Numerical control machine tool ai-fmeca method based on large language model
By integrating large language models, knowledge graphs, and graph neural networks, a fault knowledge graph is constructed and risk propagation is quantified. This solves the problems of low efficiency and inaccurate results of traditional FMECA methods in CNC machine tools, and realizes efficient and traceable intelligent analysis of faults in complex systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF TECH
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-12
AI Technical Summary
Existing FMECA methods in CNC machine tools suffer from problems such as reliance on human experience, low efficiency, insufficient data integration, and lack of factual basis for analysis results, making it difficult to effectively quantify complex fault mechanisms and cross-subsystem coupling effects.
A fault knowledge graph is constructed by integrating Large Language Model (LLM), Knowledge Graph (KG), and Graph Neural Network (GNN). Risk propagation is calibrated and quantified through GNN, and controlled reasoning and parameter scoring of LLM are combined to form a complete analysis process.
It enables comprehensive and accurate analysis of failure risks in complex systems, improves analysis efficiency and decision-making quality, overcomes the limitations of traditional methods, and provides traceable and highly reliable results.
Smart Images
Figure CN122196427A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of reliability assessment and fault diagnosis technology for complex equipment. Specifically, it relates to a Failure Mode, Effects and Criticality Analysis (FMECA) method that integrates Large Language Model (LLM), Knowledge Graph (KG), and Graph Neural Network (GNN). It is particularly suitable for intelligent reliability analysis of complex systems such as CNC machine tools. Background Technology
[0002] As the core of high-end manufacturing equipment, CNC machine tools are characterized by system complexity and tight coupling between subsystems, resulting in numerous failure modes and intertwined causal links. This makes FMECA (Fault-Tolerant Analysis) crucial for ensuring their reliable operation. However, existing methods face multiple challenges in practical applications: Traditional FMECA relies heavily on human experience, leading to strong subjectivity and low efficiency, making it difficult to fully quantify complex failure mechanisms and cross-subsystem coupling effects; although artificial intelligence, such as large language models, offers the possibility of automated analysis, its direct application in engineering diagnosis can easily produce "illusions," resulting in a lack of factual basis and structured constraints in failure reasoning and risk assessment results; at the same time, existing technologies have failed to effectively integrate unstructured texts such as operation and maintenance logs and technical manuals with multi-source information such as system topology and sensor data, resulting in the failure to fully explore the failure mechanisms and related knowledge hidden in the data, limiting the depth and breadth of analytical capabilities. Summary of the Invention
[0003] The purpose of this invention is to provide an FMECA analysis method that integrates LLM, KG, and GNN.
[0004] The technical solution adopted in this invention is as follows: This method constructs a fault knowledge graph based on LLM, and uses GNN to perform reasoning and risk propagation quantification on the graph. Finally, under the constraint of KG, it drives LLM to achieve FMECA intelligent analysis, thus forming a complete analysis process from knowledge construction and risk quantification to intelligent decision-making. The specific steps are as follows:
[0005] Step 1: Construction of a Fault Knowledge Graph Based on LLM. This step aims to construct a unified feature space and fault knowledge graph by processing multi-source heterogeneous reliability data. Specifically, it includes the following sub-steps:
[0006] Step 1.1 Data Preprocessing and Architecture Construction. Collect equipment design specifications, operation and maintenance logs, etc., and standardize the corresponding data, including unified coding and correction of redundant terminology. Utilize LLM to extract system, subsystem, and component information from unstructured text based on prompts for different types of data processing, and establish a layered system architecture skeleton.
[0007] Step 1.2 Two-stage knowledge extraction and AI review. A hybrid extraction strategy combining rules and LLM is used to extract triple information such as fault modes, causes, and effects. An initial knowledge graph is formed, and then an "AI expert review and synchronous correction" mechanism is introduced, using LLM to logically verify the generated knowledge graph entries.
[0008] Step 1.3 Entity Alignment and Graph Fusion. Fault knowledge, reviewed by AI, is injected into the graph. Topological rigor is ensured through structural normalization, and entity alignment and conflict resolution are achieved based on comprehensive similarity calculation. Conflict adjudication follows the principles of "prioritizing source credibility and timeliness." Alignment is performed by calculating the comprehensive similarity between entities. The formula is: In the formula, The weighting coefficients are determined through cross-validation. When they are identified as the same entity, a fusion operation is triggered.
[0009] Step Two: Graph Calibration and Risk Propagation Quantification Based on GNN. This step utilizes a Graph Neural Network (GNN) to perform secondary calibration of the graph quality and quantify the risk of fault propagation within the system. Specifically, it includes the following sub-steps:
[0010] Step 2.1 Node Feature Construction and Knowledge Graph Structure Calibration. An initial feature vector is constructed for each node in the knowledge graph, including TF-IDF text features, structural features such as degree and clustering coefficients, and data source confidence features. A graph convolutional network (LMN) is used for node classification, and the logical consistency of the knowledge graph structure constructed by the LM is checked and calibrated twice to ensure the rigor of the knowledge graph's semantics. To achieve this task, we first construct an information-rich initial feature representation for each node in the knowledge graph. The knowledge graph consists of entities such as systems, subsystems, components, and failure modes, and the relation types are derived from... The image is recorded as follows .remember .node The initial representation is obtained by concatenating three types of features: text, structure, and confidence. Text features Obtained from component / failure mode descriptions using TF-IDF or sentence vector models. Structural features. Includes: node in-degree / out-degree Hierarchical Depth root node arrive Shortest path length; Neighbor clustering coefficient An approximation using an undirected skeleton: In the formula, for The number of triangles among neighbors. Finally, the node features are represented as: In the formula, Indicates standardization by dimension. Confidence feature. It originates from the knowledge extraction and graph processing stages, including the confidence scores of nodes. Data source reliability These multi-source features are input into a two-layer graph convolutional network for representation learning. Its propagation formula is: In the formula, To add a self-loop adjacency matrix, For degree matrix, For the first Layer weights, This is a non-linear activation function. After multiple rounds of information propagation, the final embedding vector of a node not only contains its own features but also incorporates contextual information from its multi-step reachable neighbors in the network, thus effectively capturing cross-subsystem dependencies and potential fault propagation paths. This classification task uses a weighted cross-entropy loss function for end-to-end training to handle potential class imbalance problems. In the formula, Embedded for nodes, For category weights.
[0011] Step 2.2 Calculate the transmission risk factor ( Based on the calibrated graph topology, the propagation risk factor for each fault node is calculated iteratively using a GNN. This is used to quantify the coupling propagation effect of faults across levels and subsystems. The calculation formula is: In the formula, This is the attenuation coefficient, used to ensure the gradual attenuation of risk propagation. This represents the maximum number of propagation steps. This factor reflects the cumulative and amplified effect of risk over multiple steps, identifying critical failure risk nodes in the system.
[0012] Step 3: FMECA driven by a large language model under knowledge graph constraints.
[0013] This step aims to combine propagation risk factors under the semantic constraints of KG. To achieve accurate quantification and risk ranking of FMECA parameters, the following sub-steps are included:
[0014] Step 3.1 Controlled Knowledge Reasoning and Parameter Quantification. Using cueing engineering, relevant local subgraphs, fault causal chains, and preset reliability scoring criteria from the knowledge graph constructed in Step 1 are injected into the LLM. By strictly limiting the LLM's reasoning space to the entities and logic defined in the graph, the LLM is guided to quantify the severity, occurrence, and detectability of each fault mode, ensuring that the scoring results are based on factual evidence and avoiding the illusion problem.
[0015] Step 3.2 Calculate the revised Risk Priority Number (RPN). Introduce the propagation risk factor calculated in Step 2. The traditional risk priority number calculation formula is modified to obtain the modified risk priority number. The calculation formula is: ,in Indicates seriousness. Indicates the probability of occurrence. This indicates the difficulty of detection. This formula can identify high-risk fault modes that, while having a low probability of occurrence, are located on the core path of fault propagation and can cause a chain reaction in the system.
[0016] Step 3.3 Evidence Tracing and Logic Auditing. The LLM (Local Level Management) system is required to retrieve and return textual evidence or logical paths supporting the score from the knowledge graph while outputting the scoring result. The system automatically compares the content generated by the LLM with the original records in the knowledge graph, verifies logical consistency, and manually intervenes and marks results with low confidence or logical contradictions, thus ensuring the auditability of the analysis process. Based on this audited risk assessment result with a complete chain of evidence, the system can not only support the quantitative analysis and comparison of the overall risk level of subsystems and accurately locate key failure modes, but also provide a clear basis for the formulation of preventive maintenance strategies. This significantly improves the quality of analysts' judgments in fault risk assessment, enabling them to focus on high-value decision-making processes and ultimately achieve more comprehensive, efficient, and traceable management and response to fault risks in complex systems.
[0017] The advantages and positive effects of this invention are as follows: By deeply fusing multi-source heterogeneous data and extracting knowledge in a two-stage manner, combined with AI review and entity alignment mechanisms, a high-quality, structurally rigorous fault knowledge graph is constructed, achieving a more comprehensive and accurate semantic representation of the failure mechanisms of complex systems. The introduction of graph neural networks for graph calibration and propagation risk factor calculation effectively quantifies the coupling and amplification effects of faults across subsystems and levels, overcoming the inherent limitations of traditional FMECA methods in assessing propagation path risks. Under the strong semantic constraints of the knowledge graph, a large language model is driven to perform controlled reasoning and parameter scoring, significantly suppressing the "illusion" problem and ensuring the logical consistency and traceability of the analysis process. Its highly reliable output directly supports subsystem risk comparison, key failure mode localization, and preventative maintenance strategy formulation. The entire method forms a closed-loop framework of "construction-calibration-analysis-audit," which not only improves the accuracy of analysis and the scientific nature of decision-making under small sample conditions but also achieves intelligent, auditable analysis of the entire process of complex system fault risk from perception and quantification to control, significantly improving the analysis efficiency and decision-making quality of reliability engineering. Attached Figure Description
[0018] Figure 1 This is the AI-FMECA method flowchart.
[0019] Figure 2 This is a flowchart of the construction process for a fault knowledge graph.
[0020] Figure 3 This is a flowchart of graph calibration and risk propagation quantification based on GNN.
[0021] Figure 4 It is a flowchart of FMECA driven by a large language model under the constraints of a knowledge graph. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0023] This embodiment takes a certain type of vertical milling and turning five-axis composite machining center as the analysis object. This equipment includes multiple complex subsystems such as the tool spindle system, feed system, hydraulic system, and electrical control system. Figure 1The figure illustrates the overall process of the "AI-FMECA method for CNC machine tools based on a large language model" proposed in this invention. The implementation of this method begins with collecting fault handling data from similar models of a vertical milling and turning five-axis machining center. This data typically includes the specific location of the machine tool fault, a description of the cause, and records of corresponding handling measures.
[0024] The specific implementation steps of this invention are as follows: Step 1: Construction of a fault knowledge graph based on LLM. This step aims to construct a unified feature space and fault knowledge graph by processing multi-source heterogeneous reliability data. The detailed process of this step is as follows: Figure 2 As shown, it specifically includes the following sub-steps:
[0025] Step 1.1 Data Preprocessing and Architecture Construction. Collect equipment design specifications, operation and maintenance logs, etc., and standardize the corresponding data, including unified coding and correction of redundant terminology. Utilize LLM to extract system, subsystem, and component information from unstructured text based on prompts for different types of data processing, and establish a layered system architecture skeleton.
[0026] Step 1.2 Two-stage knowledge extraction and AI review. A hybrid extraction strategy combining rules and LLM is used to extract triple information such as fault modes, causes, and effects. An initial knowledge graph is formed, and then an "AI expert review and synchronous correction" mechanism is introduced, using LLM to logically verify the generated knowledge graph entries.
[0027] Step 1.3 Entity Alignment and Graph Fusion. Fault knowledge, reviewed by AI, is injected into the graph. Topological rigor is ensured through structural normalization, and entity alignment and conflict resolution are achieved based on comprehensive similarity calculation. Conflict adjudication follows the principles of "prioritizing source credibility and timeliness." Alignment is performed by calculating the comprehensive similarity between entities. The formula is: In the formula, The weighting coefficients are determined through cross-validation. When they are identified as the same entity, a fusion operation is triggered.
[0028] Step Two: Graph Calibration and Risk Propagation Quantification Based on GNN. This step utilizes a graph neural network to perform secondary calibration of the graph quality and quantify the risk of fault propagation within the system. The process for this step is as follows: Figure 3 As shown. Specifically, it includes the following sub-steps:
[0029] Step 2.1 Node Feature Construction and Knowledge Graph Structure Calibration. An initial feature vector is constructed for each node in the knowledge graph, including TF-IDF text features, structural features such as degree and clustering coefficients, and data source confidence features. A node classification task is performed using GCN, and the logical consistency of the knowledge graph structure constructed by LLM is checked and calibrated twice to ensure the rigor of the knowledge graph's semantics. To achieve this task, we first construct an information-rich initial feature representation for each node in the knowledge graph. The knowledge graph consists of entities such as systems, subsystems, components, and failure modes, and the relation types are derived from... The image is recorded as follows .remember .node The initial representation is obtained by concatenating three types of features: text, structure, and confidence. Text features Obtained from component / failure mode descriptions using TF-IDF or sentence vector models. Structural features. Includes: node in-degree / out-degree Hierarchical Depth root node arrive Shortest path length; Neighbor clustering coefficient An approximation using an undirected skeleton: In the formula, for The number of triangles among neighbors. Finally, the node features are represented as: In the formula, Indicates standardization by dimension. Confidence feature. It originates from the knowledge extraction and graph processing stages, including the confidence scores of nodes. Data source reliability These multi-source features are input into a two-layer graph convolutional network for representation learning. Its propagation formula is: In the formula, To add a self-loop adjacency matrix, For degree matrix, For the first Layer weights, This is a non-linear activation function. After multiple rounds of information propagation, the final embedding vector of a node not only contains its own features but also incorporates contextual information from its multi-step reachable neighbors in the network, thus effectively capturing cross-subsystem dependencies and potential fault propagation paths. This classification task uses a weighted cross-entropy loss function for end-to-end training to handle potential class imbalance problems. In the formula, Embedded for nodes, This represents the category weight.
[0030] Step 2.2 Calculate the transmission risk factor ( Based on the calibrated graph topology, the propagation risk factor for each fault node is calculated iteratively using a GNN. This is used to quantify the coupling propagation effect of faults across levels and subsystems. The calculation formula is: In the formula, This is the attenuation coefficient, used to ensure the gradual attenuation of risk propagation. This represents the maximum number of propagation steps. This factor reflects the cumulative and amplified effect of risk over multiple steps, identifying critical failure risk nodes in the system.
[0031] Step 3: FMECA driven by a large language model under knowledge graph constraints.
[0032] This step aims to combine propagation risk factors under the semantic constraints of KG. This enables accurate quantification and risk ranking of FMECA parameters. The process for this step is as follows: Figure 4 As shown.
[0033] Step 3.1 Controlled Knowledge Reasoning and Parameter Quantification. Using cueing engineering, relevant local subgraphs, fault causal chains, and preset reliability scoring criteria from the knowledge graph constructed in Step 1 are injected into the LLM. By strictly limiting the LLM's reasoning space to the entities and logic defined in the graph, the LLM is guided to quantify the severity, occurrence, and detectability of each fault mode, ensuring that the scoring results are based on factual evidence and avoiding the illusion problem.
[0034] Step 3.2 Calculate the corrected RPN. Introduce the propagation risk factor calculated in Step 2. The traditional risk priority number calculation formula is modified to obtain the modified risk priority number. The calculation formula is: ,in Indicates seriousness. Indicates the probability of occurrence. This indicates the difficulty of detection. This formula can identify high-risk fault modes that, while having a low probability of occurrence, are located on the core path of fault propagation and can cause a chain reaction in the system.
[0035] Step 3.3 Evidence Tracing and Logic Auditing. The LLM (Local Level Management) system is required to retrieve and return textual evidence or logical paths supporting the score from the knowledge graph while outputting the scoring result. The system automatically compares the content generated by the LLM with the original records in the knowledge graph, verifies logical consistency, and manually intervenes and marks results with low confidence or logical contradictions, thus ensuring the auditability of the analysis process. Based on this audited risk assessment result with a complete chain of evidence, the system can not only support the quantitative analysis and comparison of the overall risk level of subsystems and accurately locate key failure modes, but also provide a clear basis for the formulation of preventive maintenance strategies. This significantly improves the quality of analysts' judgments in fault risk assessment, enabling them to focus on high-value decision-making processes and ultimately achieve more comprehensive, efficient, and traceable management and response to fault risks in complex systems.
[0036] This invention constructs a high-quality, structurally rigorous fault knowledge graph through deep fusion of multi-source heterogeneous data and two-stage knowledge extraction, combined with AI review and entity alignment mechanisms. This enables a more comprehensive and accurate semantic representation of the failure mechanisms of complex systems. By introducing graph neural networks for graph calibration and propagation risk factor calculation, it effectively quantifies the coupling and amplification effects of faults across subsystems and levels, overcoming the inherent limitations of traditional FMECA methods in assessing propagation path risks. By driving a large language model for controlled reasoning and parameter scoring under the strong semantic constraints of the knowledge graph, it significantly suppresses the "illusion" problem, ensuring the logical consistency of the analysis process and the traceability of the results. Its highly reliable output directly supports subsystem risk comparison, key failure mode localization, and preventive maintenance strategy formulation, thus forming a closed-loop framework of "construction-calibration-analysis-audit." This enables intelligent, auditable analysis of complex system fault risks from perception and quantification to control, significantly improving the analysis efficiency and decision-making quality of reliability engineering.
[0037] This specification and accompanying drawings are merely illustrative examples of the present invention and are to be considered as covering any and all modifications, variations, combinations or equivalents within the scope of the present invention.
Claims
1. A CNC machine tool AI-FMECA method based on a large language model, characterized in that, Includes the following steps: Step 1: Constructing a fault knowledge graph based on LLM. By processing multi-source heterogeneous reliability data, a unified fault knowledge graph is constructed. Multi-source heterogeneous CNC machine tool reliability data is collected, and information is extracted and fused using the Large Language Model (LLM) to construct a structured fault knowledge graph. Step 2: Graph calibration and risk propagation quantification based on GNN. The fault knowledge graph is calibrated a second time using a graph neural network, and the propagation risk of the fault within the system is quantified to obtain the propagation risk factor. The fault knowledge graph is input into a graph neural network (GNN), and the graph structure is calibrated for consistency through a graph node classification task. Based on the calibrated topology, the propagation risk factor of each fault node is iteratively calculated to quantify the coupling propagation effect of the fault in the system. Step 3: FMECA driven by a large language model under the constraints of the knowledge graph. Under the semantic constraints of the fault knowledge graph, combined with the propagation risk factors, the FMECA parameters are quantified and the risk ranking is achieved. The fault knowledge graph is used as a constraint to input into the large language model LLM, which guides it to perform FMECA parameter quantification scoring on fault modes; and the traditional risk priority number is modified using the propagation risk factor to generate a modified risk assessment result.
2. The AI-FMECA method for CNC machine tools based on a large language model according to claim 1, characterized in that, Step one includes: Step 1.1 Data Preprocessing and Architecture Construction: Collect equipment design specifications and operation and maintenance logs, and standardize the corresponding data, including unified coding and correction of redundant terms; use LLM to extract system, subsystem and component information from unstructured text based on prompts for different types of data processing, and establish a layered system architecture skeleton. Step 1.2 Two-stage knowledge extraction and AI review: A hybrid extraction strategy combining rules and LLM is adopted to extract triple information such as fault mode, cause and effect; an initial knowledge graph is formed, and then an AI expert review and synchronous correction mechanism is introduced, and LLM is used to perform logical verification on the generated knowledge graph entries; Step 1.3 Entity Alignment and Graph Fusion: Fault knowledge reviewed by AI is injected into the graph. Topological rigor is ensured through structural normalization, and entity alignment and conflict resolution are achieved based on comprehensive similarity calculation. The comprehensive similarity between entities is calculated for alignment; the formula is: In the formula, The weighting coefficients are determined through cross-validation; when When they are identified as the same entity, a fusion operation is triggered.
3. The AI-FMECA method for CNC machine tools based on a large language model according to claim 1, characterized in that, Step two includes: Step 2.1 Node Feature Construction and Knowledge Graph Structure Calibration: An initial feature vector is constructed for each node in the knowledge graph, including TF-IDF text features, structural features such as degree and clustering coefficients, and data source confidence features. A node classification task is performed using a graph convolutional network (PWM). Logical consistency verification and secondary calibration are conducted on the graph structure constructed by the LLM to ensure the rigor of the knowledge graph's semantics. First, an information-rich initial feature representation is constructed for each node in the knowledge graph. The knowledge graph consists of entities such as systems, subsystems, components, and failure modes; the relation types are derived from... The image is recorded as follows ;remember ;node The initial representation is obtained by concatenating three types of features: text, structure, and confidence; text features Structural features are obtained from the component / failure mode description statements via TF-IDF or sentence vector modeling. Includes: node in-degree / out-degree Hierarchical Depth root node arrive Shortest path length; Neighbor clustering coefficient An approximation using an undirected skeleton: In the formula, for The number of triangles among neighbors; finally, the node features are represented as: In the formula, Indicates dimensional standardization; confidence feature It originates from the knowledge extraction and graph processing stages, including the confidence scores of nodes. Data source reliability These multi-source features are input into a two-layer graph convolutional network for representation learning; its propagation formula is: In the formula, To add a self-loop adjacency matrix, For degree matrix, For the first Layer weights, The function is a non-linear activation function. After multiple rounds of information propagation, the final embedding vector of a node not only contains its own features but also incorporates the contextual information of its multi-step reachable neighbors in the network, thus effectively capturing cross-subsystem dependencies and potential fault propagation paths. This classification task uses a weighted cross-entropy loss function for end-to-end training to handle potential class imbalance problems. In the formula, Embedded for nodes, Category weights; Step 2.2 Calculate the transmission risk factor Based on the calibrated graph topology, the propagation risk factor of each fault node is calculated iteratively using a GNN. To quantify the coupling propagation effect of faults across levels and subsystems; the calculation formula is: In the formula, This is the attenuation coefficient, used to ensure the gradual attenuation of risk propagation. The maximum propagation steps; this factor can reflect the cumulative and amplified effect of risk in a multi-step neighborhood, and identify key failure risk nodes in the system.
4. The AI-FMECA method for CNC machine tools based on a large language model according to claim 1, characterized in that, Step three includes: Step 3.1 Controlled knowledge reasoning and parameter quantification: Using prompting engineering, relevant local subgraphs, fault causal chains, and preset reliability scoring criteria from the knowledge graph constructed in Step 1 are injected into the LLM; Step 3.2 Calculate the revised risk priority number RPN; introduce the propagation risk factor obtained in Step 2. The traditional risk priority number calculation formula is modified to obtain the modified risk priority number. The calculation formula is: ,in Indicates seriousness. Indicates the probability of occurrence. Indicates the difficulty of the test; Step 3.3 Evidence tracing and logical auditing: The LLM is required to retrieve and return textual evidence or logical paths supporting the score from the knowledge graph while outputting the scoring results; the system automatically compares the content generated by the LLM with the original records in the graph to verify logical consistency, and manually intervenes and marks results with low confidence or logical contradictions to achieve auditability of the analysis process.