A heterogeneous multi-agent collaborative diagnostic system and method for CNC machine tool operation and maintenance

Abstract

This invention provides a heterogeneous multi-agent collaborative diagnostic system and method for CNC machine tool operation and maintenance. It includes a heterogeneous agent scheduling engine that dynamically calls and manages workflows and data flows between specialized agent capabilities based on the type of input data and the task processing stage; specialized agent capabilities built based on a large language model; a knowledge graph database for storing and managing the entity attributes and semantic relationships of CNC machine tools; and corresponding diagnostic methods based on the diagnostic system. By constructing multiple specialized agents, the system configures optimal models for different diagnostic sub-tasks and achieves human-machine collaborative feedback and knowledge graph self-evolution, thereby significantly improving the accuracy, interpretability, and automation level of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent operation and maintenance of numerical control machine tools, and particularly relates to a heterogeneous multi-agent collaborative diagnosis system and method for operation and maintenance of numerical control machine tools. BACKGROUND

[0002] Current fault diagnosis of numerical control machine tools mainly relies on experienced engineers or rule-based expert systems, but faces problems such as high labor cost, incomplete rules, and difficulty in maintaining expert systems. In recent years, with the development of large language models (LLM), some studies have attempted to combine LLM with knowledge graphs and fault networks for machine tool fault reasoning, or use multi-modal large models to analyze vibration, image and other signals. However, the existing technology still has the following significant defects: the model lacks generalization ability, the general large model lacks deep understanding of the terms and professional diagnosis logic in the field of numerical control machine tools, resulting in low diagnosis accuracy; task processing is single, most solutions use a single LLM to process data cleaning, feature description, knowledge query and other subtasks, which cannot match the optimal model for different task characteristics, causing resource waste and performance bottleneck; system capability is fixed, knowledge graphs and large models are mostly pre-constructed, and the system is not updated after going online, making it difficult to automatically learn new knowledge from daily maintenance conversations and work order records and dynamically update, resulting in the system being unable to adapt to new fault modes; human-computer interaction is shallow, existing systems mostly use "question and answer" interaction, lack of counter-question clarification mechanism and feedback learning mechanism, and cannot form a closed loop of diagnosis-feedback-optimization.

[0003] Therefore, there is an urgent need for an intelligent diagnosis system that can deeply integrate field knowledge, adapt to multi-modal input, support professional role division, and have continuous evolution capability. SUMMARY

[0004] To solve the above problems, the application provides a heterogeneous multi-agent collaborative diagnosis system and method for operation and maintenance of numerical control machine tools, which configures optimal models for different diagnosis subtasks by constructing multiple specialized agents, and realizes human-machine collaborative feedback and knowledge graph self-evolution, thereby significantly improving the accuracy, explainability and automation level of fault diagnosis.

[0005] The specific scheme is as follows:

[0006] A heterogeneous multi-agent collaborative diagnosis system for operation and maintenance of numerical control machine tools, the system comprising: a heterogeneous agent scheduling engine, a special capability agent, and a knowledge graph database, the heterogeneous agent scheduling engine being used to dynamically call and manage the workflow and data flow between special capability agents according to the type of input data and the task processing stage; the special capability agent being constructed based on a large language model, comprising:

[0007] A data engineer agent for parsing instructions, pre-processing multi-source data, and outputting structured data summaries;

[0008] A multi-modal analysis agent for analyzing vibration spectrograms, thermal imaging diagrams, and abnormal sound audio, generating natural language descriptions of mechanical fault characteristics, and extracting key quantitative indicators;

[0009] A knowledge graph agent for querying related information from a pre-built knowledge graph database and updating the knowledge graph database;

[0010] A causal reasoning agent for integrating the output results of the data engineer agent, the knowledge graph agent, and the multi-modal analysis agent, inferring fault propagation paths, and generating an interpretable causal graph with confidence;

[0011] A human-machine collaborative agent for interacting with engineers and integrating the analysis results of the various specialized capability agents to generate a diagnostic report;

[0012] The knowledge graph database is used to store and manage the entity attributes and semantic relationships of the numerical control machine tool.

[0013] Further, the data engineer agent is based on a Coder model, calls Python commands through a Coder model context protocol (MCP), and runs in a container through an MCP server to clean, align, and feature engineer multi-source data, and output structured data summaries.

[0014] Further, the multi-modal analysis agent uses a multi-modal large model fine-tuned based on industrial sensor data, divides images or spectrograms into sub-blocks, maps the sub-blocks to visual tokens through a position encoding method containing rotation operations, and inputs the tokens into a large language model after splicing with text tokens for processing.

[0015] Further, the knowledge graph agent queries related information from a pre-built knowledge graph database, including component relationships related to the current fault phenomenon, historical maintenance cases, and spare parts information, extracts entities and relationships from the feedback maintenance text, and updates the knowledge graph database after confirmation.

[0016] Further, the causal reasoning agent retrieves fault analysis knowledge from the knowledge graph database, integrates the output results of the data engineer agent, the knowledge graph agent, and the multi-modal analysis agent, infers fault propagation paths based on a causal reasoning optimization model, and generates an interpretable causal graph with confidence.

[0017] Furthermore, the human-machine collaborative intelligent agent interacts with the engineer and integrates the analysis results of various specialized intelligent agents to generate a diagnostic report that includes fault location, cause analysis, maintenance suggestions, and operation instructions. When generating maintenance suggestions, the report is classified according to the severity of the fault and includes graphic or video operation instructions.

[0018] Furthermore, the heterogeneous agent scheduling engine transforms engineer feedback data into training samples, and trains and adjusts them through online learning of the dedicated models corresponding to each specialized capability agent. This online learning employs low-rank adaptive technology for parameter fine-tuning; during the fine-tuning phase, the original weights of the large model are frozen. Only two low-rank factorization matrices are trained. and Through matrix Multiplication represents the parameter update matrix ,in, For the input vector dimension, For the output vector dimension, To fine-tune the settings, take The rank of the product of two matrices The adjusted weights are expressed as follows:

[0019] .

[0020] The present invention also provides a heterogeneous multi-agent collaborative diagnosis method for CNC machine tool operation and maintenance. The method is based on the above-mentioned heterogeneous multi-agent collaborative diagnosis system for CNC machine tool operation and maintenance, and includes the following steps;

[0021] S1. Receive multi-source heterogeneous operation and maintenance data of CNC machine tools and user-input descriptions of fault phenomena;

[0022] S2. Clarify and define the fault by interacting with the user through a human-machine collaborative intelligent agent;

[0023] Employing a multi-turn dialogue intent recognition mechanism, and optimizing interaction efficiency based on text similarity calculation, the system sets up user input questions. After word segmentation, in dimension The TF-IDF vectors under the unified vocabulary are ,in, Indicates the first in the vocabulary list The term in the question The TF-IDF weight in the text, i.e., the weight of the term in the text. The product of the word frequency in the text and its inverse document frequency in the entire standard question template corpus; similarly, the product of the frequency of each template in the standard question template corpus. The TF-IDF vector is ,in, Indicates the first in the vocabulary list individual terms in template issues The TF-IDF weights in the above vector representation are used to calculate the semantic matching degree through cosine similarity:

[0024] ;

[0025] Based on preset threshold Human-machine collaborative intelligent agents take priority Furthermore, the standard template with the highest matching degree is selected. If no matching result is found, multiple rounds of questioning are used to clarify the details of the fault until a standard problem template is successfully matched.

[0026] S3. The heterogeneous intelligent agent scheduling engine calls the data engineer intelligent agent to perform cleaning, alignment and feature engineering according to the input data type, and outputs vibration spectrum and structured data summary.

[0027] S4. Analyze the vibration spectrum and workpiece image through a multimodal analysis agent to generate a natural language description of the fault characteristics;

[0028] S5. The knowledge graph agent is invoked to query relevant historical repair cases, component relationships, and spare parts information from the knowledge graph database based on the fault phenomenon description and natural language description of fault characteristics; when the knowledge graph is updated, new entity relationships are extracted. The confidence level is:

[0029] ;

[0030] in, For entities that already exist in the knowledge graph. For entity similarity function, Let r be the historical frequency of occurrence. The weighting coefficients for entity similarity and historical frequency of occurrence satisfy the following conditions: ;

[0031] S6. Call the causal reasoning agent to fuse the output results of the data engineer agent, knowledge graph agent and multimodal analysis agent, infer the fault propagation path and generate an interpretable causal graph with confidence.

[0032] The causal confidence level from faulty node A to node B is:

[0033] ;

[0034] in, Let A and B be the co-occurrence probability. For historical prior probability, For smoothing coefficients;

[0035] In fault propagation path reasoning, let the fault propagation path be... The total confidence of the path is:

[0036] ;

[0037] in, The probability of initial failure occurrence is inferred jointly from the knowledge graph and real-time data;

[0038] S7. By integrating and analyzing the results through a human-machine collaborative intelligent agent, a diagnostic report is generated that includes fault location, cause analysis, maintenance suggestions, and operation instructions.

[0039] Furthermore, in step S4, the multimodal analysis agent analyzes the vibration spectrum and the finished workpiece image, dividing them into 16*16 pixel blocks. Each pixel block is mapped to a vector of uniform length through a linear layer. , where the vector dimension d is a preset model hyperparameter;

[0040] For the two-dimensional position information of pixel blocks ,Will The space is divided into several four-dimensional subspaces, with two consecutive dimensions forming a group. After a rotation operation, the position-encoded vector is obtained as follows: ;

[0041] in, ;

[0042] The location-encoded vector is input into the Transformer model to obtain multiple visual tokens with the same dimension as the text tokens. The text token is concatenated with the text token and input into the large language model. The large language model integrates the text token with the visual token of the vibration spectrum and outputs a text description of the key frequency components. After identifying the key processing quality features in the finished workpiece image, it outputs a natural language description.

[0043] Furthermore, it also includes: step S8, whereby the knowledge graph agent extracts new entities and relationships from the feedback maintenance text, updates the knowledge graph database, and fine-tunes the dedicated models corresponding to each dedicated capability agent using the feedback data through the online learning module.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. High diagnostic accuracy and level of specialization;

[0046] By adopting a heterogeneous multi-agent collaborative architecture, dedicated, domain-fine-tuned large language models are configured for different sub-tasks such as data preprocessing, multimodal analysis, knowledge query, causal reasoning, and human-computer interaction. This overcomes the problems of insufficient domain knowledge and bias in professional logic understanding that exist when a single general-purpose large model is used to deal with complex industrial scenarios. It can give full play to the performance advantages of each dedicated model in its area of ​​expertise, thereby significantly improving the accuracy of fault location and the precision of feature description.

[0047] 2. Possesses continuous evolution and self-adaptation capabilities;

[0048] This invention innovatively integrates dynamic knowledge graph updates with an online learning mechanism based on low-rank adaptive learning (LoRA). It automatically extracts new knowledge from each human-computer interaction and engineer feedback to update the knowledge graph, and efficiently fine-tunes the model parameters of each specialized agent using feedback data. This forms a closed loop of "diagnosis → feedback → learning → optimization," enabling the system to continuously accumulate operational experience, adapt to new failure modes, and effectively solve the pain points of traditional expert systems and static models, such as knowledge solidification and difficulty in maintenance, thus giving the system long-term vitality.

[0049] 3. The diagnostic results are highly interpretable and practical;

[0050] The system's final output goes beyond simple fault conclusions. The causal reasoning agent generates an interpretable causal graph with confidence, clearly demonstrating the propagation path and causes of the fault, answering the question "why it broke." At the same time, the diagnostic report generated by the human-machine collaborative agent includes maintenance suggestions that are automatically graded according to the severity of the fault, as well as specific graphic or video operation instructions, directly guiding on-site maintenance work and answering the question "how to fix it," greatly enhancing the practical value and operability of the diagnostic results in engineering practice.

[0051] 4. Significantly improved automation and operational efficiency;

[0052] This invention achieves a high degree of automation throughout the entire process from data access to report generation. The data engineer agent completes the tedious data preprocessing work by automatically generating and executing code. The heterogeneous agent scheduling engine enables seamless collaboration and task flow between agents, greatly reducing reliance on senior engineers, lowering labor costs, shortening the time for fault diagnosis and decision-making, enabling predictive maintenance and rapid response of CNC machine tools, and improving overall operation and maintenance efficiency. Attached Figure Description

[0053] Figure 1 This is a diagram of the overall architecture of the diagnostic system of the present invention;

[0054] Figure 2 This is a schematic diagram of the diagnostic process of the present invention.

[0055] The components in the attached diagram are labeled as follows: Multi-source data access module 1, Heterogeneous intelligent agent scheduling engine 2, Specialized capability intelligent agent 3, Data engineer intelligent agent 31, Multimodal analysis intelligent agent 32, Knowledge graph intelligent agent 33, Causal reasoning intelligent agent 34, Human-machine collaborative intelligent agent 35, Knowledge graph database 4, Online learning module 5, Human-computer interaction interface 6. Detailed Implementation

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1:

[0058] like Figure 1 Figure 2 As shown, this embodiment provides a heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance. The system includes:

[0059] The system includes: 1. Multi-source data access module; 2. Heterogeneous intelligent agent scheduling engine; 3. Specialized capability intelligent agent; 4. Knowledge graph database; and 6. Human-computer interaction interface.

[0060] Heterogeneous agent scheduling engine 2 is used to dynamically invoke and manage workflows and data flows between specialized agent capabilities based on the type of input data and the stage of task processing.

[0061] Specialized capability agent 3 is built based on a large language model, including;

[0062] Data Engineer Agent 31 is based on dedicated Coder models such as DeepSeek-Coder and Qwen-Coder series. The Coder model calls Python commands through MCP, and is run in a container by the MCP server to clean, align, and feature-engineer multi-source data, outputting structured data summaries. To improve the success rate of tool calls, it pre-installs commonly used Python toolkits for industrial data analysis such as pandas, scipy, and sklearn, and encapsulates commonly used tools such as Fast Fourier Transform (FFT), Wavelet Transform (WT), Short Time Fourier Transform (STFT), and Order Tracking (OT) into standardized function interfaces.

[0063] The multimodal analysis agent 32 employs large multimodal models such as Qwen-VL and LLaVA, which are fine-tuned based on industrial sensor data. It divides images or spectrograms into several sub-blocks, maps the sub-blocks to visual tokens through a position encoding method that includes rotation operations, and concatenates them with text tokens before inputting them into a large language model for processing. It is used to analyze vibration spectrograms, thermal imaging images, and abnormal noise audio, generate natural language descriptions of mechanical fault characteristics, and extract key quantitative indicators.

[0064] The knowledge graph agent 33 is based on the Llama3 or Qwen series and uses CNC machine tool maintenance manuals and SPARQL query corpora for instruction fine-tuning. It is used to query relevant information from a pre-built knowledge graph database, including component relationships related to the current fault phenomenon, historical maintenance cases and spare parts information. At the same time, it extracts entities and relationships from the feedback maintenance text and updates the knowledge graph database after confirmation.

[0065] The causal reasoning agent 34 adopts a model based on Qwen-Max, fine-tuned using Fault Tree Analysis (FTA) and Failure Mode and Effects Analysis (FMEA) documents. It combines Retrieval-Augmented Generation (RAG) technology to retrieve fault analysis knowledge from the knowledge graph database 4, and integrates the output results of the data engineer agent 31, knowledge graph agent 33, and multimodal analysis agent 32. Based on the causal reasoning optimization model, it infers the fault propagation path and generates an interpretable causal graph with confidence.

[0066] The human-machine collaborative intelligent agent 35 interacts with the engineer through the human-machine interaction interface 6, receives questions from the human engineer, and actively communicates with the engineer through multiple rounds of dialogue to clarify ambiguous information. It also integrates the analysis results of various specialized intelligent agents 3 to generate a diagnostic report that includes fault location, cause analysis, maintenance suggestions and operation instructions. When generating maintenance suggestions, it classifies them according to the severity of the fault and includes graphic or video operation instructions.

[0067] Knowledge graph database 4 is used to store and manage the entity attributes and semantic relationships of CNC machine tools.

[0068] The heterogeneous agent scheduling engine also transforms engineer feedback data into training samples, which are then trained and adjusted through online learning of the dedicated models corresponding to each specialized capability agent. This online learning employs low-rank adaptive techniques for parameter fine-tuning; during the fine-tuning phase, the original weights of the large model are frozen. Only two low-rank factorization matrices are trained. and Through matrix Multiplication represents the parameter update matrix ,in, For the input vector dimension, For the output vector dimension, To fine-tune the settings, take The rank of the product of two matrices The adjusted weights are expressed as follows:

[0069] .

[0070] Example 2:

[0071] This embodiment provides a heterogeneous multi-agent collaborative diagnostic method for CNC machine tool operation and maintenance. Based on the diagnostic system in Embodiment 1, the method takes machining with burrs and abnormal spindle noise as an example, and includes the following steps:

[0072] S1. Receive multi-source heterogeneous operation and maintenance data of CNC machine tools and user-input descriptions of fault phenomena;

[0073] In this embodiment, the multi-source heterogeneous operation and maintenance data includes text logs, time-series vibration data (sampling rate 10kHz, lasting 5 minutes), machine tool operation status text logs, and images of completed workpieces (showing severe burrs on the edge of the holes).

[0074] S2. The human-machine collaborative intelligent agent interacts with the user. The user inputs the command "There are burrs in the machining and abnormal noise from the spindle. What is the reason?" and uploads relevant vibration data, machine tool operation status text log and finished workpiece image. The human-machine collaborative intelligent agent 35 receives and organizes the information provided by the user and actively asks the user until the fault phenomenon and problem description are clear enough to clarify and identify the fault problem.

[0075] S3. The heterogeneous intelligent agent scheduling engine 2 calls the data engineer intelligent agent 31 to write Python code according to the input data type. The request is constructed through the MCP protocol. The MCP server parses the request and runs the Python code in the container to perform cleaning, alignment and feature engineering. The root mean square (RMS) and spectral kurtosis index of the stable processing stage are extracted to create a vibration spectrum. The MCP server returns the program running results in JSON (JavaScript Object Notation) format, outputting the vibration spectrum and structured data summary.

[0076] S4. The multimodal analysis agent 32 analyzes the vibration spectrum and workpiece image to generate a natural language description of the fault characteristics.

[0077] The multimodal analysis agent 32 analyzes the vibration spectrum and the image of the finished workpiece, dividing them into 16*16 pixel blocks. Each pixel block is mapped to a vector of uniform length through a linear layer. , where the vector dimension d is a preset model hyperparameter;

[0078] For the two-dimensional position information of pixel blocks ,Will The space is divided into several four-dimensional subspaces, with two consecutive dimensions forming a group. After a rotation operation, the position-encoded vector is obtained as follows: ;

[0079] in, ;

[0080] The location-encoded vector is input into the Transformer model to obtain multiple visual tokens with the same dimension as the text tokens. The text token is concatenated with the text token and input into the large language model. The large language model integrates the text token with the visual tokens from the vibration spectrum diagram and the finished workpiece image, and outputs a text description of the key frequency components: "The vibration spectrum shows a significant peak at 2310 Hz, which is consistent with the fault characteristic frequency of the outer ring of the angular contact ball bearing (theoretical value 2280 Hz)". Then, it identifies the key processing quality features in the finished workpiece image and outputs a natural language description: "The photo of the processed part shows that there are periodic burrs on the edge of the hole, which are synchronized with the spindle rotation."

[0081] S5. The knowledge graph agent 33 is invoked to query relevant historical maintenance cases, component relationships, and spare parts information from the knowledge graph database 4 based on the natural language description of the fault phenomenon and fault characteristics. The knowledge graph agent 33 queries the Neo4j database for related cases and maintenance steps in the knowledge graph database 4: "Related fault: [spindle bearing outer ring wear]; maintenance case: 2024-05-10, #M302 machine tool, replace SKF 6205-2RS bearing, downtime 2.5 hours; spare part model: SKF 6205-2RS".

[0082] S6. Call the output results of the causal reasoning agent 34, the data engineer agent 31, the knowledge graph agent 33, and the multimodal analysis agent 32 to infer the fault propagation path and generate an interpretable causal graph with confidence: "Insufficient lubrication → 0.92 → Wear of the outer ring of the spindle bearing → 0.89 → Increased vibration (peak value of 2310Hz) → 0.85 → Decreased dynamic stiffness of the spindle → 0.8 → Increased positional deviation during cutting → 0.78 → Out-of-tolerance machining hole diameter (burrs)".

[0083] S7. By integrating and analyzing the results through the human-machine collaborative intelligent agent 35, a diagnostic report is generated, including fault location, cause analysis, maintenance suggestions, and operation instructions, as follows:

[0084] "Fault location: Wear on the outer ring of the spindle front bearing (model SKF 6205-2RS). Root cause: Insufficient lubrication due to excessively long lubrication cycles. Repair recommendation: [Immediately stop the machine] Replace the bearing and check the lubrication system piping."

[0085] Step S8: The knowledge graph agent 33 extracts new entities and relationships from the feedback maintenance text, updates the knowledge graph database 4, and fine-tunes the dedicated models corresponding to each dedicated capability agent 3 using the feedback data through the online learning module 5.

[0086] The engineer added that "the root cause is the blockage of the lubrication pump filter." The online learning module 5 transformed the original multi-source heterogeneous operation and maintenance data and feedback information in this case into training samples. The knowledge graph agent 33 extracted new entities and relationships from the feedback information of the human engineer: "blockage of the lubrication pump filter → leads to insufficient lubrication," and updated the knowledge graph database 4.

[0087] The above embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Although the invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of the invention do not depart from the spirit and scope of the invention and should be covered within the scope of the claims.

Claims

1. A heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance, characterized in that, The system includes: a heterogeneous agent scheduling engine, specialized capability agents, and a knowledge graph database; The heterogeneous intelligent agent scheduling engine is used to dynamically call and manage the workflow and data flow between intelligent agents with special capabilities according to the type of input data and the stage of task processing; The specialized capability agent described above is constructed based on a large language model and includes: The data engineer agent is used to parse instructions, preprocess multi-source data, and output structured data summaries. A multimodal analysis agent is used to analyze vibration spectrum diagrams, thermal imaging diagrams, and abnormal noise audio, generate natural language descriptions of mechanical fault characteristics, and extract key quantitative indicators. Knowledge graph agents are used to query relevant information from a pre-built knowledge graph database and update the knowledge graph database. A causal reasoning agent is used to fuse the outputs of a data engineer agent, a knowledge graph agent, and a multimodal analysis agent to infer fault propagation paths and generate an interpretable causal graph with confidence. A human-machine collaborative intelligent agent is used to interact with engineers and integrate the analysis results of various specialized intelligent agents to generate diagnostic reports; The knowledge graph database is used to store and manage the entity attributes and semantic relationships of CNC machine tools.

2. The heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 1, characterized in that, The data engineer agent is based on the Coder model. It calls Python commands through the Coder model context protocol and runs in a container on the MCP server. It cleans, aligns, and performs feature engineering on multi-source data, and outputs a structured data summary.

3. The heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 2, characterized in that, The multimodal analysis agent adopts a large multimodal model based on industrial sensor data for fine-tuning. It divides the image or spectrogram into several sub-blocks, maps the sub-blocks into visual tokens through a position encoding method that includes rotation operations, and concatenates them with text tokens before inputting them into a large language model for processing.

4. A heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 3, characterized in that, The knowledge graph agent queries relevant information from a pre-built knowledge graph database, including component relationships related to the current fault phenomenon, historical repair cases, and spare parts information. At the same time, it extracts entities and relationships from the feedback repair text and updates the knowledge graph database after confirmation.

5. A heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 4, characterized in that, The causal reasoning agent retrieves fault analysis knowledge from the knowledge graph database and integrates the outputs of the data engineer agent, the knowledge graph agent, and the multimodal analysis agent. Based on the causal reasoning optimization model, it infers the fault propagation path and generates an interpretable causal graph with confidence.

6. A heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 5, characterized in that, The human-machine collaborative intelligent agent interacts with the engineer and integrates the analysis results of various specialized intelligent agents to generate a diagnostic report that includes fault location, cause analysis, maintenance suggestions and operation instructions. When generating maintenance suggestions, the report is classified according to the severity of the fault and includes graphic or video operation instructions.

7. A heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance according to claim 6, characterized in that, The heterogeneous agent scheduling engine also transforms the feedback data from engineers into training samples, and trains and adjusts them through online learning of the dedicated models corresponding to each dedicated capability agent. The online learning uses low-rank adaptive technology for parameter fine-tuning. Freeze the original weights of the large model during the fine-tuning phase. Only two low-rank factorization matrices are trained. and Through matrix Multiplication represents the parameter update matrix ,in, For the input vector dimension, For the output vector dimension, To fine-tune the settings, take The rank of the product of two matrices The adjusted weights are expressed as follows: 。 8. A heterogeneous multi-agent collaborative diagnostic method for CNC machine tool operation and maintenance, characterized in that, This method, based on the heterogeneous multi-agent collaborative diagnostic system for CNC machine tool operation and maintenance as described in any one of claims 1 to 7, includes the following steps: S1. Receive multi-source heterogeneous operation and maintenance data of CNC machine tools and user-input descriptions of fault phenomena; S2. Clarify and define the fault by interacting with the user through a human-machine collaborative intelligent agent; Employing a multi-turn dialogue intent recognition mechanism, and optimizing interaction efficiency based on text similarity calculation, the system sets up user input questions. After word segmentation, in dimension The TF-IDF vectors under the unified vocabulary are ,in, Indicates the first in the vocabulary list The term in the question The TF-IDF weight in the text, i.e., the weight of the term in the text. The product of the word frequency in the text and its inverse document frequency in the entire standard question template corpus; similarly, the product of the frequency of each template in the standard question template corpus. The TF-IDF vector is ,in, Indicates the first in the vocabulary list individual terms in template issues The TF-IDF weights in the above vector representation are used to calculate the semantic matching degree through cosine similarity: ； Based on preset threshold Human-machine collaborative intelligent agents take priority Furthermore, the standard template with the highest matching degree is selected. If no matching result is found, multiple rounds of questioning are used to clarify the details of the fault until a standard problem template is successfully matched. S3. The heterogeneous intelligent agent scheduling engine calls the data engineer intelligent agent to perform cleaning, alignment and feature engineering according to the input data type, and outputs vibration spectrum and structured data summary. S4. Analyze the vibration spectrum and workpiece image through a multimodal analysis agent to generate a natural language description of the fault characteristics; S5. The knowledge graph agent is invoked to query relevant historical repair cases, component relationships, and spare parts information from the knowledge graph database based on the fault phenomenon description and natural language description of fault characteristics; when the knowledge graph is updated, new entity relationships are extracted. The confidence level is: ； in, For entities that already exist in the knowledge graph. For entity similarity function, Let r be the historical frequency of occurrence. The weighting coefficients for entity similarity and historical frequency of occurrence satisfy the following conditions: ; S6. Call the causal reasoning agent to fuse the output results of the data engineer agent, knowledge graph agent and multimodal analysis agent, infer the fault propagation path and generate an interpretable causal graph with confidence. The causal confidence level from faulty node A to node B is: ； in, Let A and B be the co-occurrence probability. For historical prior probability, For smoothing coefficients; In fault propagation path reasoning, let the fault propagation path be... The total confidence of the path is: ； in, The probability of initial failure occurrence is inferred jointly from the knowledge graph and real-time data; S7. By integrating and analyzing the results through a human-machine collaborative intelligent agent, a diagnostic report is generated that includes fault location, cause analysis, maintenance suggestions, and operation instructions.

9. A heterogeneous multi-agent collaborative diagnosis method for CNC machine tool operation and maintenance according to claim 8, characterized in that, In step S4, the multimodal analysis agent analyzes the vibration spectrum and the image of the finished workpiece, dividing them into 16*16 pixel blocks. Each pixel block is mapped to a vector of uniform length through a linear layer. , where the vector dimension d is a preset model hyperparameter; For the two-dimensional position information of pixel blocks ,Will The space is divided into several four-dimensional subspaces, with two consecutive dimensions forming a group. After a rotation operation, the position-encoded vector is obtained as follows: ; in, ; The location-encoded vector is input into the Transformer model to obtain multiple visual tokens with the same dimension as the text tokens. The text token is concatenated with the text token and input into the large language model. The large language model integrates the text token with the visual token of the vibration spectrum and outputs a text description of the key frequency components. After identifying the key processing quality features in the workpiece image, it outputs a natural language description.

10. A heterogeneous multi-agent collaborative diagnostic method for CNC machine tool operation and maintenance according to claim 9, characterized in that, It also includes: step S8, where the knowledge graph agent extracts new entities and relationships from the feedback maintenance text, updates the knowledge graph database, and fine-tunes the dedicated models corresponding to each dedicated capability agent using the feedback data through the online learning module.

Images