or the task with the highest pheromone concentration. When the task has no subsequent tasks or reaches the preset confidence threshold, the path terminates. Then, the score of the decomposition path is comprehensively evaluated based on pheromone concentration, urgency, and resource cost, and the decomposition path that exceeds the preset score threshold is selected as the optimal decomposition task. For the long - dialogue scenario, which is a scenario of multiple conversations with the audit learning agent and generates context dialogue information, positive enhancement and negative attenuation of pheromones are performed for the context dialogue. When a subtask is correctly executed and passes the verification, that is, the auditor or user clearly approves the result generated by the system, the pheromone of this subtask is enhanced, and the value of the incentive function is the same as the confidence of the subtask. At the same time, in the actual implementation, the situations that can be considered for positive enhancement include: the subtask has been successfully executed multiple times in history, forming a stable high - success - rate record, and multiple subtasks are continuously successful due to logical association. At the same time, when the user is not satisfied with the execution of a certain subtask, negative attenuation is performed on this subtask. Also, if a subtask has not been mentioned within the preset time, negative attenuation will be performed in a timely manner to ensure the optimal utilization of resources. At the same time, a shared pool is set up to store the decomposed subtasks. When the user or auditor raises a similar audit question, the same subtasks can be extracted from the shared pool to avoid wasting resources caused by repeated calculations. A preset similarity threshold is set, and the setting standard of the threshold is based on the overlap rate of subtasks. That is, after the user re - inputs an audit question and it is divided into multiple subtasks, the overlap rate of multiple subtasks with the subtasks of historical audit questions is calculated. If the overlap rate is higher than the similarity threshold, it is determined as a similar audit question, and the relevant data of the subtasks of the historical audit question can be directly used.

[0019] In one embodiment, step S1 of classifying the decomposed subtasks by the knowledge classification method through multi - model fusion includes: S14. Input the optimal decomposition task into a word - segmentation system to convert it into text data and divide it into multiple single - topic words; S15. Compare the multiple single - topic words with the audit - domain knowledge graph to determine whether the single - topic words exist; S16. If not, classify using the KG-BERT classification model and output the classification result; S17. The KG-BERT classification model is a classification model that integrates the knowledge graph of the audit field and the large-scale pre-trained language model; S18. If it exists, determine whether the single-topic vocabulary is unique. If it is unique, output the classification result; S19. If the single-topic vocabulary maps to multiple topics, classify using the topic weight determination model and output the classification result; S110. The topic weight determination model calculates the weights of multiple topics using the TF-IDF algorithm and selects the topics with weights exceeding the preset weight as the classification result for output; S111. The classification result includes the topic category, topic entity, and topic association; S112. Match the classification result with the audit knowledge base to obtain the matching results of defined audit knowledge and undefined audit knowledge; As described in the above steps S14 - S112, first, the optimal disassembly task is input into the word segmentation system to be converted into text data. Then, continuous sentences or paragraphs are disassembled into multiple single - topic words. Word segmentation can adopt rule - based word segmentation methods such as forward maximum matching and reverse maximum matching to ensure the accuracy of the words. The generated single - topic words are compared with the audit domain knowledge graph to determine whether these words already exist in the audit domain knowledge graph. The audit domain knowledge graph usually contains information such as entities, relationships, and attributes in the audit domain, which can provide rich context for text classification. If the word exists in the knowledge graph, it enters the next step of processing; if not, it needs to be classified through the KG - BERT classification model. Among them, KG - BERT is a classification model that integrates the knowledge graph and BERT. BERT is a pre - trained language model based on the Transformer architecture, which can capture the context information of the text. On the basis of BERT, KG - BERT introduces the entity and relationship information in the audit domain knowledge graph, improving the classification performance of the model in a specific domain. For words not found in the knowledge graph, KG - BERT will classify them into the most suitable audit topic according to the context information. The classification results include topic categories, topic entities, and topic associations. If the topic word exists in the knowledge graph but maps to multiple topics, that is, one word corresponds to multiple topics, it is further processed through the topic weight determination model. The topic weight determination model uses the TF - IDF algorithm to calculate the weight of each topic. TF - IDF evaluates the importance of a word to a certain topic by calculating the frequency of the word's appearance in the audit question and its inverse document frequency in the entire corpus. According to the TF - IDF calculation results, the system will select the topics with weights exceeding the preset weight for output to ensure the accuracy and rationality of the classification results. The final classification results are matched with the audit knowledge base to determine whether they already exist in the audit knowledge base. The audit knowledge base contains knowledge related to audits in the power industry, including audit rules, case analyses, audit standards, etc. If the classification results are consistent with the existing knowledge in the audit knowledge base, it means that the classification results have been recognized and recorded by the system and are used as defined audit knowledge. If no matching item is found in the knowledge base for the classification results, it may indicate new audit knowledge or audit information not yet recognized by the system and is used as undefined audit knowledge.

[0020] In one embodiment, step 112 of the processing of the defined audit knowledge includes: S1121. Design a knowledge reuse mechanism for the defined audit knowledge; S1122. Adopt a meta - learning framework to extract historical audit meta - knowledge from historical audit tasks, and use the historical audit meta - knowledge to train the meta - learning neural network; S1123. Analyze the defined audit knowledge using a meta-learning neural network to give answers to questions and output them to the user; As described in the above steps S1121 - S1123, in order to improve the utilization efficiency of audit knowledge, a knowledge reuse mechanism is designed. The core idea of this mechanism is to extract valuable knowledge from historical audit tasks and apply it to new audit tasks to avoid repetitive labor and resource waste. The defined audit knowledge is stored in a structured knowledge base, which can include audit rules, cases, standards, etc. The audit knowledge base is usually stored using a graph database such as Neo4j or a relational database such as MySQL to ensure efficient querying and updating of knowledge. In the actual implementation scheme, a meta-learning framework is used to achieve knowledge reuse. Meta-learning is a technology that enables a model to learn how to learn and can quickly adapt to new tasks from a small amount of data. In this scheme, the meta-learning framework is used to extract meta-knowledge from historical audit tasks and use this meta-knowledge to train a meta-learning neural network. The knowledge reuse mechanism first analyzes a large number of historical audit tasks to extract meta-knowledge. Meta-knowledge can include general patterns of audit tasks, solutions to common problems, applicable conditions of audit rules, etc. Through self-supervised learning techniques, high-level features are extracted from historical audit tasks, and these features can reflect the essential laws of audit tasks. For example, pre-trained models such as BERT are used to embed the text data of historical audit tasks to extract semantic features. The extracted high-level features are used as a dataset to train the meta-learning neural network. The meta-learning neural network usually adopts a memory-based neural network. Further, in order to improve the generalization ability of the model, the knowledge reuse mechanism can adopt multi-task learning to enable the model to learn multiple different types of audit tasks simultaneously to ensure that it can handle diverse audit problems. Finally, the constructed meta-learning neural network is used to analyze the defined audit knowledge to give answers to questions and output them to the user.

[0021] In one embodiment, step S2 of analyzing undefined audit knowledge by the audit agent that combines the NLP algorithm and deep learning through an adaptive improvement iteration mechanism includes: S21. Design an adaptive improvement iteration mechanism including a web crawling technology and a question-and-answer knowledge association and tracing mechanism; S22. Use the web crawling technology to obtain network audit knowledge related to the undefined audit knowledge, and screen and integrate the network audit knowledge through the question-and-answer knowledge association and tracing mechanism to obtain updated audit knowledge; S23. At the same time, use the NLP algorithm to analyze the updated audit knowledge and the undefined audit knowledge for multi-level semantic analysis to obtain the structured information of the undefined audit knowledge; S24. Construct an audit knowledge learning model based on the Transformer structure, construct an attention layer, and define a query set , a set of keys , a set of values , where each query vector , each key vector , each value vector ; S25. Then the dot product calculation result is expressed as: S26. Among them, is the dot product calculation result, ATT-Mask is used to limit the attention of each query vector to key-value pairs. If the attention of the query vector to the key vector is reduced, then , otherwise ; S27. At the same time, introduce a generative adversarial network: S28. Among them, D represents the discriminative network, G represents the generative network, represents the distribution of the input data , represents the input noise variable; S29. Input the structured information into the audit knowledge learning model to learn the question answer and output it to the user, and at the same time store the updated audit knowledge in the audit knowledge base; As described in the above steps S21 - S29, when dealing with undefined audit knowledge, an adaptive improvement iteration mechanism is introduced. By supporting web crawling and question - answer knowledge association tracing, the system can automatically expand its knowledge base to achieve intelligent response to emerging audit problems. Specifically, the system first obtains relevant audit knowledge from multi - source data through web crawling technology, and then uses the question - answer knowledge association tracing mechanism to screen and integrate the acquired information. Through multiple iterations, the system can gradually improve its audit knowledge base, form a deep understanding and effective application of new knowledge. At the same time, NLP algorithms are used to perform multi - level semantic analysis on the updated audit knowledge and undefined audit knowledge to ensure that the intelligent agent can deeply understand the question - answer content and make accurate responses. Then, an audit knowledge learning model based on the Transformer structure is constructed to implement the attention mechanism. The introduction of the attention mechanism enables the audit knowledge learning model to dynamically learn and focus on the most important parts of the input data, and the weights of different inputs can be adaptively adjusted, avoiding the equal processing of traditional models and improving the processing efficiency. To enhance the knowledge learning ability of the intelligent agent, this solution also introduces an adversarial training method. By designing a generative adversarial network, the intelligent agent can improve its audit knowledge generation and reasoning ability in continuous adversarial training. The generative network is responsible for generating new audit knowledge, and the discriminative network verifies and screens the generated audit knowledge. Through the confrontation and cooperation of the two, the intelligent agent can gradually optimize its audit knowledge base and improve the quality and applicability of knowledge generation. In actual implementation, the adaptive improvement iteration mechanism can externally expand the audit knowledge base, and the generative adversarial network can internally expand the audit knowledge base, further improving the efficiency of the audit process.

[0022] In one embodiment, step S3 of the dynamic permission control includes: S31. Define that the dynamic permission control includes a role environment function, role methods, a machine learning prediction algorithm, and a dynamic permission database; S32. The role environment function includes environment parameters, user roles, and task types; S33. Define the initial role according to the user's function, assign basic permissions, and preset a static permission set for each role; S34. When the user logs in to the audit intelligent agent, the audit intelligent agent calls the role environment function and calls the corresponding role method according to the output of the environment function to generate real - time permissions; S35. At the same time, during the user's usage process, use the machine learning prediction algorithm to update the user's real - time permissions in real - time according to the user's historical behavior data and the current task context; S36. Write the adjusted real - time permissions into the dynamic permission database in real - time; As described in the above steps S31 - S36, dynamic permission control mainly consists of several core components, including a role environment function for dynamically capturing and analyzing the user's environment information and role type, a role method for implementing corresponding permission policies according to different roles, a machine learning prediction algorithm for real - time updating and adjusting permissions based on user behavior and task context, and a dynamic permission database for storing and managing user permission information and supporting real - time updates. First, each user is assigned an initial role according to their function, such as auditor, administrator, or technical support. Basic permissions are defined for each role to confirm the basic operation permissions of the role in the system, and a static permission set is preset for each role. When the user logs in to the audit agent, the system first performs identity authentication and identifies the user role. Then the system calls the role environment function to collect the required environmental parameters and generates real - time permissions through the role method. The real - time permissions will be dynamically adjusted based on the environmental parameters to ensure that the user can only access the information they truly need and have permission to access. At the same time, during the user's use of the audit agent, the system continuously collects the user's historical behavior data, analyzes the user's behavior and current task context through the machine learning prediction algorithm, and predicts the permissions they may need in real - time. Finally, a dynamic permission database is constructed. The dynamic permission database internally includes a user permission table, a permission change record table, a role information table, etc., and can quickly respond to permission query and update requests. When the user's permissions change, the new permission configuration and change records should be immediately written into the dynamic permission database to ensure the latest status of all permissions. At the same time, all permission changes should have an audit ability to record the historical changes of permissions to meet compliance requirements.

[0023] In one embodiment, step S3 of the Q&A traceability mechanism includes: S37. Construct a six - layer blockchain architecture including a data layer, a network layer, a consensus layer, an incentive layer, a contract layer, and an application layer; S38. When the user initiates an audit question, generate a unique Q&A ID based on the Q&A traceability mechanism, and record the initial data including the timestamp, user ID, and question content as the block data fingerprint; S39. The network layer broadcasts the audit question to all nodes, triggering the consensus layer to verify the legitimacy of the request; S310. The nodes verify the legitimacy of the question through the consensus algorithm, package the verified audit question and the subsequent generated question answer into a block, and add it to the blockchain; S311. Synchronize the block to all nodes, update the distributed ledger, and at the same time perform intelligent retrieval of the block through the semantic matching algorithm for audit question traceability; As described in the above steps S37 - S311, a six - layer blockchain architecture is constructed to implement the Q&A traceability mechanism, ensuring the security, transparency, and traceability of each audit question. Combining smart contracts, consensus mechanisms, and semantic matching algorithms, highly reliable audit question processing and retrieval are achieved.

[0024] As described above, this is only the specific implementation manner of the present application. However, the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claimed rights.

Claims

1. A method for training a knowledge learning agent based on a deep learning model, characterized in that: Through intelligent strategies, audit problems are decomposed to obtain the optimal decomposition tasks, and the optimal decomposition tasks are classified through the knowledge classification method of multi-model fusion to obtain defined audit knowledge and undefined audit knowledge; Through the adaptive improvement iterative mechanism, the audit agent integrating NLP algorithm and deep learning analyzes the undefined audit knowledge and generates answers to questions; Dynamic permission control and question-and-answer tracing mechanism are used to control data security during the use of audit knowledge learning agents.

2. The method for training a knowledge learning agent based on a deep learning model according to claim 1, characterized in that: The steps of decomposing the audit problem by intelligent strategy to obtain the optimal decomposition task are: Define the level of audit questions as M; Use a one-to-one splitting strategy to divide the audit problem into M parts, each of which is a subtask; Consider each subtask as a disassembly path, define the pheromone of each subtask and determine the optimal disassembly task.

3. The method for training a knowledge learning agent based on a deep learning model according to claim 2, characterized in that: The steps of defining the pheromone of each subtask and determining the optimal disassembly task are: Each subtask corresponds to a pheromone, which is expressed based on the task characteristics as follows: in, Indicates pheromone, represents the domain weight, represents the historical success rate, is an adjustable parameter; The score of the evaluation path is based on the comprehensive pheromone concentration, task urgency and resource cost: in, Indicates the rating, Indicates the urgency of the task. represents resource cost; Sort tasks by score, and assign disassemblers to paths with high scores first to obtain the best disassembly tasks; At the same time, in the long dialogue scenario, the updating of pheromones based on the contextual dialogue includes positive reinforcement and negative attenuation; Define the preset confidence. When the subtask is executed and exceeds the preset confidence, the pheromone is positively enhanced, and the increment is ,in is an adjustable parameter, represents the activation function; When a subtask is not mentioned for a preset time, it decays negatively: ,in is an adjustable parameter, Represents a subtask Pheromones that decay negatively at all times; A shared pool is set up to store the disassembled subtasks for handling audit issues that exceed a similarity threshold, where the similarity threshold is based on the overlap rate of the subtasks.

4. The method for training a knowledge learning agent based on a deep learning model according to claim 1, characterized in that: The steps of classifying the disassembly subtasks by the knowledge classification method of multi-model fusion are as follows: The optimal decomposition task is input into the word segmentation system and converted into text data and divided into multiple single-topic words; Compare multiple single-topic words with the audit domain knowledge graph to determine whether the single-topic words exist; If it does not exist, it is classified through the KG-BERT classification model and the classification result is output; The KG-BERT classification model is a classification model that integrates the auditing domain knowledge graph and a large-scale pre-trained language model; If it exists, determine whether the single-topic vocabulary is unique, and if it is unique, output the classification result; If a single topic word maps to multiple topics, the topic weight determination model is used to classify it and output the classification result; The topic weight determination model uses the TF-IDF algorithm to calculate the weights of multiple topics, and selects topics that exceed the preset weights as classification results for output; The classification results include subject categories, subject entities and subject associations; The classification results are matched with the audit knowledge base, and the matching results are defined audit knowledge and undefined audit knowledge.

5. The method for training a knowledge learning agent based on a deep learning model according to claim 4, characterized in that: The steps for processing defined audit knowledge are: Design knowledge reuse mechanisms for defined audit knowledge; A meta-learning framework is adopted to extract historical audit meta-knowledge from historical audit tasks, and the historical audit meta-knowledge is used to train the meta-learning neural network; Meta-learning neural networks are used to analyze defined audit knowledge to give answers to questions and output them to users.

6. The method for training a knowledge learning agent based on a deep learning model according to claim 1, characterized in that: The steps of analyzing undefined audit knowledge and generating answers to questions through the audit agent that adaptively improves the iterative mechanism and integrates the NLP algorithm and deep learning are as follows: Design adaptive improvement and iteration mechanisms including web crawling technology and question-answering knowledge association tracing mechanism; Use web crawling technology to obtain network audit knowledge related to undefined audit knowledge, and screen and integrate network audit knowledge through question-answer knowledge association tracing mechanism to obtain updated audit knowledge; At the same time, NLP algorithm analysis is used to conduct multi-level semantic analysis on updated audit knowledge and undefined audit knowledge to obtain structured information of undefined audit knowledge; Build an audit knowledge learning model based on Transformer structure, build an attention layer, and define a query set , a collection of keys , a set of values , where each query vector , each key vector , each value vector ; The dot product calculation result is expressed as: in, is the result of dot product calculation. ATT-Mask is used to limit the attention of each query vector to the key-value pair. If the attention of the query vector to the key vector is reduced, then ,otherwise ; At the same time, the generation of adversarial networks is introduced: Among them, D represents the discriminant network, G represents the generative network, Represents input data The distribution of represents the input noise variable; The structured information is input into the audit knowledge learning model to learn the answers to the questions and output them to the user. At the same time, the updated audit knowledge is stored in the audit knowledge base.

7. The method for training a knowledge learning agent based on a deep learning model according to claim 1, characterized in that: The steps of dynamic permission control are: Define dynamic permission control including role environment function, role method, machine learning prediction algorithm and dynamic permission database; The role environment function includes environment parameters, user roles, and task types; Define initial roles based on user functions, assign basic permissions, and preset static permission sets for each role; When a user logs in to the audit agent, the audit agent calls the role environment function, calls the corresponding role method according to the output of the environment function, and generates real-time permissions; At the same time, during the user's use, the user's real-time permissions are updated in real time using machine learning prediction algorithms based on the user's historical behavior data and current task context; The adjusted real-time permissions are written into the dynamic permissions database in real time.

8. The method for training a knowledge learning agent based on a deep learning model according to claim 1, characterized in that: The steps of the question-answer tracing mechanism are: Build a six-layer blockchain architecture including data layer, network layer, consensus layer, incentive layer, contract layer and application layer; When a user initiates an audit question, a unique question and answer ID is generated based on the question and answer tracing mechanism, and the initial data including timestamp, user ID, and question content are recorded as the block data fingerprint; The network layer broadcasts the audit question to all nodes, triggering the consensus layer to verify the legitimacy of the request; The nodes verify the legitimacy of the questions through the consensus algorithm, package the verified audit questions and the answers to the questions generated subsequently into blocks, and add them to the blockchain; Synchronize blocks to all nodes, update distributed ledgers, and intelligently retrieve blocks through semantic matching algorithms to trace audit issues.

Images