CN-121981242-A - Part detection method and system based on knowledge graph

CN121981242ACN 121981242 ACN121981242 ACN 121981242ACN-121981242-A

Abstract

The invention relates to the technical field of knowledge graphs, in particular to a part detection method and a part detection system based on a knowledge graph, comprising a data acquisition module, a data acquisition module and a data acquisition module, wherein the data acquisition module is used for preprocessing multi-mode detection data and detection requirement information to acquire a standardized data set; the method comprises the steps of analyzing a standardized data set by a knowledge graph interaction module, extracting detection elements, carrying out entity matching, relationship reasoning and knowledge completion in a constructed multi-mode detection knowledge graph library to generate a structured knowledge packet, calling an industry template by a report generation module according to the structured knowledge packet to generate a detection report, pushing the detection report to a user terminal by an output and iteration module, analyzing user feedback, and triggering the report generation module to update the detection report or carry out knowledge iteration by the knowledge graph interaction module. The scheme aims at the detection scene of the parts, solves the problems of unordered module interaction, non-uniform reasoning logic, disordered knowledge updating and poor scene suitability in the prior art, and strengthens the scene adaptation capability and the detection efficiency.

Inventors

ZHOU LIANG
HUANG JIANCHUAN
CHEN RUI
FU DAJUN
ZHANG LEI
YAN QINGQUAN
FENG NI
ZHANG SHAOLIN

Assignees

招商局检测车辆技术研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20260107

Claims (10)

1. A part detection system based on a knowledge graph is characterized in that a multi-mode detection knowledge graph library is used for storing structured knowledge to form a knowledge graph, and the knowledge graph is in bidirectional synchronization with a knowledge graph interaction module; the detection equipment cluster is used for acquiring multi-mode detection data; the user terminal is used for acquiring the detection requirement information and user feedback and also used for receiving a detection report; the data acquisition module is used for preprocessing the multi-mode detection data and the detection requirement information to acquire a standardized data set; The knowledge graph interaction module is used for analyzing the standardized data set, extracting detection elements, and carrying out entity matching, relationship reasoning and knowledge completion in the constructed multi-mode detection knowledge graph base to generate a structured knowledge packet; The report generation module is used for calling an industry template in the map according to the structured knowledge packet to generate a detection report; And the output and iteration module is used for pushing the detection report to the user terminal, analyzing the user feedback, triggering the report generation module to update the detection report or the knowledge graph interaction module to iterate the knowledge.
2. The knowledge-based item inspection system of claim 1, wherein the structure of the structured knowledge is an entity-relationship-multimodal property; the multi-mode detection knowledge graph library adopts a three-dimensional modeling mode of layering, multi-mode and fault propagation, and comprises the following steps: Hierarchical knowledge modeling, namely constructing a part hierarchical entity structure according to assembly-sub-parts; The multi-mode attribute expansion is that each entity associates multi-mode characteristic data except text attribute; modeling a fault propagation relationship, namely adding a fault propagation relationship type and describing a conduction path of a part fault; And (3) industry standard association, namely constructing a standard sub-library according to industry classification, and associating a threshold range, a detection method and a qualification judgment rule of the corresponding standard with each part entity.
3. The knowledge-based parts inspection system according to claim 2, wherein the data acquisition module sends the standardized data set to the knowledge-graph interaction module via a data interaction protocol; and the data interaction protocol KIP-001 adopts a JSON format, a sender adds a data check code, a receiver returns ACK after passing the check, and NACK and error reasons are returned after failure.
4. The knowledge-based part detection system according to claim 3, wherein the knowledge-graph interaction module comprises a knowledge preprocessing unit, an entity matching unit, a relationship reasoning unit, a knowledge complement unit and a knowledge packaging unit; The knowledge preprocessing unit is used for analyzing the standardized data set, extracting detection elements, and pre-screening in the constructed knowledge graph library to obtain candidate entities; the entity matching unit is used for carrying out three-level matching in the candidate entities according to the entity matching rule to match the entity; the relationship reasoning unit is used for carrying out relationship reasoning according to the relationship reasoning rule to obtain a relationship reasoning result; The knowledge completion unit is used for carrying out knowledge completion according to the knowledge completion rule; And the knowledge packaging unit is used for packaging results of entity matching, relationship reasoning and knowledge completion in the constructed knowledge graph base to generate a structured knowledge packet.
5. The knowledge-based item inspection system of claim 4, wherein said three-level matching comprises exact matching, semantic similarity matching, and multi-modal feature matching; detecting whether item fields of the standardized data set are consistent with names of candidate entities in the pre-screened candidate entities, if so, judging that the candidate entities are matched, associating the corresponding entities, and if not, carrying out semantic matching, wherein the consistent names are entities of item fields of the standardized data set; Semantic similarity matching, namely calculating the similarity between a candidate entity and a standardized data set through the cosine similarity enhanced by an industry term dictionary, judging that the candidate entity is matched and related to a corresponding entity if the similarity is more than or equal to a first preset value, carrying out multi-modal feature matching if the similarity is less than or equal to a second preset value and triggering manual verification if the similarity is less than or equal to the first preset value; The multi-mode feature matching is carried out, namely the local feature vector of the fault area is extracted from the image data in the standardized data set through CNN, the similarity between the local feature vector of the fault area and the feature vector of the similar fault image in the knowledge graph or the feature matching degree with the time sequence signal is calculated, if the similarity is more than or equal to a third preset value or the matching degree is more than or equal to a fourth preset value, the matching is judged, and corresponding entities are associated; if the detection object is matched with a plurality of entities, conflict processing is carried out, and a unique entity is determined by combining the detection object in the demand_info in the standard data set and the part level.
6. The knowledge-based parts detection system according to claim 4, wherein the relationship inference rule performs a two-stage inference and an anomaly handling; The double-stage reasoning comprises a rule reasoning stage and a GNN reasoning stage; triggering a predefined IF-THEN rule to generate a rule reasoning result; The GNN reasoning stage is used for executing multi-hop traversal of dynamic hop count based on rule reasoning results, wherein the hop count is adaptively determined according to the level depth of the parts and the fault association density; When the confidence level of the inference chain is lower than 85%, the historical case association is automatically supplemented, and the rapid query is performed by adopting characteristic hash and double-dimensional sequencing.
7. The knowledge-based component detection system of claim 4, wherein the knowledge completion rule performs a ranking of knowledge completion source priority as industry standard > history case > expert experience, specifically: Preferentially extracting terms from industry standard entities in the map; If no industry standard entity exists, historical case knowledge with the similarity more than or equal to 90% in 3 years is called to complete knowledge; if the industry standard entity and the historical case knowledge do not exist, triggering an expert experience input interface to complete the knowledge, and marking a source after the knowledge is completed; If the complement knowledge conflicts with the prior knowledge, judging that the latest effectiveness standard is larger than the high-grade standard, and prioritizing the international standard over the industry standard.
8. The knowledge-based parts detection system according to claim 1, wherein the report generating module comprises a template engine, a multi-mode integration unit and an NLG unit; The template engine is used for receiving the structured knowledge packet and calling an industry template in the map according to a data interaction protocol KIP-001; the multi-mode integration unit is used for carrying out association between the image red frame and the text description according to multimodal _link fields in the data interaction protocol; And the NLG unit is used for generating a detection report meeting the specification through NLG according to the called industry template.
9. The knowledge-based component detection system according to claim 1, wherein the output and iteration module comprises a report pushing unit, a feedback processing unit and an iteration control unit; the report pushing unit is used for pushing the generated detection report to the user terminal; The feedback processing unit is used for analyzing the acquired user feedback, generating a knowledge updating packet according to a knowledge updating protocol on the analysis result, transmitting the knowledge updating packet to the knowledge graph interaction module, and triggering the multi-mode detection knowledge graph base to update; And the iteration control unit is used for triggering a knowledge updating flow and coordinating the modules to cooperatively work.
10. A method for detecting parts based on a knowledge graph, characterized in that the system for detecting parts based on a knowledge graph according to any one of claims 1 to 9 is used, comprising the following steps: Acquiring multi-mode detection data and detection demand information, and performing standardization processing to generate a standardized data set; Analyzing the standardized data set, extracting detection elements, and carrying out entity matching, relationship reasoning and knowledge complement in the constructed multi-mode detection knowledge graph library to generate a structured knowledge packet; Generating a detection report according to the structured knowledge packet; And acquiring user feedback information, analyzing the user feedback, and triggering a detection report update or knowledge graph interaction module to iterate knowledge.

Description

Part detection method and system based on knowledge graph Technical Field The invention relates to the technical field of knowledge graphs, in particular to a part detection method and system based on a knowledge graph. Background The development of the vehicle part detection report generation technology always surrounds three core targets of efficiency improvement, content deepening and decision enabling, three key stages are experienced along with the industrial intelligent process, and the technology is continuously upgraded under the driving of technology fusion and scene requirements: 1. manual leading traditional stage (2015 years ago) The industrial detection at this stage takes manual operation and experience judgment as the core, and report generation is completed manually by detection personnel. The inspector needs to synchronously record equipment readings, describe fault phenomena (such as 'cracks exist in blades', 'sensor signal abnormality') and arrange conclusions and suggestions according to a paper manual, and the whole process takes hours or even longer. The method is limited by subjectivity and efficiency bottleneck of manual operation, and the report at the stage has problems of term expression confusion (such as mixed use of 'angle deviation' and 'attitude abnormality'), key information omission (such as part history faults and associated part influences) and the like, and can only serve as 'static record' of a detection result, and cannot provide deep support for maintenance decision. The technical core of this stage is data recording, and no standardized and intelligent technical system is formed yet. 2. Auxiliary phase of automatized tool (2015-2020) With the rise of the information technology of the Internet of things and laboratories, the detection industry starts to introduce automated tools to solve the problem of artificial dependence. Laboratory Informatization Management System (LIMS) becomes mainstream solution, directly links with check out test set through the API interface, realizes the automatic snatch of detection data (such as sensor numerical value, image acquisition result), reduces the manual entry work by a wide margin. Meanwhile, the tool can generate a report framework according to the fixed template, so that the automatic coverage rate of the process is remarkably improved, and the report generation period is effectively shortened. However, the technology still has significant limitations that only data stacking and format standardization can be realized, the fusion processing capability of multi-mode data (images, texts and time sequence signals) is lacked, fault cause reasoning cannot be carried out by combining domain knowledge, reports still stay at a result output layer, and transformation to decision assistance is not realized. 3. Intelligent fusion exploration phase (2020 to date) Under the dual drive of artificial intelligence and intelligent manufacturing development, the technology development enters a new stage of multi-modal fusion and intelligent reasoning. On the one hand, the multi-mode detection technology realizes breakthrough, models such as DeepSeek and the like fuse heterogeneous data such as vision, text and the like through a double-tower and bridging framework, and semantic association is established by means of a cross-mode attention mechanism, so that an AI system can not only understand fault images, but also 'analyze' text logs. Meanwhile, the global intelligent detection market scale is rapidly expanded, the hundred billions of yuan are expected to be broken through in 2025, the reported knowledge depth and the cross-domain adaptability requirements of the fields of automobile electronics, semiconductors and the like are rapidly increased, for example, the defect detection precision of the automobile industry is increased from 0.1mm to 0.05mm, and synchronous correlation fault characteristics, historical data and maintenance standards are required to be reported. However, the prior art still has key gaps that the prior scheme can process multi-mode data and generate texts, but lacks a unified knowledge support carrier, cannot realize end-to-end reasoning of fault characteristics, causes and maintenance schemes, and is difficult to quickly adapt to standard specifications (such as an automobile ISO standard and a mechanical GB standard) of different industries, and the technology falls to the double challenges of knowledge fragmentation and high adaptation cost. Common techniques in the prior art are as follows: 1. Traditional industrial detection report automation tool Such techniques are the basic scheme for report automation in industrial detection scenarios, with the core surrounding the data auto-acquisition and templated fill-in build flow. The device is connected with detection equipment (size measuring instrument, image acquisition camera, sensor and the like) in real time through a hardware interface (such as RS485 and Ethernet