Design failure mode and effect analysis method and system based on knowledge graph and large model retrieval enhancement

Abstract

The present invention discloses a design failure mode and effect analysis method based on knowledge graph and large model retrieval enhancement, comprising the following steps: Step 1: Planning and preparation stage; Step 2: Structural analysis and functional analysis stage; Step 3: Failure analysis stage: Using graph knowledge reuse components to process input data and retrieve relevant information from the full-cycle knowledge base, the data is then processed and constructed by the industrial knowledge software system to obtain a knowledge graph; Based on the knowledge graph, the large model retrieval enhancement method deeply analyzes the input information, extracts relevant information from the knowledge graph in combination with the large model prompt learning method, and generates failure analysis results; Step 4: Risk analysis and optimization stage: Determine the issues that need to be addressed first, propose and implement improvement measures; Step 5: Result documentation and system integration stage. The present invention also discloses a design failure mode and effect analysis system based on knowledge graph and large model retrieval enhancement.

Description

Technical Field

[0001] The present invention belongs to the field of failure mode and effect analysis technology, and specifically provides a design failure mode and effect analysis method and system based on knowledge graph and large model retrieval enhancement. Background Art

[0002] DFMEA (Design Failure Mode and Effects Analysis) is a type of FMEA that focuses on failure prevention and risk management during the product design phase. It systematically identifies, analyzes, and evaluates potential failure modes and their impacts in a design, aiming to uncover and eliminate design weaknesses and ensure the reliability and safety of the product in actual use. By identifying and resolving design issues early, DFMEA provides a solid foundation for the successful development and application of products. In the design and manufacturing of new energy vehicles, DFMEA is a crucial tool for ensuring product reliability and safety. However, current DFMEA practices are plagued by several common issues.

[0003] (1) The DFMEA process is highly dependent on personal experience. Traditional analysis relies primarily on the knowledge and experience of a team of experts. This dependence results in analysis results being greatly affected by the professionalism and experience of the participants, making it prone to subjective bias and potentially missing critical failure modes, thus affecting product reliability and safety.

[0004] (2) The problem of information silos is serious. In the traditional DFMEA process, information sharing between different departments and teams is insufficient. Various design data and failure cases are scattered in their own systems, making it difficult to form a unified knowledge base. This information silo phenomenon makes it difficult to fully utilize existing experience and data during the analysis process, increasing duplication of work and waste of resources, and reducing the efficiency and accuracy of failure mode identification.

[0005] (3) Insufficient dynamic updating and continuous optimization of DFMEA information. Traditional DFMEA analysis is often fixed after the design phase is completed, lacking dynamic connection with feedback and improvement measures during actual use. As new energy vehicles continue to expose new problems and failure modes in actual use, the original DFMEA analysis results may be outdated and unable to reflect the current product status. This static analysis model limits the effectiveness of DFMEA throughout the product life cycle and makes it difficult to adapt to changing product requirements and technological advances.

[0006] (4) The complexity of new energy vehicles increases the difficulty of DFMEA. New energy vehicles involve a variety of new technologies and materials, such as batteries, motors, and electronic control systems. Their design and manufacturing processes are more complex than those of traditional fuel vehicles. The interactions between each system and component are more intimate, making the analysis and prevention of failure modes more challenging. Traditional DFMEA methods are unable to fully cover these complex interactions, resulting in potential failure modes being overlooked.

[0007] (5) The intelligence and automation level of DFMEA tools and methods is low. Most current DFMEA tools are still at the manual recording and analysis stage, lacking intelligent data processing and analysis methods. This not only increases the workload but also easily leads to human errors and omissions. With the development of big data and artificial intelligence technology, DFMEA tools urgently need to be upgraded to improve the automation and intelligence level of analysis, thereby improving the efficiency and accuracy of analysis. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a design failure mode and impact analysis method and system based on knowledge graph and large model retrieval enhancement, which can accurately identify and predict possible failure modes, improve the speed and accuracy of analysis, reduce human errors and omissions, and continuously improve the reliability and safety of the design throughout the product life cycle.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] The present invention first proposes a design failure mode and effect analysis method based on knowledge graph and large model retrieval enhancement, which includes the following steps:

[0011] Step 1: Planning and preparation phase

[0012] Conduct planning and preparation, clarify analysis objectives, and form a multidisciplinary team;

[0013] Step 2: Structural analysis and functional analysis stage

[0014] Conduct structural analysis of the product, identify the functions of each subsystem, assembly and component and their interrelationships, and fully understand the design intent of the product and the interaction between its parts;

[0015] Step 3: Failure Analysis Phase

[0016] Input design plans, design standards, two-dimensional models, and existing failure case information. Graph knowledge reuse components are used to process the input data and retrieve relevant information from the full-cycle knowledge base. The data is then processed and constructed using the industrial knowledge software system to generate a knowledge graph. A large-model retrieval enhancement method based on the knowledge graph deeply analyzes the input information. Combined with a large-model prompt learning method, relevant information is extracted from the knowledge graph to generate failure analysis results including failure impacts, failure modes, and failure causes.

[0017] Step 4: Risk Analysis and Optimization Phase

[0018] Evaluate the severity, frequency, and detection difficulty of each failure mode, calculate the risk priority number, and determine the issues that need to be addressed first; propose and implement improvement measures including design improvements, process adjustments, and increased detection methods to reduce or eliminate high-risk failure modes;

[0019] Step 5: Result documentation and system integration stage

[0020] The analysis results including failure modes, failure effects, failure causes, risk assessment and optimization measures are documented and stored in the FMEA knowledge base. By integrating with the data center, bill of materials (BOM) system and product lifecycle management (PLM) system, a full-cycle knowledge base is formed, and the full-cycle knowledge base is continuously updated and optimized throughout the product lifecycle.

[0021] Furthermore, in step 1, the planning and preparation steps are as follows:

[0022] 11) Clarify the objectives and scope of FMEA analysis and ensure that the basic information of the project is complete;

[0023] 12) Establish a multidisciplinary team including design, manufacturing, quality and service to ensure the comprehensiveness and professionalism of FMEA analysis;

[0024] 13) Develop a detailed project plan, clarify the time schedule, resource allocation and responsible persons for each task, and ensure that the FMEA analysis is carried out as planned.

[0025] Furthermore, in step 2, the structural analysis method is:

[0026] In the design failure mode and effect analysis, identify and define the product structure and clarify the interface design of each subsystem, component and part;

[0027] In process failure mode and effects analysis, critical steps in the manufacturing and assembly process are identified to ensure the stability of process inputs and outputs.

[0028] Furthermore, in step 2, the method of functional analysis is:

[0029] In the design failure mode and effect analysis, define the functions and characteristics of the product to ensure that the design can meet the expected performance and reliability requirements;

[0030] In process failure mode and effects analysis, key functions in the process are identified to ensure that each step can achieve its intended function and avoid failures caused by process variation.

[0031] Furthermore, in step three, the failure analysis method steps are:

[0032] 31) Identify possible failure modes of products or processes;

[0033] 32) Evaluate the potential impact of each failure mode on the system, subsystem, component, and user;

[0034] 33) Analyze potential causes of failure modes.

[0035] Furthermore, in step 4, the risk analysis method steps are:

[0036] 41) Identify and develop preventive measures to reduce the probability of failure;

[0037] 42) Identify and implement detection measures to improve the ability to detect failures after they occur;

[0038] 43) Score the severity S, frequency O and detection difficulty D of each failure mode, and re-evaluate the risk priority number based on the implemented prevention and detection measures to ensure that the risks are effectively controlled.

[0039] Furthermore, in step 4, the risk optimization method steps are:

[0040] 44) Identify and record high-risk failure modes that require improvement;

[0041] 45) Propose specific optimization measures including design improvement, material replacement and process adjustment;

[0042] 46) Implement optimization measures and verify their effectiveness;

[0043] 47) Score and evaluate optimization measures to ensure they significantly reduce risk priority numbers.

[0044] Furthermore, in step three, the large model retrieval enhancement method based on the knowledge graph is: starting from the design plan and standards, a problem combination representation is generated, and passed to the large model through Prompt for knowledge retrieval; the large model extracts relevant knowledge from the knowledge graph through in-depth analysis of the prompts to generate failure analysis results.

[0045] Furthermore, in step 3, the large model prompt learning method is:

[0046] Input the question, and then provide a batch of examples from the analog pool and the supplementary pool to the large model; repeat this step until the input sequence length limit is reached;

[0047] Restate the query text and input the original triple sequence, and use the large model to re-rank the top candidate entities in the original triple sequence;

[0048] The answer given by the large model is parsed into an ordered list, replacing the first few entities in the original triple sequence as the final answer, thereby completing the knowledge reasoning task.

[0049] The present invention also proposes a large model retrieval-enhanced design failure mode and effect analysis system based on knowledge graph, including a data acquisition system, a design failure mode and effect analysis system and a user interaction system;

[0050] The data acquisition system is used to acquire data and includes a data center, a bill of materials (BOM) system, and a product lifecycle management (PLM) system;

[0051] The design failure mode and effects analysis system obtains failure analysis results using the design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement as described above, performs risk analysis and optimization on the failure analysis results, and finally documents and stores the analysis results in the FMEA knowledge base;

[0052] The user interaction system is used to display the analysis results to the user.

[0053] The beneficial effects of the present invention are:

[0054] The design failure mode and effect analysis method based on knowledge graph and large model retrieval enhancement has the following technical effects:

[0055] (1) Improve the objectivity and comprehensiveness of failure analysis. By leveraging the deep analytical capabilities of knowledge graphs and large models, we can effectively integrate multiple information sources such as design plans, design standards, two-dimensional models, and failure cases to generate comprehensive and accurate failure analysis results. This overcomes the problem of traditional FMEA being heavily dependent on personal experience and data silos, ensuring continuous optimization and updating of DFMEA information throughout the product life cycle, and improving the reliability and safety of product design and manufacturing processes.

[0056] (2) Achieve dynamic management and application throughout the entire life cycle. The system integrates with the data center, bill of materials (BOM) system, and product lifecycle management (PLM) system to form a full-cycle knowledge base covering the entire product life cycle. FMEA information can be continuously updated and optimized during product design, manufacturing, use, and maintenance. This full-cycle dynamic management approach ensures that failure modes and improvement measures can reflect the current product status in real time, improving product reliability and safety.

[0057] (3) Improve the efficiency and accuracy of information retrieval and failure analysis. The large-model retrieval enhancement method based on the knowledge graph significantly improves the efficiency of engineering design and problem solving through rich semantic understanding and rapid information retrieval. Combined with the large-model prompt learning method, the system can extract the most relevant information from the knowledge graph, generate failure analysis results, assess risks, and propose optimization measures. This efficient and accurate information processing capability makes the failure analysis process faster and more reliable, improving overall operation and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0059] Figure 1 A flowchart of the design failure mode and effect analysis method based on knowledge graph and large model retrieval enhancement of the present invention;

[0060] Figure 2 Develop a technology roadmap for designing failure mode and effects analysis;

[0061] Figure 3 A seven-step flowchart for failure mode and effects analysis;

[0062] Figure 4 Flowchart of the enhanced method for large model retrieval;

[0063] Figure 5 Flowchart for cue learning of large models;

[0064] Figure 6 Flowchart of the software-based approach for industrial knowledge;

[0065] Figure 7 A flowchart for constructing a hierarchical multimodal industrial knowledge ontology system;

[0066] Figure 8 It is a structural diagram of the hierarchical ontology system;

[0067] Figure 9 Schematic diagram of the industrial knowledge mining framework for modal feature-natural language description generation;

[0068] Figure 10 Schematic diagram for cross-modal industrial knowledge fusion and industrial knowledge graph formation throughout the product life cycle;

[0069] Figure 11 This is a schematic diagram of the knowledge graph-based large model retrieval enhanced design failure mode and effect analysis system of the present invention. DETAILED DESCRIPTION

[0070] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0071] like Figure 1-3 As shown, the design failure mode and effect analysis method based on knowledge graph and large model retrieval enhancement in this embodiment includes the following steps.

[0072] Step 1: Planning and preparation phase

[0073] Carry out planning and preparation, clarify analysis objectives, and form a multidisciplinary team.

[0074] In the initial stage of DFMEA analysis, detailed planning and preparation are crucial. Specifically, in this embodiment, the planning and preparation steps are as follows:

[0075] 11) Clarify the objectives and scope of FMEA analysis and ensure that the basic information of the project is complete;

[0076] 12) Establish a multidisciplinary team including design, manufacturing, quality and service to ensure the comprehensiveness and professionalism of FMEA analysis;

[0077] 13) Develop a detailed project plan, clarify the time schedule, resource allocation and responsible persons for each task, and ensure that the FMEA analysis is carried out as planned.

[0078] Step 2: Structural analysis and functional analysis stage

[0079] Conduct structural analysis of the product, identify the functions of each subsystem, assembly and component and their interrelationships, and fully understand the design intent of the product and the interaction of its parts.

[0080] The purpose of the structural analysis phase is to identify the various components of the product and their interrelationships, laying the foundation for subsequent failure analysis. Specifically, structural analysis methods include: Design Failure Mode and Effects Analysis (DFMEA) identifies and defines the product structure, clarifying the interface design of each subsystem, component, and part; and Process Failure Mode and Effects Analysis (PFMEA) identifies key steps in the manufacturing and assembly process to ensure the stability of process inputs and outputs. This helps provide a comprehensive understanding of the product's structure and function, enabling more effective failure analysis.

[0081] The purpose of functional analysis is to identify the functions and requirements of a product or process, ensuring that each function is fully considered in subsequent steps. Specifically, the functional analysis method is as follows: in the design failure mode and effects analysis, the functions and characteristics of the product are defined to ensure that the design can meet the expected performance and reliability requirements; in the process failure mode and effects analysis, the key functions of the process are identified to ensure that each step can achieve its intended function and avoid failures caused by process variation.

[0082] Step 3: Failure Analysis Phase

[0083] Input design plans, design standards, two-dimensional models and existing failure case information, use graph knowledge reuse components to process the input data and retrieve relevant information from the full-cycle knowledge base, and then process and construct the knowledge graph through the industrial knowledge software system; the large model retrieval enhancement method based on the knowledge graph deeply analyzes the input information, and combines the large model prompt learning method to extract relevant information from the knowledge graph to generate failure analysis results including failure impact, failure mode and failure cause.

[0084] Failure analysis is the core step of FMEA. It provides a basis for subsequent risk assessment and optimization by identifying potential failure modes, failure effects, and failure causes. In this embodiment, the failure analysis method steps are as follows:

[0085] 31) Identify possible failure modes of products or processes, such as fracture, wear, corrosion, etc.;

[0086] 32) Evaluate the potential impact of each failure mode on the system, subsystem, component, and user;

[0087] 33) Analyze potential causes of failure modes, such as design defects, material problems, improper manufacturing processes, etc.

[0088] The failure analysis step helps to fully understand the potential failure risks and their root causes.

[0089] Step 4: Risk Analysis and Optimization Phase

[0090] Evaluate the severity, frequency of occurrence and detection difficulty of each failure mode, calculate the risk priority number, and determine the issues that need to be addressed first; propose and implement improvement measures including design improvements, process adjustments and increased detection methods to reduce or eliminate high-risk failure modes.

[0091] During the risk analysis phase, the severity, frequency, and detection difficulty of each failure mode are evaluated to calculate the risk priority number (RPN) and determine the issues that need to be addressed first. Specifically, in this embodiment, the risk analysis method steps are as follows:

[0092] 41) Identify and develop preventive measures to reduce the probability of failure;

[0093] 42) Identify and implement detection measures to improve the ability to detect failures after they occur;

[0094] 43) Score the severity S, frequency O and detection difficulty D of each failure mode, and re-evaluate the risk priority number based on the implemented prevention and detection measures to ensure that the risks are effectively controlled.

[0095] The purpose of the optimization phase is to propose and implement improvement measures to reduce or eliminate high-risk failure modes. Specifically, the risk optimization method steps are:

[0096] 44) Identify and record high-risk failure modes that require improvement;

[0097] 45) Propose specific optimization measures including design improvement, material replacement and process adjustment;

[0098] 46) Implement optimization measures and verify their effectiveness;

[0099] 47) Score and evaluate optimization measures to ensure they significantly reduce risk priority numbers.

[0100] The optimization phase helps to continuously improve the reliability and safety of a product or process.

[0101] Step 5: Result documentation and system integration stage

[0102] The analysis results including failure modes, failure effects, failure causes, risk assessment and optimization measures are documented and stored in the FMEA knowledge base. By integrating with the data center, bill of materials (BOM) system and product lifecycle management (PLM) system, a full-cycle knowledge base is formed, and the full-cycle knowledge base is continuously updated and optimized throughout the product lifecycle.

[0103] The purpose of results documentation is to document all results and conclusions of FMEA analysis and store them in an FMEA knowledge base. By integrating with the data center, bill of materials (BOM) system, and product lifecycle management (PLM) system, comprehensive information sharing and utilization is achieved. Detailed information from the FMEA analysis, including failure modes, failure effects, failure causes, risk assessments, and optimization measures, is recorded. A control plan for the product or process is developed to ensure effective control at every stage. FMEA analysis results are incorporated into the knowledge base to form a full-cycle knowledge base, ensuring continuous updating and optimization throughout the product lifecycle.

[0104] Regarding the failure analysis stage, the following describes the large model retrieval enhancement method based on knowledge graph, the large model prompt learning method, and the industrial knowledge software method.

[0105] (1) Large model retrieval enhancement method based on knowledge graph

[0106] The large model retrieval enhancement based on the knowledge graph is an intelligent knowledge retrieval method that combines large-scale pre-trained language models and knowledge graph technology. The knowledge graph is a structured knowledge base that represents knowledge units and their relationships in the form of nodes and edges. The large model has powerful natural language processing and understanding capabilities, and can learn and extract knowledge from large amounts of data. By combining the knowledge graph with the large model, enhancements in rich semantic understanding, fast information retrieval, dynamic knowledge expansion, and deep knowledge association can be achieved. The large model retrieval enhancement method based on the knowledge graph achieves efficient knowledge retrieval and failure analysis by combining the powerful natural language processing capabilities of the large-scale pre-trained language model and the structured knowledge base of the knowledge graph. Specifically, in this embodiment, the large model retrieval enhancement method based on the knowledge graph is: starting from the design plan and standards, a problem combination representation is generated, and passed to the large model through Prompt for knowledge retrieval; the large model extracts relevant knowledge from the knowledge graph through in-depth analysis of the prompts, and generates failure analysis results, including failure impacts, patterns, and causes. This method can quickly and accurately provide rich semantic understanding and information retrieval, improving the efficiency and reliability of engineering design and problem solving, such as Figure 4 shown.

[0107] (2) Large model prompt learning method

[0108] Hint learning is a contextual learning strategy designed for knowledge graph completion tasks. By encoding part of the knowledge graph as examples, it significantly improves the performance of large models (LLM) on knowledge reasoning tasks. In this process, knowledge hints use the analogy pool (Da) and the supplementary pool (Ds) to construct examples for each query (h, r, ?), thereby providing richer and more relevant context information, such as Figure 5 shown.

[0109] The analogy pool Da contains triplets that have the same relationship as the query. These triplets help the LLM better understand the semantics of the query through analogy. To ensure the diversity of examples and allow the LLM to learn from a variety of analogy examples, the triplets in the analogy pool Da are selected using a sorting strategy that promotes diversity. Specifically, a zero counter is set for each entity, a triplet in Da is randomly selected as an example, and the counter of its related entity is incremented by 1. Subsequently, the triplet with the smallest related entity counter value is iteratively selected as an example, and its counter value is incremented accordingly. This process is repeated until all triplets in Da are used up, and the final list of examples is recorded as La.

[0110] The supplementary pool Ds provides supplementary information about the query head entity h, focusing on selecting examples relevant to the query. Specifically, all triples in Ds are ranked using the BM25 score, which is used to assess the textual relevance of each example to the query. The resulting list, denoted as Ls, contains all triples in Ds. Queries with specific relations or head entities use shared analogies or supplementary pools. To prevent duplicate calculations, these example pools are created and ranked during the data preprocessing phase. During inference, the corresponding example pool is used based on the query.

[0111] In practice, a unified prompt template is used to convert queries and examples into plain text with the same format. Multiple rounds of interaction guide the large model in re-ranking. During this process, the large model summarizes natural text descriptions from given examples using the original descriptions, generating easy-to-understand prompt templates and achieving self-alignment of the text. At this point, the large model processes the question, ranks candidate answers based on plausibility, and reviews the feedback to ensure accurate understanding of the task.

[0112] Specifically, in this embodiment, the large model prompt learning method is: input a question (corresponding to the query text), and then provide a batch of examples from the analogy pool and the supplementary pool to the large model; repeat this step until the input sequence length limit is reached; restate the query text, and input the original triple sequence, and use the large model to re-rank the top candidate entities in the original triple sequence; parse the answer given by the large model into an ordered list, replace the first few entities in the original triple sequence as the final answer, thereby completing the knowledge reasoning task.

[0113] This large-model-based hint learning method greatly enhances the performance of large models in complex knowledge reasoning tasks by making full use of analogies and supplementary information in knowledge graphs.

[0114] (3) Industrial knowledge software method

[0115] The industrial knowledge software system uses industrial knowledge software methods to process data, specifically, Figure 6 As shown, the industrial knowledge softwareization method includes the following steps.

[0116] S1: Obtain the sources and needs of multimodal industrial data throughout the product life cycle, and perform cluster analysis on the multimodal industrial data.

[0117] Specifically, in this embodiment, the method steps for cluster analysis of multimodal industrial data are as follows:

[0118] S11) Analyze the sources, knowledge carriers, and application requirements of multimodal industrial data, perform data cleaning, relationship definition, combination induction, and integration processing on multimodal industrial knowledge sources, build a full-cycle multimodal industrial database, and realize abnormal data cleaning and multi-source heterogeneous data integration of multimodal industrial knowledge sources.

[0119] S12) Based on the multimodal industrial data types and the adapted similarity metric, combined with the full-cycle multimodal industrial knowledge software requirements, a bottom-up agglomerative multimodal clustering algorithm is selected.

[0120] S13) Initialize the multimodal industrial data, regard each data point as an independent cluster, and globally measure the distance or similarity matrix between all clusters.

[0121] S14) judging whether the distance or similarity between adjacent clusters meets the inter-cluster merging condition: if so, executing step 15); if not, the inter-cluster merging is terminated, a clustering hierarchy is constructed, and executing step 16).

[0122] In this embodiment, the inter-cluster merging condition is that the distance or similarity threshold is less than or equal to a preset distance threshold. In this embodiment, the distance is used to judge whether adjacent clusters meet the inter-cluster merging condition. The distance includes the nearest neighbor distance (single chain method), the farthest neighbor distance (full chain method) and the average distance (average chain method).

[0123] S15) Select two clusters with the smallest distance or the smallest similarity to merge, and update the distance or similarity matrix between clusters to reflect the relationship between the newly formed cluster and other clusters; and execute step 14) repeatedly.

[0124] S16) Using a dendrogram to represent the hierarchical clustering results of full-cycle multimodal industrial data, to enhance the understanding of the gradual aggregation process of the data and the hierarchical structure of the clusters.

[0125] S2: Based on the clustering results of multimodal industrial data, existing ontologies and standards are reused and mapped to build a hierarchical multimodal industrial knowledge ontology system.

[0126] In this embodiment, the product life cycle multimodal industrial knowledge ontology system is based on the hierarchical clustering results. The specific ontology system construction process is as follows: Figure 7 As shown. The ontology system construction method of this embodiment is an improved seven-step method, which is an improvement based on the Stanford seven-step method. Specifically, in this embodiment, the method steps for constructing a hierarchical multimodal industrial knowledge ontology system are:

[0127] S21) The ontology domain is clearly defined as the industrial domain, and the application purpose of the ontology is clearly defined as providing top-level guidance for the mining and softwareization of industrial knowledge.

[0128] S22) Check whether there are reusable concepts, mainly involving existing ontologies and standards: if there are reusable ontologies, modify the reusable ontologies; for reusable standards, map the required content to standard ontologies based on the needs of industrial knowledge software.

[0129] S23) Domain experts evaluate the reused ontologies and standard ontologies based on the evaluation criteria, and introduce the reused ontologies and standard ontologies that meet the requirements into the reused ontology library. For ontologies that do not meet the requirements, they are revised or mapped again until they meet the requirements or are eliminated.

[0130] S24) Match the concepts in the reused ontology library with the full-cycle multimodal industrial data clustering results. For the clustering results and related concepts without reused ontology, determine the important terms of the industrial knowledge ontology system with the guidance of domain experts.

[0131] S25) Based on the hierarchical clustering results, define the hierarchical structure of the ontology classes and evaluate the ontology hierarchical structure.

[0132] S26) Determine the ontology attributes and their properties, and perform detailed definition and facet processing on the attributes.

[0133] S27) Build a bridging ontology to integrate ontologies of different levels, modalities, and types into a multimodal industrial knowledge ontology system for the entire product life cycle.

[0134] like Figure 8 As shown, in this embodiment, the multimodal industrial knowledge ontology system of the product's entire life cycle is a hierarchical ontology system, which is divided into three layers. The top layer includes the broadest and most abstract industrial concepts, such as "productivity", "automation", "sustainable development", etc., involving the core concepts and universal principles of the entire industrial field; the second layer is close to the actual industrial scene, but still maintains a certain degree of abstraction, including specific production processes, industrial machine types, production efficiency and other concepts; the bottom layer is targeted at specific industrial application scenarios, including the operating details of specific machines, specific maintenance guidelines, troubleshooting procedures, etc.

[0135] S28) Create an instance based on the defined ontology and attributes.

[0136] S3: Under the guidance of the multimodal industrial knowledge ontology system, multimodal industrial data is converted into text modality using text as a modal bridge; multimodal industrial data knowledge mining is realized using text knowledge mining methods to form triple clusters of industrial knowledge in each modality.

[0137] like Figure 9 As shown, in this embodiment, the method steps for multimodal industrial data knowledge mining using the text knowledge mining method are as follows:

[0138] S31) To address the challenges of standardizing knowledge mining methods for data from different modalities, given the diverse presentation methods, types, and carriers of industrial data throughout the product lifecycle, we explore natural language description methods for multimodal industrial data and establish an industrial knowledge mining framework based on consistent generation of modal features and natural language descriptions. Specifically, the industrial knowledge mining framework constructed in this embodiment targets five common modalities in industrial sites: text, images, graphics, signals, and formulas. However, the concept of this mining framework can be extended to other modal data.

[0139] S32) Utilizing an industrial knowledge mining framework generated by modal feature-natural language description consistency to convert non-text modal data into text or graph form, and implementing multimodal knowledge mining through text extraction methods, including:

[0140] For text modality: First, the PL-Marker strategy is used to encode text sentences or paragraphs, then Albert is used for embedding representation, and finally a hypergraph is used to model the embedded representation, and a hypergraph neural network is constructed to extract text modality knowledge and form a text modality triple group.

[0141] Note: There are various methods for text modality extraction. This embodiment is only an example. In this framework, the text modality knowledge extraction method is the most basic method and is applied to other modalities.

[0142] For image modality: select several multi-image description models, perform image description tasks on the same image, obtain multiple image description texts, embed and represent the description texts and images, perform feature recognition, and rank the description texts by quality. The top-k description texts are combined into prompts and input into the large language model to generate a fused image description. Use text modality knowledge extraction methods to extract image modality industrial knowledge and form image modality triple groups.

[0143] For graphic modalities: first, all 3D models are converted into standard formats such as OBJ or STEP, and then the PCL library is introduced to perform point cloud processing on the 3D models to obtain the 3D point cloud of the model; then, 3D object detectors such as Pointnet++ are used to obtain point cloud-related features, and point cloud description text is generated based on multi-head cross-attention; text modal knowledge extraction methods are used to realize graphic modal industrial knowledge extraction and form graphic modal triple groups.

[0144] For signal modality: first, the signal data is stored in a standard database such as CSV, and then the time-frequency domain statistical methods and signal processing methods such as Fourier transform and wavelet transform are called to process the signal, and the typical time-frequency domain features of the signal are extracted; under the guidance of prior knowledge such as sensors, signal-text association rules are constructed; the text modality knowledge extraction method is used to realize the extraction of signal modality industrial knowledge and form a signal modality triple group.

[0145] Regarding expression modality: Expression data is categorized into unstructured and semi-structured based on the different forms in which expressions exist. Unstructured data primarily refers to data in the form of images, while semi-structured data refers to data in the form of expressions such as latex and XML. Specifically, for semi-structured expression data, discrete information is directly extracted and a graph structure is formed. For unstructured expression data, expressions in images are first identified through OCR, then discretized. With human intervention, a scene graph of discrete elements is constructed, ultimately forming a group of expression modal triples.

[0146] Note: All text knowledge extraction is carried out under the guidance of the ontology system, and the parts that require human participation are also operated in accordance with the system specified by the ontology.

[0147] S4: Globally encode the triplet clusters of industrial knowledge of each modality, and perform modal fusion and cross-modal alignment to achieve cross-modal industrial knowledge fusion and form a multimodal industrial knowledge graph for the entire product life cycle.

[0148] After the multimodal industrial data passes through the industrial knowledge mining framework generated by multimodal features and natural language descriptions, each modality forms its own modal triple group. In order to better achieve the complementarity of each modality's industrial knowledge, cross-modal industrial knowledge fusion is required. Figure 10As shown. In this embodiment, the method for forming a multimodal industrial knowledge graph for the entire life cycle of a product is as follows: based on the association relationship of the modal triple groups, the modal original data features are introduced, and the modal triple group graph features such as the graph topology information, relationship information and attribute information of each triple cluster are combined and globally encoded, and weights are assigned to each part of the features through dynamic cross-modal weighting, and then modal fusion and cross-modal alignment operations are performed to achieve cross-modal knowledge fusion and form a multimodal industrial knowledge graph for the entire life cycle of the product. The multimodal industrial knowledge graph for the entire life cycle of a product uses the URI number as the core node, and is loaded with text attribute descriptions (i.e., information on each modal triple cluster), modal attribute descriptions (i.e., the number of modalities involved under the node and related information), standardized data attributes (i.e., the standardized data that can represent the node, and non-text data can be represented by URL mounting) and feature attributes (i.e., abstract features extracted by various algorithm models in the mining framework).

[0149] S5: Using the multimodal industrial knowledge graph of the entire product life cycle as the foundation for knowledge reuse, it supports the personalized configuration and development of industrial software for ubiquitous deployment of multi-scenario intelligent business, and realizes the software-based standard process of industrial knowledge acquisition, mining, and reuse.

[0150] like Figure 11 As shown, this embodiment also proposes a large model retrieval-enhanced design failure mode and effect analysis system based on knowledge graph, including a data acquisition system, a design failure mode and effect analysis system and a user interaction system. Specifically, the data acquisition system is used to acquire data and includes a data center, a bill of materials BOM system and a product lifecycle management PLM system. After the design failure mode and effect analysis system obtains the failure analysis results using the design failure mode and effect analysis method based on knowledge graph and large model retrieval-enhanced as described in this embodiment, it performs risk analysis and optimization on the failure analysis results, and finally documents the analysis results and stores them in the FMEA knowledge base. The user interaction system is used to display the analysis results to the user.

[0151] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement, characterized by: The steps include: Step 1: Planning and preparation phase Conduct planning and preparation, clarify analysis objectives, and form a multidisciplinary team; Step 2: Structural analysis and functional analysis stage Conduct structural analysis of the product, identify the functions of each subsystem, assembly and component and their interrelationships, and fully understand the design intent of the product and the interaction between its parts; Step 3: Failure Analysis Phase Input design plans, design standards, two-dimensional models, and existing failure case information. Use the graph knowledge reuse component to process the input data and retrieve relevant information from the full-cycle knowledge base. Then, the industrial knowledge software system processes and constructs a knowledge graph. The large-model retrieval enhancement method based on the knowledge graph deeply analyzes the input information and combines it with the large-model prompt learning method to extract relevant information from the knowledge graph to generate failure analysis results including failure impact, failure mode and failure cause; Step 4: Risk Analysis and Optimization Phase Assess the severity, frequency and detection difficulty of each failure mode, calculate the risk priority number, and determine the issues that need to be addressed first; Propose and implement improvement measures including design improvements, process adjustments and additional testing methods to reduce or eliminate high-risk failure modes; Step 5: Result documentation and system integration stage The analysis results including failure modes, failure effects, failure causes, risk assessment and optimization measures are documented and stored in the FMEA knowledge base. By integrating with the data center, bill of materials (BOM) system and product lifecycle management (PLM) system, a full-cycle knowledge base is formed, and the full-cycle knowledge base is continuously updated and optimized throughout the product lifecycle.

2. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In step 1, the planning and preparation steps are as follows: 11) Clarify the objectives and scope of FMEA analysis and ensure that the basic information of the project is complete; 12) Establish a multidisciplinary team including design, manufacturing, quality and service to ensure the comprehensiveness and professionalism of FMEA analysis; 13) Develop a detailed project plan, clarify the time schedule, resource allocation and responsible persons for each task, and ensure that the FMEA analysis is carried out as planned.

3. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In the step 2, the structural analysis method is: In the design failure mode and effect analysis, identify and define the product structure and clarify the interface design of each subsystem, component and part; In process failure mode and effects analysis, critical steps in the manufacturing and assembly process are identified to ensure the stability of process inputs and outputs.

4. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 3 is characterized by: In the step 2, the method of functional analysis is: In the design failure mode and effect analysis, define the functions and characteristics of the product to ensure that the design can meet the expected performance and reliability requirements; In process failure mode and effects analysis, key functions in the process are identified to ensure that each step can achieve its intended function and avoid failures caused by process variation.

5. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In step 3, the failure analysis method steps are: 31) Identify possible failure modes of products or processes; 32) Evaluate the potential impact of each failure mode on the system, subsystem, component, and user; 33) Analyze the potential causes of failure modes.

6. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In step 4, the risk analysis method steps are: 41) Identify and develop preventive measures to reduce the probability of failure; 42) Identify and implement detection measures to improve the ability to detect failures after they occur; 43) Score the severity S, frequency O and detection difficulty D of each failure mode, and re-evaluate the risk priority number based on the implemented prevention and detection measures to ensure that the risks are effectively controlled.

7. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 6 is characterized by: In step 4, the risk optimization method steps are: 44) Identify and record high-risk failure modes that require improvement; 45) Propose specific optimization measures including design improvement, material replacement and process adjustment; 46) Implement optimization measures and verify their effectiveness; 47) Score and evaluate optimization measures to ensure they significantly reduce risk priority numbers.

8. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In step 3, the large model retrieval enhancement method based on the knowledge graph is as follows: starting from the design plan and standards, a combination of question representations is generated and passed to the large model through prompts for knowledge retrieval; The large model conducts in-depth analysis of prompts, extracts relevant knowledge from the knowledge graph, and generates failure analysis results.

9. The design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement according to claim 1 is characterized by: In step 3, the large model prompt learning method is: Input the question, and then provide a batch of examples from the analog pool and the supplementary pool to the large model; repeat this step until the input sequence length limit is reached; Restate the query text and input the original triple sequence, and use the large model to re-rank the top candidate entities in the original triple sequence; The answer given by the large model is parsed into an ordered list, replacing the first few entities in the original triple sequence as the final answer, thereby completing the knowledge reasoning task.

10. A knowledge graph-based large model retrieval enhanced design failure mode and effect analysis system, characterized by: Including data acquisition system, design failure mode and effect analysis system and user interaction system; The data acquisition system is used to acquire data and includes a data center, a bill of materials (BOM) system, and a product lifecycle management (PLM) system; The design failure mode and effects analysis system obtains failure analysis results by using the design failure mode and effects analysis method based on knowledge graph and large model retrieval enhancement as described in any one of claims 1 to 9, performs risk analysis and optimization on the failure analysis results, and finally documents and stores the analysis results in a FMEA knowledge base; The user interaction system is used to display the analysis results to the user.

Images