Numerical control system fault diagnosis method and system based on knowledge injection

Abstract

The invention belongs to the technical field of fault diagnosis, and particularly discloses a numerical control system fault diagnosis method and system based on knowledge injection, and the method comprises the steps: carrying out the fault diagnosis based on an RAG frame, the RAG frame comprises a knowledge base and a large model, the knowledge base is a knowledge graph constructed based on the fault diagnosis historical data of a numerical control system, and the large model is a knowledge graph constructed based on the fault diagnosis historical data of the numerical control system; retrieving a fault path related to the user question from the knowledge base, and then comprehensively retrieving the fault path and the user question by a large model to generate a fault diagnosis result; after each fault diagnosis, judging whether a user problem is solved, and if the user problem is solved, directly adding a fault path into a feedback database; if not, adding the problem fed back by the engineer into a feedback database in combination with the fault path and the feedback knowledge of the engineer; and performing knowledge injection on the large model through the data in the feedback database, and updating the knowledge graph. According to the method, continuous learning optimization of a large model and a knowledge base can be realized, and the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and more specifically, relates to a numerical control system fault diagnosis method and system based on knowledge injection. Background Art

[0002] At present, with the advent of the era of general artificial intelligence, artificial intelligence big model technology is empowering all walks of life. In the manufacturing industry, especially in the field of CNC system fault diagnosis, the RAG (Retriever-Augmented Generation) framework has gradually attracted attention and has been initially applied due to its advantages in reducing model training costs and improving answer accuracy and interpretability.

[0003] Traditional RAG knowledge bases are mainly composed of document bases. However, in the manufacturing industry, knowledge is incomplete, dynamically updated, and dependent on expert experience. Therefore, the knowledge base needs to be constantly learned and updated to inject new knowledge and integrate expert knowledge and experience. However, unstructured text-based knowledge bases have many limitations: their knowledge updates are usually carried out in the form of large paragraphs of text, resulting in insufficient content refinement and high repetition; at the same time, document-based data is difficult to effectively associate with experiential knowledge and difficult to support deep interaction with experts. Therefore, the construction of knowledge bases in the manufacturing industry needs to break through the limitations of traditional document bases and build knowledge bases that can be dynamically updated and deeply integrated with expert experience.

[0004] In addition, in traditional methods, the knowledge base is only used as a retrieval tool to help the large model understand the context, but the large model itself does not have a deep understanding and learning of the knowledge base content. This leads to excessive reliance on the quality of the retrieval results during the generation process, which is prone to biased understanding of the knowledge content, thus affecting the accuracy and consistency of the answers. Therefore, the lack of effective integration and complementarity between the knowledge base and the large model has become a bottleneck limiting the accuracy of the system. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a CNC system fault diagnosis method and system based on knowledge injection, which aims to achieve continuous learning and optimization of large models and knowledge bases and improve the accuracy of fault diagnosis.

[0006] To achieve the above object, according to one aspect of the present invention, a method for fault diagnosis of a numerical control system based on knowledge injection is proposed, comprising the following steps:

[0007] Fault diagnosis is carried out based on the RAG framework, which includes a knowledge base and a large model. Among them, the knowledge base is a knowledge graph constructed based on the historical data of numerical control system fault diagnosis. Fault paths related to the user's problem are retrieved from the knowledge base, and then the large model synthesizes the retrieved fault paths and the user's problem to generate a fault diagnosis result;

[0008] After each fault diagnosis, it is judged whether the user's problem is solved. If it is solved, the fault path is directly added to the feedback database; if it is not solved, the problems feedback by engineers, combined with the fault path and the knowledge feedback by engineers, are added to the feedback database;

[0009] Knowledge injection is performed on the large model through the data in the feedback database, and the knowledge graph is updated.

[0010] As a further preference, the fault path includes: numerical control model, alarm number, alarm phenomenon, cause, and solution.

[0011] As a further preference, for the data in the feedback database, if there is no feedback from engineers, the data is divided into the enhanced training dataset; if there is feedback from engineers, the data is divided into the optimized training dataset.

[0012] As a further preference, knowledge injection into the large model through the data in the feedback database includes:

[0013] For the data in the enhanced training dataset, the LoRA fine-tuning method is used to perform knowledge injection on the large model in the RAG framework;

[0014] For the data in the optimized training dataset, the In-context learning method is used to combine the fault path and the knowledge feedback by engineers to obtain an optimized fault path; based on the optimized fault path, the LoRA fine-tuning method is used to perform knowledge injection on the large model, and the knowledge graph is updated.

[0015] As a further preference, for the unsolved user problems, the user subsequently solves the problems through the way of expert remote consultation or engineer on-site operation and maintenance, and finally obtains a work order automatically; based on the work order, a new fault path is newly constructed through the named entity recognition method, and the newly added fault path is added to the feedback database and divided into the enhanced training dataset.

[0016] As a further preference, updating the knowledge graph through the data in the feedback database includes:

[0017] For each newly added and optimized fault path, first accurately match the numerical control model in the knowledge graph according to the fault path. If the match fails, a new numerical control model is added, and the obtained numerical control model is used as the starting node of the path;

[0018] Then, starting from the starting node, sequentially match the alarm number, alarm phenomenon, cause, and solution nodes in the knowledge graph; among them, exact matching is used for the alarm number, and vectorized fuzzy matching is used for the alarm phenomenon, cause, and solution; during this process, if a certain node fails to match successfully, a new path is constructed starting from the previous node and added to the knowledge graph to achieve an update.

[0019] As a further preference, set double thresholds of quantity and time for the feedback database. When the data volume in the feedback database reaches the preset quantity threshold, or the time since the last data processing exceeds the preset time threshold, extract the data in the feedback database to perform knowledge injection into the large model.

[0020] According to another aspect of the present invention, a numerical control system fault diagnosis system based on knowledge injection is provided, including a RAG framework and a learning and updating module, wherein:

[0021] The RAG framework includes a knowledge base and a large model, wherein the knowledge base is a knowledge graph constructed based on the historical data of numerical control system fault diagnosis, and the knowledge base is used to retrieve the fault path related to the user's problem; the large model is used to generate a fault diagnosis result by synthesizing the retrieved fault path and the user's problem;

[0022] The learning and updating module is used to update the large model and the knowledge graph according to user feedback, including: if the user feedbacks that the problem has been solved, directly add the fault path to the feedback database; if the user feedbacks that the problem has not been solved, select the problem with engineer feedback, combine the fault path and the engineer feedback knowledge and add them to the feedback database; perform knowledge injection into the large model through the data in the feedback database, and update the knowledge graph.

[0023] As a further preference, the fault path includes: numerical control model, alarm number, alarm phenomenon, cause, and solution.

[0024] As a further preference, for the data in the feedback database, if there is no engineer feedback, divide the data into the enhanced training data set; if there is engineer feedback, divide the data into the optimized training data set;

[0025] Furthermore, perform knowledge injection into the large model through the data in the feedback database, including: for the data in the enhanced training data set, use the LoRA fine-tuning method to perform knowledge injection into the large model in the RAG framework; for the data in the optimized training data set, adopt the In-context learning method to combine the fault path and the engineer feedback knowledge to obtain an optimized fault path; based on the optimized fault path, use the LoRA fine-tuning method to perform knowledge injection into the large model and update the knowledge graph.

[0026] Generally speaking, compared with the prior art, the above technical solutions conceived by the present invention mainly have the following technical advantages:

[0027] 1. The present invention relies on the RAG framework and the knowledge graph-based knowledge base. During the fault diagnosis process, knowledge injection is performed on the large model in the RAG framework through the feedback from users and engineers, and the knowledge base is updated. Thus, a knowledge base deeply integrated with expert experience is obtained, and continuous learning and optimization of the large model and the knowledge base are realized, effectively improving the accuracy of fault diagnosis.

[0028] 2. For the data with problems solved and no engineer feedback, the present invention uses the LoRA-based multi-task fine-tuning method to inject knowledge into the large model, so as to enhance its understanding of specific path knowledge while maintaining the overall stability of the model. For the data with engineer feedback, first, the In-context learning method is used to process the fault path and engineer feedback knowledge, and the optimized fault path is output. Then, the LoRA-based multi-task fine-tuning method is used to inject knowledge into the large model, enabling the model to effectively absorb and understand the updated content and improving the diagnostic accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart of the fault diagnosis method for a numerical control system based on knowledge injection according to an embodiment of the present invention;

[0030] Figure 2 is a flowchart of knowledge injection for the feedback database according to an embodiment of the present invention;

[0031] Figure 3 is a flowchart of knowledge graph update according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0033] A fault diagnosis method for a numerical control system based on knowledge injection provided by an embodiment of the present invention, as Figure 1 shown, includes the following steps:

[0034] (1) Perform fault diagnosis based on the RAG framework, which includes a knowledge base and a large model. Retrieve in the knowledge base based on the user's question, and the large model outputs the fault diagnosis result according to the retrieved fault path and the user's question.

[0035] Furthermore, a knowledge graph for CNC system fault diagnosis is constructed as a knowledge base, which mainly stores historical fault cases. The knowledge graph contains entity elements such as "CNC model" - "alarm number" - "alarm phenomenon" - "cause" - "solution", forming a complete fault path. During each fault diagnosis process, the large model will combine the path extracted from the knowledge graph with the user's question to generate an answer.

[0036] Specifically, during the knowledge base retrieval process, first, the CNC model and alarm number are extracted from the user's question through named entity recognition. Then, the CNC model is precisely matched in the knowledge graph, and the alarm number is accurately located through the "trigger" relationship. Next, the "manifested as" relationship is used to perform fuzzy matching of the alarm phenomenon through vector calculation. Finally, the cause and solution are traversed through breadth-first search (BFS) or depth-first search (DFS) to generate a complete knowledge graph path. The entities and relationships in the knowledge graph are shown in Table 1.

[0037] Table 1 Knowledge Graph in the Field of CNC System Fault Diagnosis

[0038]

[0039]

[0040] (2) Construct a feedback database to store the new knowledge obtained during the fault diagnosis process, including the knowledge graph paths retrieved during each fault diagnosis process, the newly constructed knowledge graph paths, and their corresponding engineer feedback knowledge. The knowledge in the feedback database can be used to inject into the large model and enhance or update its knowledge base to improve the model's fault diagnosis ability.

[0041] Specifically, the structure of the feedback database is shown in Table 2.

[0042] Table 2 Example of CNC System Fault Diagnosis Feedback Database

[0043]

[0044] Furthermore, after one fault diagnosis, a question-and-answer pair consisting of the user's question and the large model's answer is obtained, as well as the knowledge graph path retrieved from the knowledge base during the diagnosis process. This path includes the CNC model, alarm number, alarm phenomenon, cause, and solution. The generation of the knowledge path stems from the user's question about the alarm number and alarm phenomenon of a certain CNC device, and is gradually retrieved in the knowledge graph through multiple rounds of question and answer.

[0045] After fault diagnosis through the RAG framework, the obtained question-and-answer pairs are classified according to whether the problem is solved. The judgment of whether the user's problem is solved can be fed back by the user after the question and answer.

[0046] For the Q&A pairs classified as solved problems, the corresponding knowledge graph paths are regarded as correct knowledge. Extract the corresponding knowledge graph fault paths (including CNC model, alarm number, alarm phenomenon, cause, and solution), add them to the feedback database, and at the same time mark them as "no feedback" in the engineer feedback information to indicate that this path is correct and does not need to be modified. These correct knowledge graph paths will be used for the reinforcement learning of the large model to further improve the diagnostic accuracy and reliability of the model.

[0047] For the Q&A pairs classified as unsolved problems, during the fault diagnosis process, engineers can call the engineer feedback interface to supplement or provide feedback on the Q&A pairs that have not been solved. The corresponding knowledge graph paths of these Q&A pairs are regarded as incorrect or incomplete knowledge; if there is no engineer feedback knowledge for this knowledge graph path, it will not be processed and will not be added to the feedback database; if there is feedback, combine the knowledge graph path and the engineer feedback knowledge and add them to the feedback database. Specifically, if the engineer provides feedback information, add the knowledge graph path and the corresponding engineer feedback knowledge to the feedback database together to indicate that this path needs to be modified or improved. Subsequently, adjust and update this path according to the engineer's feedback knowledge and inject the updated knowledge into the large model, thereby improving the model's fault diagnosis ability and enhancing its response level in similar problems.

[0048] In addition, for unsolved problems, users can subsequently solve this fault problem through expert remote consultation or engineer on-site operation and maintenance, and finally obtain a work order automatically. The content of the work order includes the CNC model, alarm number, alarm phenomenon, cause, and solution, which can be constructed into a fault path through the named entity recognition method. This path is processed according to the method of the above correct knowledge, that is, add the constructed fault path to the feedback database and mark it as "no feedback".

[0049] (3) Inject knowledge into the large model in the RAG framework through the data in the feedback database and update the knowledge base.

[0050] Furthermore, as Figure 2 shown, classify the data in the feedback database according to whether there is engineer feedback, that is, judge by the content of the engineer feedback field. Specifically, if the value of the engineer feedback field in this piece of data is "no feedback", then divide this piece of data into the enhanced training dataset for enhancing the training of the large model to enhance the large model's understanding of correct knowledge and improve its accuracy and response speed for similar problems. Otherwise, divide this piece of data into the optimization training dataset for optimizing the training of the large model.

[0051] Furthermore, for the data in the enhanced training dataset, a LoRA-based multi-task fine-tuning method is used to inject knowledge into the large model, thereby enhancing the model's understanding of these fault paths.

[0052] Specifically, for the data in the enhanced training dataset, it is necessary to enhance the model's understanding of the key logical relationships in the fault paths. First, it is necessary to understand each triple relationship in the knowledge graph path, and at the same time, it is necessary to understand the overall logic of the fault path. Therefore, a triple understanding task and a multi-hop entity association task are designed. During the task process, the LoRA fine-tuning method is used to freeze the parameters of most language models and introduce two low-rank matrices, and only the parameters of the low-rank matrices are fine-tuned. Thus, on the premise of maintaining the overall stability of the model, its understanding of specific path knowledge is enhanced. Next, the task design is introduced.

[0053] 1) Triple understanding task:

[0054] This task aims to help the large model understand the relationships expressed by triples in the form of a knowledge graph and in the fault paths. A fault path contains multiple triples. The form of a triple is expressed as:

[0055] t = [e 1 , r, e 2

[0056] where t represents the triple set, e 1 is the head entity, r is the relationship, and e 2 is the tail entity. For the triple understanding task, the constructed training set is:

[0057] z = [instruction, e 1 , r, e 2

[0058] where instruction is the instruction. The training task can be expressed by the formula:

[0059] e 2 ’ = M(z)

[0060] where e 2 ’ is the expected output of the model. Among them, the large model outputs e 1 , r according to the instruction instruction and e 2 , and optimizes the large model M by calculating the loss between e 2 ' and e 2 .

[0061] 2) Multi-hop entity association task:

[0062] ​​The task is to help the large model understand the logical relationships between fault cases. The model input is multiple entity sets, and the training objective of the model is to analyze all entities and generate a case text containing all entities and correctly describe the relationships between the entities. If the logical relationship is completely correct, it indicates that the model has mastered the associations between the entities. The task can be represented by the formula:

[0063] Text’=M(instruction,e 1 ,e 2 ,...,e n )

[0064] where e is the entity and Text’ is the text output by the model, which describes the relationships between all entities. The model M is optimized by calculating the loss between Text’ and Text.

[0065] Furthermore, for the data in the optimized training dataset, in-context learning and LoRA-based multi-task fine-tuning methods are adopted. The knowledge graph path and engineer feedback knowledge modification path are combined and injected into the large model.

[0066] Specifically: for the data in the optimized training dataset, first, an in-context learning context is constructed by combining the knowledge graph path to be modified with the engineer feedback knowledge. This in-context learning context consists of three parts: task description, engineer input, and task examples, clearly indicating that the model needs to understand the engineer feedback knowledge and adjust the knowledge graph path accordingly, so that the model can give accurate answers to the updated path knowledge in future diagnoses. The task examples, as the core part of the ICL context, are derived from the successful cases of engineer feedback tasks in daily work. The core function is to help the model understand the target task through an analogical reasoning mechanism. The in-context learning context constructed in this embodiment is shown in Table 3:

[0067] Table 3 In-context learning context

[0068]

[0069] For the optimized and updated path, knowledge injection is performed again based on the LoRA-based multi-task fine-tuning method to ensure that the large model can effectively absorb and understand the updated content. By designing triple understanding tasks and multi-hop entity association tasks for the updated path and using the LoRA fine-tuning method to keep the main parameters frozen and only fine-tune the introduced low-rank matrix parameters, the model can better understand the knowledge contained in the new path and improve the diagnostic accuracy of the model.

[0070] Furthermore, for the updated and newly added knowledge graph paths, they need to be updated into the knowledge base (knowledge graph). For example, Figure 3 as shown: First, accurately match the numerical control model in the knowledge graph as the starting node of the path. If the match fails, a new numerical control model node is added. Then, starting from the starting node, accurately match the alarm numbers in sequence; perform vectorized fuzzy matching on nodes such as alarm phenomena, causes, and solutions. Specifically, vectorize the text of the fault phenomenon, fault cause, and solution, and calculate the similarity with the corresponding nodes. If the similarity is greater than 0.75, it is considered a successful match. If there are multiple nodes with a similarity greater than 0.75, select the node with the highest similarity for matching. If the match fails, start from the previous successfully matched node and construct a new path to complete the update.

[0071] Furthermore, set double thresholds of quantity and time for the feedback database. When the data volume in the feedback database reaches the preset quantity threshold, or although the data volume has not reached it, but the time since the last data processing has exceeded the preset time threshold, extract the data in the feedback database for processing operations.

[0072] A numerical control system fault diagnosis system based on knowledge injection provided by an embodiment of the present invention relies on the RAG framework and includes a knowledge base, a large model, and a learning and updating module, where:

[0073] The knowledge base is a knowledge graph constructed based on the historical data of numerical control system fault diagnosis. The knowledge base is used to output fault paths according to user questions; the large model is used to output fault diagnosis results according to the retrieved fault paths and user questions; the learning and updating module is used to update the large model and the knowledge base according to user feedback, including: if the user feedbacks that the problem has been solved, directly add the fault path to the feedback database; if the user feedbacks that the problem has not been solved, add the fault path combined with the engineer's feedback knowledge to the feedback database; perform knowledge injection on the large model through the data in the feedback database, and update the knowledge base. For the specific method, refer to the aforementioned numerical control system fault diagnosis method based on knowledge injection, which will not be elaborated here.

[0074] Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for fault diagnosis of numerical control system based on knowledge injection, characterized in that: The steps include: Fault diagnosis is performed based on the RAG framework, which includes a knowledge base and a large model. The knowledge base is a knowledge graph constructed based on the historical data of fault diagnosis of the numerical control system. Fault paths related to user problems are retrieved from the knowledge base, and then the large model integrates the retrieved fault paths and user problems to generate fault diagnosis results. After each fault diagnosis, determine whether the user's problem is solved. If it is solved, the fault path is directly added to the feedback database. If it is not solved, the problem reported by the engineer will be added to the feedback database in combination with the fault path and engineer feedback knowledge. Knowledge is injected into the large model through data from the feedback database, and the knowledge graph is updated.

2. The method for fault diagnosis of numerical control system based on knowledge injection as claimed in claim 1, characterized in that: The fault path includes: CNC model, alarm number, alarm phenomenon, cause and solution.

3. The method for fault diagnosis of numerical control system based on knowledge injection as claimed in claim 2, characterized in that: For the data in the feedback database, if there is no feedback from engineers, the data will be divided into the enhanced training data set; if there is feedback from engineers, the data will be divided into the optimized training data set.

4. The method for fault diagnosis of numerical control system based on knowledge injection as claimed in claim 3, characterized in that: Knowledge is injected into the large model through the data in the feedback database, including: For the data in the enhanced training dataset, the LoRA fine-tuning method is used to inject knowledge into the large model in the RAG framework; For the data in the optimized training data set, the In-context learning method is used to combine the fault path and engineer feedback knowledge to obtain the optimized fault path; based on the optimized fault path, the LoRA fine-tuning method is used to inject knowledge into the large model and update the knowledge graph.

5. The method for NC system fault diagnosis based on knowledge injection as claimed in claim 4, characterized in that: For unresolved user issues, users can eventually solve the problems through remote consultation with experts or on-site maintenance by engineers, and automatically receive a work order. Based on the work order, a new fault path is constructed using the named entity recognition method, and the newly added fault path is added to the feedback database and divided into the enhanced training data set.

6. The method for fault diagnosis of numerical control system based on knowledge injection as claimed in claim 5, characterized in that: Update the knowledge graph through the data in the feedback database, including: For each newly added and optimized fault path, first accurately match the CNC model in the knowledge graph according to the fault path. If the match is not successful, add a new CNC model and use the obtained CNC model as the starting node of the path; Then, starting from the starting node, the alarm number, alarm phenomenon, cause, and solution nodes are matched in the knowledge graph in turn; among them, exact matching is used for the alarm number, and vectorized fuzzy matching is used for the alarm phenomenon, cause, and solution; in this process, if a node is not matched successfully, starting from the previous node, a new path is constructed to join the knowledge graph to achieve update.

7. The method for fault diagnosis of a numerical control system based on knowledge injection according to any one of claims 1 to 6, characterized in that: A dual threshold of quantity and time is set for the feedback database. When the amount of data in the feedback database reaches the preset threshold, or the time since the last data processing exceeds the preset time threshold, the data in the feedback database is extracted to inject knowledge into the large model.

8. A CNC system fault diagnosis system based on knowledge injection, characterized in that: Includes RAG framework and learning update modules, including: The RAG framework includes a knowledge base and a large model, wherein the knowledge base is a knowledge graph constructed based on the historical data of fault diagnosis of the numerical control system, and the knowledge base is used to retrieve the fault path related to the user problem; the large model is used to comprehensively retrieve the fault path and user problem to generate the fault diagnosis result; The learning update module is used to update the large model and knowledge graph according to user feedback, including: if the user feedback problem has been solved, the fault path is directly added to the feedback database; if the user feedback problem is not solved, the problem with engineer feedback is selected, and the fault path and engineer feedback knowledge are combined to add to the feedback database; knowledge is injected into the large model through the data in the feedback database, and the knowledge graph is updated.

9. The numerical control system fault diagnosis system based on knowledge injection as claimed in claim 8, characterized in that: The fault path includes: CNC model, alarm number, alarm phenomenon, cause and solution.

10. The numerical control system fault diagnosis system based on knowledge injection as claimed in claim 9, characterized in that: For the data in the feedback database, if there is no feedback from engineers, the data will be divided into the enhanced training data set; if there is feedback from engineers, the data will be divided into the optimized training data set; Then, knowledge is injected into the big model through the data in the feedback database, including: for the data in the enhanced training data set, the LoRA fine-tuning method is used to inject knowledge into the big model in the RAG framework; for the data in the optimized training data set, the In-context learning method is used to combine the fault path and engineer feedback knowledge to obtain the optimized fault path; based on the optimized fault path, the LoRA fine-tuning method is used to inject knowledge into the big model and update the knowledge graph.

Images