Multi-dimensional knowledge management method and system based on multi-agent collaboration

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-dimensional knowledge management method and system based on multi-agent collaboration. The multi-dimensional knowledge management method based on multi-agent collaboration comprises the following steps: acquiring knowledge data, and classifying and mining the knowledge data; a hierarchical hybrid expert architecture is constructed, acceleration processing is performed on an expert model, and dependency on computing power resources is reduced; the general teacher-level expert model understands an input task instruction, generates a field-level agent establishment workflow, calls the research-level expert model to complete industry-level task disassembly, generates an industry-level agent establishment workflow, and sends an instruction to call the engineer-level expert model to generate a function-level agent establishment workflow; and a corresponding tool is called to complete the task. According to the multi-dimensional knowledge management method and system based on multi-agent collaboration, an artificial intelligence means is used for replacing manual work to complete knowledge management service work, the automation degree is high, the dependency on computing power resources is low, the management cost is reduced, and the management efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-dimensional knowledge management method and system based on multi-agent collaboration. Background Art

[0002] With the deepening of informatization, massive amounts of data are exploding. Managing and storing this vast amount of data, analyzing and mining the valuable information within it, and improving people's work and learning efficiency are becoming increasingly pressing issues. To meet this demand, knowledge bases have emerged. They aggregate vast amounts of data and describe the connections between objects in the objective world in the form of a semantic web. Currently, knowledge bases have been applied in numerous fields and are playing a crucial role.

[0003] In the operational use of knowledge bases, knowledge base construction is the foundation for knowledge base systems to provide knowledge services and knowledge management, and directly impacts knowledge base system performance. Knowledge base construction involves tasks such as knowledge cognition extraction, knowledge annotation and classification, and knowledge mining governance, resulting in complex processes and a high workload. However, effective tools are often lacking in knowledge base construction, requiring much of the work to be completed manually, which consumes significant time.

[0004] In response to the needs of knowledge base construction, the present invention proposes a multi-dimensional knowledge management method and system based on multi-agent collaboration. Summary of the Invention

[0005] In order to remedy the deficiencies of the prior art, the present invention provides a simple and efficient multi-dimensional knowledge management method and system based on multi-agent collaboration.

[0006] The present invention is achieved through the following technical solutions: A multi-dimensional knowledge management method based on multi-agent collaboration includes the following steps: Step S1: Build a tool set for data extraction, annotation, and analysis, including a data agile acquisition tool set, a data rapid annotation tool set, and a data analysis and mining tool set to obtain multi-dimensional knowledge data and perform classification and mining; In step S1, the data agile acquisition tool set integrates a face recognition model, an image recognition model, a speech recognition model, and an optical character recognition model, which are used to acquire knowledge data from images, speech, and files to form a knowledge data set, including personnel information data, natural resource data, operation and maintenance data, scientific research data, and organizational relationship data.

[0007] In step S1, the data rapid annotation tool set includes an unsupervised clustering tool, a decision tree tool, an SVM classifier, and a neural network classifier; Use unsupervised clustering tools to perform horizontal clustering on the unclassified data in the knowledge dataset, classify the data belonging to different data subjects according to data types, and obtain the number of data types; Decision tree tools, SVM classifiers and / or neural network classifiers are used to vertically classify unclassified data in the knowledge dataset, and data belonging to the same data subject are classified according to data type.

[0008] In step S1, the data analysis and mining tool set includes a Gaussian process regression algorithm tool, a correlation analysis algorithm tool, a fuzzy calculation tool, and a deep learning regression algorithm tool, which customize the mining and analysis of the classified data according to the application scenario.

[0009] Step S2: Build a hierarchical hybrid expert architecture, dividing the expert models into chief engineer-level expert models, researcher-level expert models, and engineer-level expert models, which are responsible for task decomposition and agent construction and orchestration at different levels respectively; The chief engineer-level expert model is used to understand user instructions, complete domain-level task decomposition, generate domain-level intelligent agents to establish workflows, and call the researcher-level expert model; The researcher-level expert model is used to disassemble the instruction tasks issued by the domain-level intelligent agent, generate the industry-level intelligent agent to establish the workflow, and call the engineer-level expert model; The engineer-level expert model is used to disassemble the instruction tasks issued by the industry-level intelligent agent, generate the functional-level intelligent agent to establish the workflow, and complete the task.

[0010] Step S3: Accelerate the expert model using model training and inference acceleration technology based on block matrix compression decomposition to reduce the dependence on computing resources during model deployment and application; In step S3, the specific process is as follows: Step S3.1: Use the stacking grouping technique to group and decompose the large-dimensional matrix in the expert model into a stacked block matrix structure, where the dimension of the smallest matrix block in the stacked block matrix structure is less than the custom dimension threshold min_n; In step S3.1, the specific steps are as follows: Step S3.1.1, set the minimum matrix block dimension threshold during the group decomposition process, denoted as min_n; Step S3.1.2: Perform custom row and column splitting of equal dimensions on the matrix to obtain the first-layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.3: Perform custom row and column splitting of the diagonal blocks of the current layer grouping matrix with equal dimensions to obtain the next layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.4: Repeat step S3.1.3 until the minimum matrix block dimension is less than the dimension threshold min_n, and then stop the operation.

[0011] Step S3.2: Use the block matrix decomposition technique based on Schmidt orthogonalization to transform The non-diagonal block matrix Z is compressed and decomposed into two orthogonal matrices and Multiply, the dimensions are and , , two orthogonal matrices and The required memory size is ; In step S3.2, the specific steps are as follows: Step S3.2.1. Normalize the first row of matrix Z and assign it to the first row of matrix B. ,Right now: , row index ; Step S3.2.2: Determine and find the first column index , making ; Step S3.2.3, the impedance matrix After column normalization, we get the first column of the U matrix ,Right now: ; Step S3.2.4: Find the kth row index , making ; Step S3.2.5: The impedance matrix The kth row of the B matrix is obtained by orthogonalizing the first k-1 rows of the B matrix. ,Right now: , , Step S3.2.6: Determine and find the kth column index , making ; Step S3.2.7, the matrix Z The kth column of the U matrix is obtained by orthogonalizing the first k-1 columns of the U matrix. ,Right now: , ; Step S3.2.8: Determine whether the loop is finished. The determination formula is as follows: , in, is the truncation error threshold, which is controlled by The size of controls the size of the cutoff rank k; when The smaller it is, the higher the accuracy is; the larger k is, the more memory is required; If the above judgment formula is satisfied, the loop ends; otherwise, the k value is updated to , repeat steps S3.2.4 to S3.2.8 until the above judgment formula is satisfied; Step S3.2.9: After the loop is completed, the U matrix and the r rows of the original impedance matrix are obtained by solving the following equation: , Get the V matrix.

[0012] Step S3.3: recompress the obtained orthogonal matrices U and V using the singular value decomposition method, and the compression effect is achieved by controlling the size of the truncation precision; The specific steps are as follows: Step S3.3.1: The matrices U and V obtained in step S3.2 are respectively and , perform QR decomposition, expressed as: , , Among them, the superscript represents the dimension of the matrix, and the subscript represents the decomposed matrix. The computational complexity in the above formula is ; Step S3.3.2: Change the two dimensions in step S3.3.1 to Matrix and Multiply to get a new matrix , expressed as: , Step S3.3.3, matrix Perform singular value decomposition and truncate the middle diagonal matrix. The truncation accuracy is , after truncation, the rank is t, and we get: , The computational complexity of steps S3.3.2 and S3.3.3 is ; Step S3.3.4, get the original matrix The approximate expression of is: , Memory consumption is further reduced to ,in , which obviously achieves the goal of reducing computing time and storage resources.

[0013] Step S4: Based on the task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; Step S5: The researcher-level expert model understands the received instructions, decomposes the industry-level tasks, generates an industry-level agent to establish a workflow, and issues instructions to call the engineer-level expert model; Step S6: The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

[0014] A multi-agent collaborative multi-dimensional knowledge management system for implementing the above method includes: The tool set module includes agile data acquisition tools, rapid data annotation tools, and data analysis and mining tools, which are used to acquire multi-dimensional knowledge data and perform classification and mining. The model acceleration module is responsible for accelerating the processing of expert models using model training and inference acceleration technology based on block matrix compression decomposition to reduce the dependence on computing resources during model deployment and application; The expert model management module is responsible for establishing a hierarchical hybrid expert architecture for storing and running chief engineer-level expert models, researcher-level expert models, and engineer-level expert models; For task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; The researcher-level expert model understands the instructions it receives, decomposes industry-level tasks, generates industry-level intelligent agents to establish workflows, and issues instructions to invoke the engineer-level expert model. The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

[0015] A multi-dimensional knowledge management device based on multi-agent collaboration includes a memory and a processor; the memory is used to store a computer program, and the processor is used to implement the method when executing the computer program.

[0016] A readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0017] The beneficial effects of the present invention are: the multi-dimensional knowledge management method and system based on multi-agent collaboration uses artificial intelligence to replace manual knowledge management service work, has a high degree of automation, low dependence on computing resources, reduces management costs, improves management efficiency, and realizes efficient and intelligent knowledge management services, which can effectively promote the digital transformation of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Attachment Figure 1 Schematic diagram of the multi-dimensional knowledge management method based on multi-agent collaboration of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0021] The multi-dimensional knowledge management method based on multi-agent collaboration includes the following steps: Step S1: Build a tool set for data extraction, annotation, and analysis, including a data agile acquisition tool set, a data rapid annotation tool set, and a data analysis and mining tool set to obtain multi-dimensional knowledge data and perform classification and mining; Faced with a variety of data sources, including faces, ID photos, pictures, videos, work orders, files, tables, etc., in step S1, the data agile acquisition tool set integrates face recognition models, image recognition models, speech recognition models and optical character recognition models to obtain knowledge data from images, voices and files to form a knowledge data set, including personnel information data, natural resource data, operation and maintenance data, scientific research data and organizational relationship data.

[0022] In step S1, the data rapid annotation tool set includes an unsupervised clustering tool, a decision tree tool, an SVM classifier, and a neural network classifier; In a certain knowledge dataset, there is no strictly classified data. Unsupervised clustering tools are used to perform horizontal clustering on the unclassified data in the knowledge dataset. Data belonging to different data subjects are classified according to data type. For example, data in the personnel information knowledge dataset can be classified into identity ID, name, age, gender, place of origin, occupation, marital status, education level, etc., and the number of data types is obtained. Use decision tree tools, SVM classifiers and / or neural network classifiers to perform vertical classification on the unclassified data in the knowledge dataset, and classify the data belonging to the same data subject according to the data type. For example, in personnel information knowledge, the identity ID, name, age, gender, place of origin, occupation, marital status, education level and other data of the same person are classified together.

[0023] In step S1, the data analysis and mining tool set includes a Gaussian process regression algorithm tool, a correlation analysis algorithm tool, a fuzzy calculation tool, and a deep learning regression algorithm tool, which performs mining and analysis on the classified data according to the application scenario for application in different scenarios. According to the data classification information, the Gaussian process regression, correlation analysis, fuzzy calculation, deep learning regression and other algorithm tools in the data analysis and mining tools are used to perform mining and analysis on the classified data according to the application scenario for application in different scenarios, such as fault diagnosis and early warning, natural disaster judgment, violation prediction, resource utilization efficiency evaluation, scientific research hotspot evaluation, etc.

[0024] Step S2: Build a hierarchical hybrid expert architecture, dividing the expert models into chief engineer-level expert models, researcher-level expert models, and engineer-level expert models, which are responsible for task decomposition and agent construction and orchestration at different levels respectively; The chief engineer-level expert model is used to understand user instructions, complete domain-level task decomposition, generate domain-level intelligent agents to establish workflows, and call the researcher-level expert model; The researcher-level expert model is used to disassemble the instruction tasks issued by the domain-level intelligent agent, generate the industry-level intelligent agent to establish the workflow, and call the engineer-level expert model; The engineer-level expert model is used to disassemble the instruction tasks issued by the industry-level intelligent agent, generate the functional-level intelligent agent to establish the workflow, and complete the task.

[0025] Step S3: Using the model training and inference acceleration technology based on block matrix compression decomposition to accelerate the expert model, so as to reduce the dependence on computing resources during the model deployment and application process; In step S3, the specific process is as follows: Step S3.1: Use the stacking grouping technique to group and decompose the large-dimensional matrix in the expert model into a stacked block matrix structure, where the dimension of the smallest matrix block in the stacked block matrix structure is less than the custom dimension threshold min_n; In step S3.1, the specific steps are as follows: Step S3.1.1, set the minimum matrix block dimension threshold during group decomposition, denoted as min_n; Step S3.1.2: Perform custom row and column splitting of equal dimensions on the matrix to obtain the first-layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.3: Perform custom row and column splitting of the diagonal blocks of the current layer grouping matrix with equal dimensions to obtain the next layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.4: Repeat step S3.1.3 until the minimum matrix block dimension is less than the dimension threshold min_n, and then stop the operation.

[0026] Step S3.2: Use the block matrix decomposition technique based on Schmidt orthogonalization to transform The non-diagonal block matrix Z is compressed and decomposed into two orthogonal matrices and Multiply, the dimensions are and , , two orthogonal matrices and The required memory size is ; Generally r is much smaller than m and n, so and The memory required to store the two matrices is much smaller than ; In step S3.2, the specific steps are as follows: Step S3.2.1. Normalize the first row of matrix Z and assign it to the first row of matrix B. ,Right now: , row index ; Step S3.2.2: Determine and find the first column index , making ; Step S3.2.3, the impedance matrix After column normalization, we get the first column of the U matrix ,Right now: ; Step S3.2.4: Find the kth row index , making ; Step S3.2.5: The impedance matrix The kth row of the B matrix is obtained by orthogonalizing the first k-1 rows of the B matrix. ,Right now: , , Step S3.2.6: Determine and find the kth column index , making ; Step S3.2.7, the matrix Z The kth column of the U matrix is obtained by orthogonalizing the first k-1 columns of the U matrix. ,Right now: , ; Step S3.2.8: Determine whether the loop is finished. The determination formula is as follows: , in, is the truncation error threshold, which is controlled by The size of controls the size of the cutoff rank k; when The smaller it is, the higher the accuracy is; the larger k is, the more memory is required; If the above judgment formula is satisfied, the loop ends; otherwise, the k value is updated to , repeat steps S3.2.4 to S3.2.8 until the above judgment formula is satisfied; Step S3.2.9: After the loop is completed, the U matrix and the r rows of the original impedance matrix are obtained by solving the following equation: , Get the V matrix.

[0027] Step S3.3: recompress the obtained orthogonal matrices U and V using the singular value decomposition method, discarding some information with very little influence to achieve the purpose of compression. The compression effect is achieved by controlling the size of the truncation precision; The specific steps are as follows: Step S3.3.1: The matrices U and V obtained in step S3.2 are respectively and , perform QR decomposition, expressed as: , , Among them, the superscript represents the dimension of the matrix, and the subscript represents the decomposed matrix. The computational complexity in the above formula is ; Step S3.3.2: Change the two dimensions in step S3.3.1 to Matrix and Multiply to get a new matrix , expressed as: , Step S3.3.3, matrix Perform singular value decomposition and truncate the middle diagonal matrix. The truncation accuracy is , after truncation, the rank is t, and we get: , The computational complexity of steps S3.3.2 and S3.3.3 is ; Step S3.3.4, get the original matrix The approximate expression of is: , After the above steps, the original matrix The memory required is , after compression using the block matrix decomposition technique based on Schmidt orthogonalization, the required memory is reduced to , and then compressed by the singular value decomposition method, its memory consumption is further reduced to ,in , which obviously achieves the goal of reducing computing time and storage resources.

[0028] Step S4: Based on the task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; For example, if a user inputs the task prompt "Analyze the hot research directions and important achievements in the field of artificial intelligence in the past five years," the chief engineer-level expert model will understand the task prompt and conclude that "This is a hot spot analysis task in the scientific research field. The main task is to collect and classify the technical research achievements in the field of artificial intelligence in the past five years (XXXX-XXXX), and summarize the hot research directions. There is no cross-domain situation." Then, the scientific research field agent is generated to call the researcher-level expert model in the scientific research field to continue to perform the task.

[0029] Step S5: The researcher-level expert model understands the received instructions, decomposes the industry-level tasks, generates an industry-level agent to establish a workflow, and issues instructions to call the engineer-level expert model; Continuing with the example in step S4, when the researcher-level expert model in the scientific research field receives the instruction to call the scientific research field agent, it begins to analyze the conclusion prompt word output by the chief engineer-level expert model, "This is a hot spot analysis task in the scientific research field. The main task is to collect and classify the technical research results in the field of artificial intelligence in the past five years (XXXX-XXXX), and summarize the hot spots of technical research. There is no cross-domain situation." The task is concluded to be the "scientific research results search" + "scientific research hot spot analysis" task workflow in the scientific research field, and then generates the scientific research results search agent and the scientific research hot spot analysis agent, and calls the engineer-level expert model respectively to refine the task and continue execution.

[0030] Step S6: The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

[0031] Continuing with the example in step S5, according to the domain-level task workflow, the "Research Results Search" task is first executed. The Research Results Search Agent calls the engineer-level expert model to establish the task workflow and generates a function-level agent to call relevant tools to complete the task, such as: search keyword extraction - content search (academic journal website search, academic forum search, academic news report search, local research results library search) - content integration and clustering - output search content. Next, the "Research Hotspot Analysis" task is executed. The Research Hotspot Analysis Agent calls the engineer-level expert model to establish the task workflow and generates a function-level agent to call relevant tools to complete the task, such as: extracting key information for each research result category (keywords, number of results, number of citations, research funding amount, number of academic experts) - calling data analysis and mining tools to analyze hotspots - generating an analysis report and visualizing it.

[0032] The multi-agent collaborative multi-dimensional knowledge management system is used to implement the above method, including: The tool set module includes agile data acquisition tools, rapid data annotation tools, and data analysis and mining tools, which are used to acquire multi-dimensional knowledge data and perform classification and mining. The model acceleration module is responsible for accelerating the processing of expert models using model training and inference acceleration technology based on block matrix compression decomposition to reduce the dependence on computing resources during model deployment and application; The expert model management module is responsible for establishing a hierarchical hybrid expert architecture for storing and running chief engineer-level expert models, researcher-level expert models, and engineer-level expert models; For task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; The researcher-level expert model understands the instructions it receives, decomposes industry-level tasks, generates industry-level intelligent agents to establish workflows, and issues instructions to invoke the engineer-level expert model. The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

[0033] A multi-dimensional knowledge management device based on multi-agent collaboration includes a memory and a processor; the memory is used to store computer programs, and the processor is used to implement a method when executing the computer programs.

[0034] A readable storage medium stores a computer program, and a method is implemented when the computer program is executed by a processor.

[0035] Compared with existing technologies, this multi-agent collaborative multi-dimensional knowledge management method and system has the following characteristics: 1) Hierarchical hybrid expert architecture design By designing a hierarchical hybrid expert architecture, we build expert models that adapt to the business rules of different industries, support multi-agent collaborative orchestration, and meet complex business needs.

[0036] 2) Knowledge service workflow construction Based on business needs, a knowledge service workflow is built, which realizes the complete process from user demand analysis to multi-agent collaboration, and can provide efficient and accurate knowledge services.

[0037] 3) Model training and inference acceleration technology The use of model training and inference acceleration technology reduces the dependence on computing resources during model deployment and application, and improves system adaptability and response speed.

[0038] 4) Multi-dimensional knowledge intelligent management system Rapidly build a multi-dimensional knowledge intelligent management system with core capabilities of knowledge cognition extraction, labeling classification, mining governance and management services, which can support the operation of industry-customized knowledge bases.

[0039] 5) Artificial intelligence replaces human labor Using artificial intelligence to complete knowledge management services frees humans from tedious processes and significantly improves the efficiency of knowledge management services.

[0040] The above is a detailed introduction to a multi-dimensional knowledge management method and system based on multi-agent collaboration in an example of the present invention. This section uses specific examples to illustrate the principles and implementation methods of the invention. The above examples are only used to help understand the core ideas of the present invention. Without departing from the principles of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts should fall within the scope of protection of the present invention.

Claims

1. A multi-dimensional knowledge management method based on multi-agent collaboration, characterized by: The steps include: Step S1: Build a tool set for data extraction, annotation, and analysis, including a data agile acquisition tool set, a data rapid annotation tool set, and a data analysis and mining tool set to obtain multi-dimensional knowledge data and perform classification and mining; Step S2: Build a hierarchical hybrid expert architecture, dividing the expert models into chief engineer-level expert models, researcher-level expert models, and engineer-level expert models, which are responsible for task decomposition and agent construction and orchestration at different levels respectively; The chief engineer-level expert model is used to understand user instructions, complete domain-level task decomposition, generate domain-level intelligent agents to establish workflows, and call the researcher-level expert model; The researcher-level expert model is used to disassemble the instruction tasks issued by the domain-level intelligent agent, generate the industry-level intelligent agent to establish the workflow, and call the engineer-level expert model; The engineer-level expert model is used to disassemble the instruction tasks issued by the industry-level intelligent agent, generate the function-level intelligent agent to establish the workflow and complete the task; Step S3: Using the model training and inference acceleration technology based on block matrix compression decomposition to accelerate the expert model, so as to reduce the dependence on computing resources during the model deployment and application process; Step S4: Based on the task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; Step S5: The researcher-level expert model understands the received instructions, decomposes the industry-level tasks, generates an industry-level agent to establish a workflow, and issues instructions to call the engineer-level expert model; Step S6: The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

2. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 1, characterized in that: In step S1, the data agile acquisition tool set integrates a face recognition model, an image recognition model, a speech recognition model, and an optical character recognition model, which are used to acquire knowledge data from images, speech, and files to form a knowledge data set, including personnel information data, natural resource data, operation and maintenance data, scientific research data, and organizational relationship data.

3. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 1 is characterized by: In step S1, the data rapid annotation tool set includes an unsupervised clustering tool, a decision tree tool, an SVM classifier, and a neural network classifier; Use unsupervised clustering tools to perform horizontal clustering on the unclassified data in the knowledge dataset, classify the data belonging to different data subjects according to data types, and obtain the number of data types; Decision tree tools, SVM classifiers and / or neural network classifiers are used to vertically classify unclassified data in the knowledge dataset, and data belonging to the same data subject are classified according to data type.

4. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 3 is characterized by: In step S1, the data analysis and mining tool set includes a Gaussian process regression algorithm tool, a correlation analysis algorithm tool, a fuzzy calculation tool, and a deep learning regression algorithm tool, which customize the mining and analysis of the classified data according to the application scenario.

5. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 1 is characterized by: In step S3, the specific process is as follows: Step S3.1: Use the stacking grouping technique to group and decompose the large-dimensional matrix in the expert model into a stacked block matrix structure, where the dimension of the smallest matrix block in the stacked block matrix structure is less than the custom dimension threshold min_n; Step S3.2: Use the block matrix decomposition technique based on Schmidt orthogonalization to transform The non-diagonal block matrix Z is compressed and decomposed into two orthogonal matrices and Multiply, the dimensions are and , , two orthogonal matrices and The required memory size is ; Step S3.3: Use the singular value decomposition method to recompress the obtained orthogonal matrices U and V. The compression effect is achieved by controlling the size of the truncation precision. The specific steps are as follows: Step S3.3.1: The matrices U and V obtained in step S3.2 are respectively and , perform QR decomposition, expressed as: ， ， Among them, the superscript represents the dimension of the matrix, and the subscript represents the decomposed matrix. The computational complexity in the above formula is ; Step S3.3.2: Change the two dimensions in step S3.3.1 to Matrix and Multiply to get a new matrix , expressed as: ， Step S3.3.3, matrix Perform singular value decomposition and truncate the middle diagonal matrix. The truncation accuracy is , after truncation, the rank is t, and we get: ， The computational complexity of steps S3.3.2 and S3.3.3 is ; Step S3.3.4, get the original matrix The approximate expression of is: ， Memory consumption is further reduced to ,in .

6. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 5, characterized in that: In step S3.1, the specific steps are as follows: Step S3.1.1, set the minimum matrix block dimension threshold during the group decomposition process, denoted as min_n; Step S3.1.2: Perform custom row and column splitting of equal dimensions on the matrix to obtain the first-layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.3: Perform custom row and column splitting of the diagonal blocks of the current layer grouping matrix with equal dimensions to obtain the next layer grouping matrix. Determine whether the dimension of each matrix block is less than the dimension threshold min_n. If so, stop the operation; otherwise, proceed to the next step. Step S3.1.4: Repeat step S3.1.3 until the minimum matrix block dimension is less than the dimension threshold min_n, and then stop the operation.

7. The multi-dimensional knowledge management method based on multi-agent collaboration according to claim 5 is characterized by: In step S3.2, the specific steps are as follows: Step S3.2.

1. Normalize the first row of matrix Z and assign it to the first row of matrix B. ,Right now: , row index ; Step S3.2.2: Determine and find the first column index , making ; Step S3.2.3, the impedance matrix After column normalization, we get the first column of the U matrix ,Right now: ； Step S3.2.4: Find the kth row index , making ; Step S3.2.5: The impedance matrix The kth row of the B matrix is obtained by orthogonalizing the first k-1 rows of the B matrix. ,Right now: ， ， Step S3.2.6: Determine and find the kth column index , making ; Step S3.2.7, the matrix Z The kth column of the U matrix is obtained by orthogonalizing the first k-1 columns of the U matrix. ,Right now: ， ； Step S3.2.8: Determine whether the loop is finished. The determination formula is as follows: ， in, is the truncation error threshold, which is controlled by The size of controls the size of the cutoff rank k; when The smaller it is, the higher the accuracy is; the larger k is, the more memory is required; If the above judgment formula is satisfied, the loop ends; otherwise, the k value is updated to , repeat steps S3.2.4 to S3.2.8 until the above judgment formula is satisfied; Step S3.2.9: After the loop is completed, the U matrix and the r rows of the original impedance matrix are obtained by solving the following equation: ， Get the V matrix.

8. A multi-dimensional knowledge management system based on multi-agent collaboration, characterized by: Used to implement the method according to any one of claims 1 to 7, comprising: The tool set module includes agile data acquisition tools, rapid data annotation tools, and data analysis and mining tools, which are used to acquire multi-dimensional knowledge data and perform classification and mining. The model acceleration module is responsible for accelerating the processing of expert models using model training and inference acceleration technology based on block matrix compression decomposition to reduce the dependence on computing resources during model deployment and application; The expert model management module is responsible for establishing a hierarchical hybrid expert architecture for storing and running chief engineer-level expert models, researcher-level expert models, and engineer-level expert models; For task input, the chief engineer-level expert model is used to understand the input task instructions, complete the domain-level task decomposition, generate the domain-level intelligent agent to establish the workflow, and issue instructions to call the researcher-level expert model; The researcher-level expert model understands the instructions it receives, decomposes industry-level tasks, generates industry-level intelligent agents to establish workflows, and issues instructions to invoke the engineer-level expert model. The engineer-level expert model understands the input task instructions, completes the functional-level task decomposition, generates a functional-level intelligent agent to establish a workflow, and calls the corresponding tools to complete the task.

9. A multi-dimensional knowledge management device based on multi-agent collaboration, characterized by: The method comprises a memory and a processor; the memory is used to store a computer program, and the processor is used to implement the method according to any one of claims 1 to 7 when executing the computer program.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images