A device knowledge management system and method

Abstract

The application discloses a kind of equipment knowledge management system and method;The system comprises: multi-modal data access module, for accessing multiple format files and extracting multi-modal elements;Knowledge processing module, for the structural treatment of extracted elements and constructs knowledge graph;Retrieval question and answer module, for executing user intent judgment based on user question input;According to the intention recognition result, when judging triggering question and answer, generate answer and its generation basis;Fault diagnosis module, for determining triggering fault diagnosis according to the intention recognition result, constructing fault tree according to knowledge graph and outputting troubleshooting scheme;Process application module, for determining triggering process application according to the intention recognition result, extracting BOM structure and process elements, generating process outline, multi-level process and SOP content, recommending associated resources based on process;Authority management module, for realizing access control based on directory and secret level.The application provides efficient and accurate equipment knowledge management platform for users.

Description

Technical Field

[0001] This application relates to the field of knowledge management technology, specifically to a device knowledge management system and method. Background Technology

[0002] With the acceleration of industrial modernization, enterprises are placing increasing emphasis on knowledge management throughout the entire equipment lifecycle. Equipment knowledge management systems play a crucial role in the operation of modern enterprises, widely serving typical scenarios in the equipment field such as knowledge management, fault diagnosis, maintenance optimization, and process application. This not only helps enterprises improve production efficiency and reduce costs but also enhances their core competitiveness.

[0003] In the past, to meet the needs of equipment knowledge management, the traditional approach was to handle related issues using multiple methods. For knowledge retrieval, keyword matching was typically relied upon, requiring users to accurately input keywords to obtain potentially relevant information. For multimodal data management, due to the lack of a unified processing mechanism, text, images, tables, and other data were often stored and managed separately without effective correlation. Fault diagnosis relied mainly on human experience and simple rules, lacking intelligent reasoning capabilities. The application of process knowledge usually required manual extraction of assembly procedures and parameter information from large amounts of process documents, resulting in low efficiency. Regarding knowledge security and access control, only simple user authentication and basic access control were typically implemented, lacking fine-grained management.

[0004] Existing technologies have significant shortcomings. Traditional knowledge retrieval methods struggle to understand users' natural language query intent, leading to inaccurate and incomplete search results and low efficiency. Current systems lack unified multimodal data processing and association capabilities, making it difficult to integrate multimodal knowledge data such as text, images, and tables related to equipment. Furthermore, the lack of intelligent reasoning capabilities based on knowledge graphs and fault trees hinders rapid and accurate fault location and root cause analysis, resulting in low levels of intelligent fault diagnosis. In addition, assembly procedures and parameter information in process documents are difficult to extract and structure quickly, leading to inefficient process optimization and maintenance decisions, and the lack of fine-grained data security levels and directory access management fails to meet enterprises' knowledge security management requirements. Summary of the Invention

[0005] To address the problems of low knowledge retrieval efficiency, difficulty in multimodal data fusion, low level of intelligent fault diagnosis, inconvenience in applying process knowledge, and insufficient knowledge security and access control in existing equipment knowledge management systems, and to better realize applications such as knowledge management, fault diagnosis, maintenance optimization, and process application, this application provides an equipment knowledge management system and method.

[0006] In a first aspect, this application provides an equipment knowledge management system, including: a multimodal data access module, a knowledge processing module, a retrieval and question answering module, a fault diagnosis module, a process application module, and a permission management module, with interactive connections established between the modules;

[0007] The multimodal data access module is used to access and parse files of various formats and extract text, images, and table elements;

[0008] The knowledge processing module is used to perform structured processing on the extracted elements and construct a knowledge graph;

[0009] The retrieval and question-answering module is used to perform user intent judgment based on user question input; when question-answering is triggered according to the intent recognition result, multi-path fact retrieval is performed, and the recalled factual evidence is fused to form a closed knowledge set; based on the closed knowledge set, a reasoning chain from fault phenomenon to solution is constructed; the user question, closed knowledge set, and reasoning chain are input as prompt words into the large language model to generate the answer and the basis for answer generation;

[0010] The fault diagnosis module is used to construct a fault tree based on the knowledge graph and output a fault elimination solution when the intent recognition result determines that fault diagnosis is triggered.

[0011] The process application module is used to extract the BOM structure and process elements from the extracted elements when the process application is triggered based on the intent recognition result, generate the process outline, multi-level process and SOP content, and recommend tooling, equipment and auxiliary materials based on the process.

[0012] The permission management module is used to implement access control based on directory and security level.

[0013] By adopting the above scheme, the system can access and parse multi-format files and extract elements, perform structured processing and knowledge graph construction on the elements, use the retrieval and question answering module to judge based on the user's question intent, recall factual evidence, construct reasoning chains and generate answers and evidence, use the fault diagnosis module to construct fault trees and output troubleshooting solutions based on intent, use the process application module to extract structural elements based on intent, generate process content and recommended resources, and rely on the permission management module to implement access control based on directory and security level, providing users with an efficient, convenient and accurate information acquisition platform.

[0014] Preferably, the retrieval question-answering module includes an intent recognition unit, a multi-way fact retrieval unit, and a closure set construction unit;

[0015] The intent recognition unit is used to perform user intent judgment based on user question input and obtain user intent; the user intent includes: question answering, fault diagnosis, and process application.

[0016] The multi-path fact retrieval unit, connected to the output of the intent recognition unit, is used to retrieve entities and relational paths related to the user's question from the knowledge graph, document fragments related to the user's question from the document vector library, and non-textual information related to the user's question from the multimodal feature library, thus completing the retrieval of structured facts and unstructured evidence. The document vector library is obtained by performing text slicing, word segmentation, and vectorization on the extracted text. The multimodal feature library is generated by extracting features from the extracted elements.

[0017] The closure set construction unit, connected to the output of the multi-way fact retrieval unit, is used to align and format the recalled structured facts and unstructured evidence, perform deduplication and preliminary similarity screening; based on the inverted index, the processed fact evidence is prioritized to obtain the Top-K fact evidence, and then reordered using a Cross-Encoder; according to the type of user question and the context of the fact evidence, the weights of the recalled fact evidence corresponding to different recall sources are dynamically adjusted, and the recalled fact evidence is merged to form a closed knowledge set.

[0018] By adopting the above scheme, an intent recognition unit is set up to accurately determine the user's intent, clarifying whether the user's intent is question answering, fault diagnosis, or process application, providing direction for subsequent operations; a multi-path fact retrieval unit is set up to retrieve relevant information from multiple channels such as knowledge graphs, document vector libraries, and multimodal feature libraries, completing the retrieval of structured facts and unstructured evidence, improving retrieval efficiency and comprehensiveness; a closure set construction unit is set up to process, sort, and adjust the weights of the recalled evidence, merging them to form a closed knowledge set, ensuring the accuracy and completeness of the knowledge set.

[0019] Preferably, the closure set construction unit is further configured to perform conflict detection on the processed factual evidence before prioritizing the processed factual evidence based on the inverted index, and combine the conflict detection results to ensure that the sum of the number of retained non-conflicting factual evidence and the number of retained conflicting factual evidence based on the conflict priority order reaches a preset number; based on the retained factual evidence, priority sorting, reordering and fusion are further performed; the conflict detection is performed by querying the corresponding domain factual evidence conflict template in the pre-built domain factual evidence conflict template library, comparing the corresponding domain factual evidence conflict template, and detecting whether there are semantic conflicts, numerical conflicts, or timeliness conflicts; for the conflicting factual evidence, the conflict priority order of the conflicting factual evidence is determined according to the factual evidence priority rules stored in the corresponding domain factual evidence conflict template.

[0020] By adopting the above scheme, conflict detection of factual evidence is performed when constructing a closed knowledge set, which effectively avoids erroneous information caused by evidence conflicts and improves the accuracy and reliability of the closed knowledge set. At the same time, retaining evidence according to the preset quantity and conflict priority can ensure the integrity and rationality of the knowledge set, enabling the system to respond to user needs more accurately.

[0021] Preferably, the retrieval question-answering module further includes a causal chain generation unit connected to the output of the closure set construction unit;

[0022] The causal chain generation unit is used to construct a factual evidence relationship graph based on a closed knowledge set; based on the constructed factual evidence relationship graph, it performs multiple path expansion, path scoring and filtering, domain rule application, and path-to-text evidence conversion to construct a reasoning chain from fault phenomena to solutions; the domain rule application includes representing domain rules as logical rules or graph patterns, matching candidate paths with domain rules, checking whether they meet the domain rules, correcting paths that do not meet the domain rules, and retaining corrected paths that meet the domain rules.

[0023] By adopting the above scheme, a reasoning chain from fault phenomena to solutions is constructed based on a factual closed knowledge set, which enhances the logical deduction ability in retrieval and question answering. Furthermore, by applying domain rules, paths that do not conform to domain rules are corrected, ensuring the rationality and effectiveness of the reasoning chain and improving the quality and reliability of the equipment knowledge management system in answering questions.

[0024] Preferably, the closure set construction unit is further configured to set a number of preset quantities that are adjacent to each other to form a number of different closed knowledge sets; the causal chain generation unit is further configured to construct fact evidence relationship graphs based on different closed knowledge sets respectively.

[0025] The retrieval question-answering module also includes an answer generation unit connected to the output of the causal chain generation unit;

[0026] The answer generation unit is also used to input the user question, each closed knowledge set, and the reasoning chain corresponding to each closed knowledge set as prompt words into the large language model to generate an answer; and to output the final answer and the basis for answer generation by performing cluster analysis on multiple answers and analyzing the centroid of the cluster with the largest weight.

[0027] By adopting the above scheme, the system can form multiple different closed knowledge sets, construct multiple fact-evidence relationship graphs, generate multiple answers based on multiple prompt words, and output the final answer and its generation basis through cluster analysis, thereby improving the accuracy and reliability of the answer.

[0028] Preferably, it also includes: a collaborative triggering unit that is connected to the retrieval question and answer module, the fault diagnosis module, and the process application module;

[0029] The collaborative triggering unit is used to receive intent recognition results, user questions, and generated answers. It uses a large language model to analyze and determine whether to recommend fault diagnosis or process application. If fault diagnosis or process application is recommended, a corresponding fault diagnosis module or process application module is generated and triggered. When the fault diagnosis module is triggered, it executes the construction of a fault tree based on the knowledge graph and outputs a fault elimination solution. When the process application module is triggered, it executes the extraction of BOM structure and process elements from the extracted elements, generates process outline, multi-level processes and SOP content, and recommends tooling, equipment and auxiliary materials based on the processes.

[0030] By adopting the above scheme, the collaborative triggering unit, combined with the large language model, intelligently determines whether to recommend fault diagnosis or process application based on the intent recognition results, user questions, and generated answers, and triggers the corresponding modules to achieve collaboration between modules, thereby improving the system's intelligence level and responsiveness to user needs.

[0031] Preferably, the fault diagnosis module is further configured to detect in real time whether the knowledge graph is updated during the process of constructing a fault tree based on the knowledge graph and outputting a fault troubleshooting solution; when an update to the knowledge graph is detected, determine whether there is a correlation between the updated knowledge graph node and the fault node in the fault tree; if there is a correlation, trigger the reconstruction of the fault tree to re-output the fault troubleshooting solution.

[0032] By adopting the above scheme, when the fault diagnosis module constructs the fault tree and outputs troubleshooting solutions, it responds promptly to the update of the knowledge graph, ensuring that the fault tree is associated with the latest knowledge graph, so that the troubleshooting solutions are always based on the latest knowledge, thereby improving the accuracy and timeliness of fault diagnosis.

[0033] Preferably, the process application module, which interfaces with the pre-built equipment digital twin model, is also used to input tooling, equipment and auxiliary materials recommended based on the process as input parameters into the equipment digital twin model, so that simulation operation is performed in the equipment digital twin model to obtain the control range of the key process parameters and provide feedback.

[0034] By adopting the above scheme, the process application module is set up to interface with the pre-built equipment digital twin model. The tooling, equipment and auxiliary materials recommended based on the process are used as input parameters. The simulation is run in the equipment digital twin model to obtain the control range of the key process parameters and provide feedback, thereby improving the accuracy and efficiency of process control.

[0035] Preferably, the knowledge processing module is further used to perform semantic association analysis on the constructed knowledge graph and generate recommendation questions by combining historical user question input data.

[0036] By adopting the above approach, semantic association analysis is performed on the constructed knowledge graph, and recommended questions are generated by combining historical user question input data to assist users in knowledge querying and improve knowledge retrieval efficiency.

[0037] Preferably, this application provides a device knowledge management method, including:

[0038] It can access and parse files of various formats and extract text, images, and table elements;

[0039] The extracted elements are structured and a knowledge graph is constructed.

[0040] Perform user intent determination based on user question input;

[0041] When the intent recognition result determines that a question-answering is triggered, multi-path fact retrieval is performed, and the recalled factual evidence is fused to form a closed knowledge set. Based on the closed knowledge set, a reasoning chain from the fault phenomenon to the solution is constructed. The user's question, the closed knowledge set, and the reasoning chain are input as prompt words into the large language model to generate the answer and the basis for the answer generation.

[0042] When fault diagnosis is triggered based on the intent recognition result, a fault tree is constructed based on the knowledge graph and a fault elimination solution is output.

[0043] When the process application is triggered based on the intent recognition results, the BOM structure and process elements are extracted from the extracted elements to generate the process outline, multi-level processes and SOP content, and tooling, equipment and auxiliary materials are recommended based on the process.

[0044] By adopting the above scheme, we can achieve effective management and utilization of multimodal data, improve the efficiency and accuracy of knowledge retrieval, and realize intelligent fault diagnosis and process application.

[0045] In summary, this application has the following beneficial effects:

[0046] By setting up the system to access and parse files of various formats and extract elements, unified processing of multimodal data is achieved. The system also structures these elements and constructs a knowledge graph, providing a data foundation for subsequent retrieval, diagnosis, and application. Furthermore, the system determines user intent based on input questions and triggers different operations accordingly. When question answering is triggered, multi-path fact retrieval is performed, merging to form a closed knowledge set, constructing a reasoning chain, and inputting it into a large language model to generate answers and supporting evidence, improving knowledge retrieval efficiency. When fault diagnosis is triggered, a fault tree is constructed based on the knowledge graph, and troubleshooting solutions are output, enhancing the intelligence of fault diagnosis. When process application is triggered, process outlines, multi-level procedures, and SOP content are generated, and tooling and other resources are recommended to facilitate the application of process knowledge. Finally, the system implements directory-based and security-level access control, addressing the issues of insufficient knowledge security and access control, and providing users with an efficient, convenient, and accurate information acquisition platform. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the structure of the device knowledge management system described in a specific embodiment;

[0048] Figure 2 This is a functional structure diagram of the device knowledge management system described in a specific embodiment;

[0049] Figure 3 This is a flowchart of the device knowledge management method described in a specific embodiment;

[0050] Figure 4 This is a schematic diagram of the device knowledge management method described in a specific embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] like Figure 1 As shown in the illustration, this application discloses an equipment knowledge management system, including: a multimodal data access module 100, a knowledge processing module 200, a retrieval and question-and-answer module 300, a fault diagnosis module 400, a process application module 500, an access control module 600, and other general modules 700. These modules establish communication and interactive connections to jointly achieve efficient management of equipment knowledge, thereby enhancing the enterprise's capabilities in equipment knowledge management, fault diagnosis, and process application. The content of each module will be described in detail below.

[0053] Firstly, the multimodal data access module 100 is used to access and parse files of various formats and extract text, images, and table elements.

[0054] Specifically, this module supports uploading various file formats, including PDF, Word, PPT, Excel, TXT, Markdown, images, audio, and video. Taking file upload as an example, common web front-end upload plugins such as FineUploader can be used, supporting single-file uploads, batch uploads, and resume uploads. For parsing various file formats, text parsing can use open-source text processing libraries, such as Python's NLTK library, to extract text content and perform text slicing, word segmentation, and vectorization. Image parsing can extract metadata, including title, abstract, source, shooting time, location, and device information, specifically using the OpenCV library to extract image features and metadata. Table parsing can use the Pandas library for convenient processing and analysis of tabular data. Furthermore, this module also supports the classification, organization, and maintenance of extracted text, images, and table elements for subsequent retrieval.

[0055] Secondly, the knowledge processing module 200 is used to perform structured processing on the extracted elements and construct a knowledge graph.

[0056] Specifically, this module achieves structured processing through various data processing operations. For text elements, it extracts triples from text fragments, constructs a "subject-relationship-object" structure, and stores it in a knowledge graph database. Simultaneously, it segments long texts into meaningful semantic fragments through text slicing, and transforms continuous text into structured word sequences through text segmentation, facilitating subsequent processing such as vector calculation to capture complex features and similarities between data, and determining data similarity by calculating the distance between vectors. For table elements, it performs structured parsing, extracting row and column relationships and numerical logic, and linking them to the knowledge graph. For image elements, it performs visual content understanding, extracting semantic information and associating it with relevant text fragments. Furthermore, based on GraphRAG technology, it enables entity linking, relationship querying, and graph visualization.

[0057] In addition, to assist users in knowledge retrieval, a knowledge processing module 200 can be set up. This module is also used to perform semantic association analysis on the constructed knowledge graph and generate recommended questions by combining historical user question input data. Specifically, for extended question recommendation generation: for the core entity E queried by the current user, the top-K entities / relationships with the highest semantic similarity to the core entity E are selected from the knowledge graph. Combined with the high-frequency historical query questions of these entities / relationships, an extended question list is finally generated. For completion question recommendation generation: for the incomplete query Q1 of the current user, the semantic similarity between the incomplete query Q1 and historical similar questions is calculated using Sentence-BERT. The historical complete questions with the highest similarity are selected. Combined with the corresponding entity attributes / relationships in the knowledge graph, the missing information of the incomplete query Q1 is completed, and a completed recommended question is generated. For high-frequency association question recommendation generation: for the current user query Q2, the top-M questions with the highest co-occurrence frequency with the current user query Q2 are selected from the historical question database. Combined with the semantic association of the knowledge graph, the co-occurring questions are re-ranked to generate a high-frequency association recommendation list. For the above three types of recommendation questions, the final recommendation list can be output by weighting and sorting according to semantic relevance (weight 0.6), historical click conversion rate (weight 0.3), and knowledge coverage (weight 0.1).

[0058] For example, extract BOM structure, constraint relationships, PMI annotations, component elements, and process experience from drawings and process data, and achieve generative retrieval through semantic fusion; generate process outlines and multi-level processes based on the BOM structure, support automatic generation of content and standard wording by process, and recommend resources such as tooling, equipment, and auxiliary materials through SOP arrangement.

[0059] Third, the retrieval question and answer module 300 specifically includes: an intent recognition unit 301, a multi-way fact retrieval unit 302, a closure set construction unit 303, a causal chain generation unit 304, and an answer generation unit 305.

[0060] The intent recognition unit 301 is used to perform user intent judgment based on user question input using natural language processing technology, such as a deep learning-based text classification model, to obtain the user intent. The user intent includes: question and answer, fault diagnosis, process application, and general application, etc. For example, if a user inputs: "The centrifugal pump vibrates too much, what could be the reason?", the user intent judgment will identify it as a question and answer intent, and then proceed with the corresponding operation of generating an answer.

[0061] The multi-path fact retrieval unit 302, connected to the output of the intent recognition unit 301, is used to perform multi-path fact retrieval when a question-and-answer session is triggered based on the intent recognition result. The multi-path fact retrieval includes: retrieving entities and relationship paths related to the user's question from the knowledge graph to obtain structured facts; and retrieving document fragments related to the user's question from the document vector library as evidence. Specifically, this can be combined with an inverted index for text fragment retrieval, completing the retrieval of both structured facts and unstructured evidence. The document vector library is obtained by processing the extracted text through text slicing, word segmentation, and vectorization.

[0062] To further enhance the accuracy of subsequent intelligent question answering and supplement the key evidence chain, the multi-path fact retrieval unit 302 is also used to retrieve non-textual information related to the user's question from the multi-modal feature library, thereby expanding the dual-path retrieval to a three-path retrieval, jointly completing the retrieval of structured facts and unstructured evidence. The multi-modal feature library is constructed by extracting features from the extracted elements. To assist in the subsequent fusion of recalled factual evidence, non-textual information such as images, audio, and sensor time-series data are incorporated into a unified retrieval space. Semantic alignment of heterogeneous data is achieved through cross-modal embedding models such as CLIP, for example, input device infrared thermal imaging images. The mapping is done as a feature vector, while simultaneously matching the description of "localized overheating" in the text report. Establish visual-text associations. An example will be provided below:

[0063] User input: Gearbox abnormal noise accompanied by oil temperature rise (text) + on-site vibration spectrum diagram (image);

[0064] Knowledge Graph Recall: Abnormal Sounds Gear wear Causal chain of oil temperature rise; Document vector library recall: maintenance record text of similar fault cases; Multimodal feature library recall: matching historical vibration spectrum library. Cases of anomalous resonance within the same frequency band were identified and correlated with oil ferrographic analysis images. Metal fragments were found.

[0065] The closure set construction unit 303 is connected to the output of the multi-way fact retrieval unit 302 and is used to fuse the recalled factual evidence to form a closed knowledge set.

[0066] To generate more accurate answers, the recalled factual evidence is preprocessed, including: converting the recalled factual evidence into a uniform text format through alignment and formatting; deduplication; and preliminary similarity screening. Specifically, by setting a similarity threshold, factual evidence is deduplicated based on semantic similarity, and low-similarity factual evidence is filtered out. Then, the preprocessed factual evidence is prioritized based on an inverted index, such as using a relevance scoring algorithm (BM25) to obtain the Top-K factual evidence. Finally, Cross-Encoder re-ranking is performed, i.e., a pre-trained Cross-Encoder model, such as a BERT-based sequence pair classification model, is used to interactively compute each processed factual evidence with the user question, obtaining a score for each piece of evidence, and then re-ranking them based on the scores.

[0067] Finally, based on the type of user question and the context of factual evidence, the weights of different recall sources are dynamically adjusted, and the recalled factual evidence is merged to form a closed knowledge set. The context of factual evidence includes: the statistical characteristics of the Cross-Encoder score of evidence in each recall source, the number of pieces of evidence in each recall source, and the type of recall source (knowledge graph, document, multimodal). The types of user questions include: fault diagnosis, operation steps, principle explanation, visual recognition, etc. A lightweight neural network is designed, taking the type of user question and the context of factual evidence as input, and outputting the weights of each recall source. The weights of the recalled factual evidence corresponding to different recall sources are dynamically adjusted, thereby merging the factual evidence recalled by the corresponding recall sources to form a closed knowledge set. For example, for fault diagnosis: knowledge graph weight = 0.6, document weight = 0.3, multimodal weight = 0.1, serving as the weights of the recalled factual evidence for the corresponding recall source.

[0068] The causal chain generation unit 304 is connected to the output of the closure set construction unit 303 and is used to construct a reasoning chain from fault phenomena to solutions based on a closed knowledge set.

[0069] Specifically, a fact-evidence relationship graph is constructed based on a closed knowledge set. Based on this graph, multi-hop path expansion, path scoring and filtering, domain rule application, and path-to-text evidence conversion are performed to build a reasoning chain from fault phenomena to solutions. The multi-hop path expansion includes: identifying fault phenomenon entities based on user question input; using these entities as nodes, performing multi-hop traversal along the relationship edges of the fact-evidence relationship graph; specifically, breadth-first search or depth-first search can be used for multi-hop traversal, while limiting the maximum number of hops to avoid path explosion; obtaining potential related entities and relationship paths to form candidate paths. The path scoring and filtering uses a pre-trained machine learning model for path scoring. Specifically, correctly validated paths from historical fault diagnosis cases are used as positive samples, and randomly generated or incorrect paths are used as negative samples to complete model training. Each candidate path is input, and paths with scores higher than a preset threshold are retained. The application of domain rules includes: representing domain rules as logical rules or graph patterns, such as converting logic gates in an FTA fault tree into constraints on paths; matching candidate paths with domain rules; checking whether the domain rules are satisfied; for AND gate rules, multiple cause nodes must appear simultaneously in the path; correcting paths that do not conform to the domain rules; and retaining corrected paths that conform to the domain rules. The path-to-text evidence conversion includes: converting the filtered paths into text evidence described in natural language, so that the text evidence can be combined into a complete reasoning chain for answer generation or direct display to the user. Examples are given below:

[0070] Starting from the node "excessive pump vibration", multiple paths are explored, leading to the following path:

[0071] Path 1: Excessive pump vibration Possible reasons Impeller imbalance Solution Dynamic balance;

[0072] Path 2: Excessive pump vibration Possible reasons Bearing wear Solution Replace the bearing;

[0073] After path scoring and filtering: Path 1 score: 0.9, Path 2 score: 0.7, two paths are retained;

[0074] Domain rule application: Apply FTA rules; both paths comply.

[0075] Evidence converted from path to text: Path 1 text: The possible cause of excessive pump vibration is impeller imbalance, and the solution is dynamic balancing; Path 2 text: The possible cause of excessive pump vibration is bearing wear, and the solution is to replace the bearing;

[0076] Combined textual evidence suggests that excessive pump vibration could be caused by two reasons. First, impeller imbalance can lead to excessive vibration, which can be resolved by dynamic balancing. Second, worn bearings can also cause excessive vibration, which can be resolved by replacing the bearings.

[0077] The answer generation unit 305 is connected to the output of the causal chain generation unit 304, and is used to input the user question, closed knowledge set, and reasoning chain together as prompt words into the large language model to generate the answer and the basis for answer generation.

[0078] Specifically, large language models can be engineered using instruction-based fine-tuning techniques, such as GPT-4. For example: Generate a structured answer, including: fault probability analysis, suggested troubleshooting steps, safety precautions, and links to relevant reference documents. Output: Structured answer with accompanying evidence citations. Furthermore, when generating answers, multimodal output is supported, such as single-text question-and-answer, image-based question-and-answer, and table-based question-and-answer formats, achieving multimodal answer output.

[0079] Fourth, the fault diagnosis module 400 is used to construct a fault tree based on the knowledge graph and output a fault elimination solution when the fault diagnosis is triggered based on the intent recognition result.

[0080] Specifically, the FTA fault tree is constructed, including: Top event confirmation and mapping: Based on the intent recognition results, core fault entities are extracted and mapped to fault event nodes in the knowledge graph, serving as the top event of the fault tree; simultaneously, the top-level BOM node corresponding to the fault event is associated, intermediate events are decomposed based on the BOM hierarchy, bottom events (root causes) are associated and supplemented, and the validity of the fault tree is verified, realizing the association between fault events and root causes through the BOM structure. Automatic fault diagnosis includes: Fault path extraction: Extracting all possible propagation paths from the top event to the bottom event of the fault tree; Quantitative calculation of troubleshooting priority: A weighted scoring model is used to rank the bottom events (root causes), with higher scores indicating higher troubleshooting priority; Fault solution generation: For the ranked root causes, "investigation steps + solutions" are generated sequentially, with content derived from the solution relationships in the knowledge graph. For example: "Investigate fuse blown → Step 1: Power off; Step 2: Open the power module casing; Step 3: Check fuse continuity; Solution: Replace with a fuse of the same specification." Visual topology display of the fault tree and troubleshooting paths.

[0081] Furthermore, to further ensure the effectiveness of real-time generation of troubleshooting solutions, this module continuously monitors whether the knowledge graph is being updated during the process of constructing the fault tree based on the knowledge graph and outputting the troubleshooting solution. This includes monitoring changes to the knowledge graph, such as newly added, modified, and deleted nodes and relationships. When an update to the knowledge graph is detected, it determines whether there is a relationship between the updated node in the knowledge graph and the fault node in the fault tree. If a relationship exists, it triggers the reconstruction of the fault tree to re-output the troubleshooting solution. Among them, the correlation detection can adopt multi-dimensional correlation judgment, such as: for knowledge graph updates: {new node, updated node, new relationship, updated relationship}, fault tree set: {fault tree 1, fault tree 2, ...}; graph structure analysis is used to determine whether there is a correlation, such as: if the newly added node is directly connected to a node in the fault tree in the knowledge graph, then there is a direct correlation; if the newly added node is connected to a node in the fault tree through multiple edges, and the relationship type on the path is the same as the relationship type used when the fault tree was constructed, then there is an indirect correlation; the attributes of the newly added node, such as fault type, equipment type, etc., match the attributes of the nodes in the fault tree; semantic correlation detection, that is, semantic similarity calculation and threshold comparison, determines whether there is a correlation.

[0082] Fifth, the process application module 500 is used to extract the BOM structure and process elements from the extracted elements when the process application is triggered based on the intent recognition result, generate a process outline, multi-level processes and SOP content, and recommend tooling, equipment and auxiliary materials based on the processes. Specifically, the fault diagnosis module and the process application module share the BOM structure data source. The process application module includes an SOP orchestration engine, which triggers a recommended list of associated resources based on process nodes. The resource list includes auxiliary material specifications, equipment models and operating safety specifications.

[0083] In addition, to further improve the accuracy and efficiency of process control, this module takes the tooling, equipment and auxiliary materials recommended based on the process as input parameters to the pre-built digital twin model of the equipment. This allows the simulation to run in the digital twin model, obtain the control range of key process parameters, and feed them back to the process application module for visualization.

[0084] Sixth, the permission management module 600 is used to implement access control based on directory and security level.

[0085] Specifically, this module allows setting different directory permissions and data security levels, controlling user access to different directories and data security levels based on user roles and permissions. For example, sensitive process knowledge and fault diagnosis information can be assigned a higher security level, allowing only users with the appropriate permissions to access them. This module also provides log management, terminology management, and user management functions to ensure the system's secure and compliant operation.

[0086] In addition, the system also includes 700 other general modules. For example, it supports general knowledge processing functions such as Chinese-English bilingual translation, text summarization, and keyword and opinion extraction. When other general applications are triggered based on intent recognition results, it completes Chinese-English bilingual translation, text summarization, and keyword and opinion extraction.

[0087] like Figure 2 As shown, by adopting the above system, the following functions can be achieved. The system can be viewed as a device knowledge base, which contains various functional modules, such as:

[0088] The chat function, or the application of the search and question-and-answer module, performs searches, asks and answers, and provides some auxiliary functions. Specifically, the search function can search for various forms of content such as text, tables, images, and file fragments. Search results are aggregated and filtered by dimensions such as type, time, and source. The question-and-answer function can provide diverse interactive formats such as single-article question-and-answer, image question-and-answer, and table question-and-answer. It supports GraphRAG technology to realize entity linking, multi-entity relationship and attribute query, and visualizes it in graph form. Auxiliary functions may include underlying supporting technologies such as text slicing, word segmentation, vector calculation, multi-way recall, and ranking models.

[0089] Business applications, namely the application of fault diagnosis module, process application module and other general modules, realize intelligent troubleshooting, automatically recommend tooling equipment and auxiliary materials, support Chinese and English bilingual fragment / document translation, and compare translation results of multiple models; at the same time, it provides text summarization, keyword and viewpoint automatic extraction functions;

[0090] The knowledge base, i.e., the application of the multimodal data access module, supports multi-classification management and maintenance of files;

[0091] The backend management module, or access control module, supports user management, access and security level management, log and terminology management, etc. The knowledge processing module supports multimodal data processing and question recommendation.

[0092] Therefore, it can be seen that through the collaborative work of multiple modules in the system, this equipment knowledge management system achieves full lifecycle management of equipment knowledge. The design and implementation of the entire system effectively solves the problems existing in traditional equipment knowledge management, such as low knowledge retrieval efficiency, difficulty in multimodal data fusion, low level of intelligent fault diagnosis, inconvenience in the application of process knowledge, and insufficient knowledge security and access control, thereby improving the efficiency and competitiveness of enterprises in equipment management.

[0093] In one specific embodiment, considering the potential conflicts among the generated factual evidence, to ensure the quality and credibility of answers generated from a closed knowledge set formed based on the factual evidence, conflict detection is performed before prioritizing the processed factual evidence based on an inverted index. Conflicting factual evidence is excluded or reasonably retained, thereby improving the accuracy and effectiveness of the system's question-and-answer retrieval. The system also includes:

[0094] The closure set construction unit 303 is further configured to perform conflict detection on the processed factual evidence before prioritizing the processed factual evidence based on the inverted index. Based on the conflict detection results, the sum of the number of retained non-conflicting factual evidence {a, b, c...} and the number of retained conflicting factual evidence {e, f, g...} based on the conflict priority order reaches a preset number. The preset number can be a normal value N obtained by statistical analysis of the factual evidence involved in historical domain dialogues. Based on the retained factual evidence, the subsequent priority sorting, reordering and fusion are then completed.

[0095] The conflict detection process requires conflict resolution based on domain characteristics. Factual evidence priority rules are established according to the characteristics of different domains. This involves querying relevant domain factual evidence conflict templates in a pre-built domain factual evidence conflict template library, comparing them, and detecting semantic, numerical, or time-related conflicts. Specifically, semantic conflict refers to two pieces of evidence being contradictory or opposite in meaning. For conflicting factual evidence, numerical conflict refers to different values ​​for the same parameter, and time-related conflict refers to conflicts between old and new standards or data. The conflict priority order of conflicting factual evidence is determined according to the factual evidence priority rules stored in the relevant domain factual evidence conflict templates, ensuring that subsequent events retain higher-priority factual evidence based on its priority.

[0096] In a specific embodiment, to further improve the accuracy and reliability of the answer, multiple different closed knowledge sets are constructed, multiple fact-evidence relationship graphs are built, and multiple answers are generated based on multiple prompt words to obtain the optimal answer. The system also includes:

[0097] The closure set construction unit 303 is also used to set a number of preset quantities that are adjacent to each other (e.g., N-1, N, N+1, etc.) to form a number of different closed knowledge sets; the causal chain generation unit 304 is also used to construct fact evidence relationship graphs based on different closed knowledge sets.

[0098] The answer generation unit 305 is also used to input the user question, each closed knowledge set, and the reasoning chain corresponding to each closed knowledge set as prompt words into the large language model to generate an answer; by performing cluster analysis on multiple answers, one or more clusters are obtained, and the centroid of the cluster with the largest weight in the obtained clusters is analyzed to output the final answer and the basis for answer generation.

[0099] In a specific embodiment, in order to further enhance the comprehensiveness and intelligence of the equipment knowledge management system in responding to user needs, and to provide users with more practical troubleshooting solutions, process outlines and resource recommendations, the system further includes a collaborative triggering unit 800 that is connected to the retrieval question and answer module 300, the fault diagnosis module 400 and the process application module 500. This is achieved by completing the collaborative work between multiple modules and determining whether to trigger the fault diagnosis module or the process application module.

[0100] The collaborative triggering unit 800 is used to receive the intent recognition result, user question and generated answer, and use a large language model to analyze and determine whether to recommend fault diagnosis or process application; if the determination result is to recommend fault diagnosis or process application, a corresponding fault diagnosis module or process application module is generated to trigger the fault diagnosis module or process application module, so that when the fault diagnosis module is triggered, it executes to build a fault tree based on the knowledge graph and outputs a fault elimination solution, or when the process application module is triggered, it executes to extract the BOM structure and process elements from the extracted elements, generate process outline, multi-level process and SOP content, and recommend tooling, equipment and auxiliary materials based on the process.

[0101] Specifically, for example: When a user inputs, "What could be causing excessive vibration in a centrifugal pump?", the question-and-answer module 300 detects that the user's question falls under the category of fault diagnosis, and the search results may contain multiple possible causes. The collaborative triggering unit 800 receives the user's question and the search results, and through large language model analysis, suggests activating the fault diagnosis module. When the fault diagnosis module 400 receives the signal suggesting activation, it performs fault diagnosis, thus achieving collaborative cooperation between the question-and-answer module 300 and the fault diagnosis module 400. Similarly, collaborative cooperation is achieved between the question-and-answer module 300 and the process application module 500.

[0102] like Figure 3 As shown, this embodiment discloses a device knowledge management method using the above system, the specific steps of which include:

[0103] S1. Access and parse files of various formats and extract text, images and table elements.

[0104] In this step, a multimodal data access module is used to complete the operation. As mentioned earlier, appropriate tools and libraries, such as file upload components, text processing libraries, image parsing libraries, and table processing libraries, are used to upload and parse files of different formats, extracting the required text, images, and table elements.

[0105] S2. The extracted elements are structured and a knowledge graph is constructed.

[0106] Specifically, the extracted elements are structured through operations such as text slicing, word segmentation, vector computation, and triplet processing. For example, a text about equipment operating procedures is sliced ​​and segmented, its vector representation is calculated, and "subject-relationship-object" information is extracted, such as "operator-operation-equipment". Then, this structured information is integrated into a knowledge graph to construct a network of connections for equipment knowledge.

[0107] S3. Perform user intent judgment based on user question input.

[0108] Specifically, this step is completed by the intent recognition unit in the retrieval question-answering module. For example... Figure 4 As shown, natural language processing technology is used to analyze and classify user input questions to determine whether the user's intent is for question answering, fault diagnosis, or process application. For example, when a user inputs "What should I do if a certain piece of equipment malfunctions?", the system determines that the user's intent is fault diagnosis; when a user inputs "What are the parameters of a certain process?", the system determines that the user's intent is process application.

[0109] S4. When the intent recognition result determines that a question-answering is triggered, multi-path fact recall is performed, and the recalled factual evidence is integrated to form a closed knowledge set. Based on the closed knowledge set, a reasoning chain from the fault phenomenon to the solution is constructed. The user's question, the closed knowledge set, and the reasoning chain are input as prompt words into the large language model to generate the answer and the basis for the answer generation.

[0110] Specifically, this step is completed through the question and answer module, which generates the answer and the basis for its generation.

[0111] S5. When the fault diagnosis is triggered based on the intent recognition result, a fault tree is constructed based on the knowledge graph and a fault elimination solution is output.

[0112] Specifically, this step is performed by the fault diagnosis module, such as... Figure 4 As shown, a fault tree is constructed based on information from the knowledge graph. Then, based on the analysis of the fault tree, a solution to troubleshoot the fault is output.

[0113] S6. When the process application is triggered based on the intent recognition result, extract the BOM structure and process elements from the extracted elements, generate the process outline, multi-level processes and SOP content, and recommend tooling, equipment and auxiliary materials based on the process.

[0114] Specifically, this step is handled by the process application module. For example... Figure 4 As shown, when the system determines that the user's intent is a process application, it extracts the BOM structure and process elements from the extracted elements. Then, based on this information, it generates a process outline, multi-level procedures, and SOP content. Based on the generated procedures, it recommends suitable tooling, equipment, and auxiliary materials.

[0115] This application also discloses a computer-readable storage medium.

[0116] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed as described in the device knowledge management method. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] This application also discloses a computer device.

[0118] Specifically, the computer device includes a memory and a processor, and the memory stores computer programs that can be loaded by the processor and executed using the aforementioned device knowledge management method.

[0119] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A device knowledge management system, characterized in that, include: The system includes a multimodal data access module, a knowledge processing module, a retrieval and question answering module, a fault diagnosis module, a process application module, and a permission management module, with interactive connections established between these modules. The multimodal data access module is used to access and parse files of various formats and extract text, images, and table elements. The knowledge processing module is used to perform structured processing on the extracted elements and construct a knowledge graph. The retrieval and question-answering module is used to perform user intent judgment based on user question input; when question-answering is triggered according to the intent recognition result, multi-path fact retrieval is performed, and the recalled factual evidence is fused to form a closed knowledge set; based on the closed knowledge set, a reasoning chain from fault phenomenon to solution is constructed; the user question, closed knowledge set, and reasoning chain are input as prompt words into the large language model to generate the answer and the basis for answer generation; The fault diagnosis module is used to construct a fault tree based on the knowledge graph and output a fault elimination solution when the intent recognition result determines that fault diagnosis is triggered. The process application module is used to extract the BOM structure and process elements from the extracted elements when the process application is triggered based on the intent recognition result, generate the process outline, multi-level process and SOP content, and recommend tooling, equipment and auxiliary materials based on the process; the permission management module is used to implement access control based on directory and security level. The retrieval question answering module further includes a causal chain generation unit connected to the output of the closure set construction unit; The causal chain generation unit is used to construct a fact-evidence relationship graph based on a closed knowledge set; based on the constructed fact-evidence relationship graph, it performs multiple path expansion, path scoring and filtering, domain rule application, and path-to-text evidence conversion to construct a reasoning chain from fault phenomena to solutions; the domain rule application includes representing domain rules as logical rules or graph patterns, matching candidate paths with domain rules, checking whether they meet the domain rules, correcting paths that do not meet the domain rules, and retaining corrected paths that meet the domain rules. It also includes: a collaborative triggering unit that is connected to the retrieval and question answering module, the fault diagnosis module, and the process application module; The collaborative triggering unit is used to receive intent recognition results, user questions, and generated answers. It uses a large language model to analyze and determine whether to recommend fault diagnosis or process application. If fault diagnosis or process application is recommended, a corresponding fault diagnosis module or process application module is generated and triggered. When the fault diagnosis module is triggered, it executes the construction of a fault tree based on the knowledge graph and outputs a fault elimination solution. When the process application module is triggered, it executes the extraction of BOM structure and process elements from the extracted elements, generates process outline, multi-level processes and SOP content, and recommends tooling, equipment and auxiliary materials based on the processes.

2. The equipment knowledge management system according to claim 1, characterized in that, The retrieval question answering module includes an intent recognition unit, a multi-way fact retrieval unit, and a closure set construction unit; The intent recognition unit is used to determine the user's intent based on the user's question input and to obtain the user's intent. The user intent includes: question answering, fault diagnosis, and process application; a multi-path fact retrieval unit, connected to the output of the intent recognition unit, is used to retrieve entities and relationship paths related to the user's question from the knowledge graph, retrieve document fragments related to the user's question from the document vector library, and retrieve non-textual information related to the user's question from the multimodal feature library, thus completing the retrieval of structured facts and unstructured evidence; the document vector library is obtained by text slicing, word segmentation, and vectorization of the extracted text; the multimodal feature library is generated by feature extraction of the extracted elements; a closure set construction unit, connected to the output of the multi-path fact retrieval unit, is used to align and format the retrieved structured facts and unstructured evidence, perform deduplication and preliminary similarity screening; the processed fact evidence is prioritized based on an inverted index to obtain Top-K fact evidence, and then reordered using a Cross-Encoder; the weights of the recalled fact evidence corresponding to different retrieval sources are dynamically adjusted according to the type of user question and the context of the fact evidence, and the recalled fact evidence is merged to form a closed knowledge set.

3. The equipment knowledge management system according to claim 2, characterized in that, The closure set construction unit is also used to perform conflict detection on the processed factual evidence before prioritizing the processed factual evidence based on the inverted index, and combine the conflict detection results to make the sum of the number of retained non-conflicting factual evidence and the number of retained conflicting factual evidence based on the conflict priority order reach a preset number. Based on the retained factual evidence, priority sorting, reordering, and fusion are further completed; the conflict detection is carried out by querying the corresponding domain factual evidence conflict template in the pre-built domain factual evidence conflict template library, comparing the corresponding domain factual evidence conflict template, and detecting whether there are semantic conflicts, numerical conflicts, or timeliness conflicts; for factual evidence with conflicts, the conflict priority order of the conflicting factual evidence is determined according to the factual evidence priority rules stored in the corresponding domain factual evidence conflict template.

4. The equipment knowledge management system according to claim 1, characterized in that, The closure set construction unit is further configured to set several preset quantities that are adjacent to each other to form several different closed knowledge sets; the causal chain generation unit is further configured to construct factual evidence relationship graphs based on different closed knowledge sets; the retrieval question answering module also includes an answer generation unit connected to the output of the causal chain generation unit; the answer generation unit is further configured to input the user question, each closed knowledge set, and the reasoning chain corresponding to each closed knowledge set as prompt words into the large language model to generate an answer; by performing cluster analysis on multiple answers, the centroid of the cluster with the largest weight obtained from the analysis is output as the final answer and the basis for answer generation.

5. The equipment knowledge management system according to claim 1, characterized in that, The fault diagnosis module is also used to detect in real time whether the knowledge graph is updated during the process of constructing a fault tree based on the knowledge graph and outputting a fault troubleshooting solution; when an update of the knowledge graph is detected, it determines whether there is a relationship between the updated node of the knowledge graph and the fault node in the fault tree. If there is a relationship, it triggers the reconstruction of the fault tree to re-output the fault troubleshooting solution.

6. The equipment knowledge management system according to claim 1, characterized in that, The process application module interfaces with the pre-built equipment digital twin model and is also used to input tooling, equipment and auxiliary materials recommended based on the process as input parameters into the equipment digital twin model, so that simulation operation is performed in the equipment digital twin model to obtain the control range of the key process parameters and provide feedback.

7. The equipment knowledge management system according to claim 1, characterized in that, The knowledge processing module is also used to perform semantic association analysis on the constructed knowledge graph and generate recommendation questions by combining historical user question input data.

8. A device knowledge management method using the system described in any one of claims 1 to 7, characterized in that, include: It can access and parse files of various formats and extract text, images, and table elements; The extracted elements are structured and a knowledge graph is constructed. Based on user question input, the system performs user intent judgment; when question answering is triggered based on intent recognition results, it performs multi-path fact retrieval, and merges the recalled factual evidence to form a closed knowledge set; based on the closed knowledge set, it constructs a reasoning chain from fault phenomena to solutions; it inputs the user question, closed knowledge set, and reasoning chain together as prompt words into a large language model to generate answers and the basis for answer generation; when fault diagnosis is triggered based on intent recognition results, it constructs a fault tree based on the knowledge graph and outputs fault elimination solutions; when process application is triggered based on intent recognition results, it extracts the BOM structure and process elements from the extracted elements, generates process outlines, multi-level processes, and SOP content, and recommends tooling, equipment, and auxiliary materials based on the processes.

Images