Methods, devices, media and electronic equipment for interpreting knowledge graphs in the machinery manufacturing industry

Abstract

This application provides a method, apparatus, medium, and electronic device for interpreting a knowledge graph in the machinery manufacturing industry, relating to the field of data processing. The method includes: in response to receiving an interpretation request, obtaining request parameters corresponding to the interpretation request; obtaining several key fields; mapping them to entities in a pre-constructed manufacturing knowledge graph to obtain a set of key entities; performing neighborhood retrieval and constraint path retrieval centered on the target quality indicator entity to obtain the shortest path and the highest weight path conforming to a preset pattern; extracting and packaging relevant original evidence corresponding to the set of key entities to obtain an original evidence package; performing similarity retrieval in several historical cases to obtain a historical evidence package; and finally encapsulating the evidence to obtain the interpretation content corresponding to the interpretation request. This application achieves an organic combination of anomaly attribution and process mechanism inference, significantly reducing engineers' reliance on experience and the cost of secondary interpretation.

Description

Technical Field

[0001] This application belongs to the field of data processing, and in particular relates to methods, devices, media and electronic equipment for interpreting knowledge graphs in the machinery manufacturing industry. Background Technology

[0002] Currently, in the process of digital and intelligent transformation, machinery manufacturing enterprises have widely deployed various business systems such as MES, SCADA, QMS, and ERP, generating a large amount of structured and semi-structured data covering process flow, equipment status, material batches, and quality inspection results. Consequently, machine learning and automated modeling platforms (AutoML) are also being gradually implemented within these enterprises, providing model support for scenarios such as production line yield prediction, fault warning, and quality grading. However, the core capabilities of existing general-purpose AutoML platforms are still mainly concentrated in the two stages of "model training and result display," lacking a deep understanding of the professional semantics and process mechanisms behind the model output.

[0003] On the one hand, existing platforms generally cannot recognize a large number of specialized terms and industry jargon in the field of mechanical manufacturing, and cannot establish stable semantic correspondences between field names, indicator names, and actual process steps, equipment components, or material properties. This makes it difficult to reuse and align model results across different systems and workshops. On the other hand, when faced with phenomena such as "a decline in a certain quality indicator" or "fluctuations in the yield of a certain batch," existing platforms can usually only provide "descriptive results" such as indicator trends and feature importance rankings. They are unable to further answer key questions such as "which process steps or equipment conditions caused this change" and "what physical mechanism is behind it." Engineers need to rely on experience to reinterpret the results, which is inefficient and highly subjective.

[0004] With the development of Large Language Model (LLM) technology, its capabilities in natural language understanding, complex reasoning, and multi-source information fusion have made it possible to provide business-level and mechanism-level explanations for manufacturing data analysis results. However, if interpretations rely solely on general large models for "free-form" explanations without considering the manufacturing industry's own knowledge system and process mechanisms, problems such as discrepancies between explanations and actual processes, inconsistent interpretation styles across different batches, and difficulty in tracing the basis for explanations can easily arise. This fails to meet the stringent requirements of the manufacturing industry for interpretability, verifiability, and feasibility. Therefore, there is an urgent need for a knowledge foundation system that can deeply integrate the knowledge graph of the machinery manufacturing industry with the interpretive capabilities of large models. Based on a full understanding of business fields and industry semantics, this system should provide structured and interconnected business explanations for analysis results and anomalies, and support the attribution of anomaly causes and inference of process mechanisms. This would achieve the goal of "making the model truly understand the process and making the explanation truly applicable" in the manufacturing context. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, apparatus, medium, and electronic device for interpreting knowledge graphs in the machinery manufacturing industry, which at least partially solves the problems existing in the prior art.

[0006] In a first aspect of this application, a method for interpreting knowledge graphs in the machinery manufacturing industry is provided, the method comprising the following steps: In response to receiving an explanation request, the request parameters corresponding to the explanation request are obtained; When the explanation request comes from the model output, the Top-K key fields are filtered in the feature field pool corresponding to the model to obtain a number of key fields; Several key fields are mapped to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities; Centered on the target quality indicator entity among the key entities, neighborhood search and constraint path search are performed based on the pre-constructed manufacturing knowledge graph to obtain the shortest path and the highest weight path that conforms to the preset pattern. Extract relevant original evidence corresponding to key entity sets from the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph and package them to obtain the original evidence package; Similar searches are conducted in several historical cases to obtain a package of historical evidence; The key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package are encapsulated to obtain the explanation content corresponding to the explanation request.

[0007] In a second aspect of this application, a knowledge graph interpretation device for the machinery manufacturing industry is provided, the device comprising: The acquisition unit is used to acquire the request parameters corresponding to the interpretation request in response to receiving the interpretation request; The filtering unit is used to filter the Top-K key fields in the feature field pool corresponding to the model when the explanation request comes from the model output, so as to obtain a number of key fields. The mapping unit is used to map several key fields to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities. The retrieval unit is used to perform neighborhood retrieval and constraint path retrieval based on the pre-built manufacturing knowledge graph, with the target quality indicator entity among the key entities as the center, in order to obtain the shortest path and the highest weight path that conforms to the preset pattern. The original evidence acquisition unit is used to extract relevant original evidence corresponding to key entity sets from the multimodal original manufacturing data corresponding to the pre-built manufacturing knowledge graph and package it to obtain the original evidence package. The historical evidence acquisition unit is used to perform similar searches in several historical cases to obtain a historical evidence package; The explanation generation unit is used to encapsulate the key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package to obtain the explanation content corresponding to the explanation request.

[0008] In a third aspect of this application, a non-transitory computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored in the storage medium, and the at least one instruction or at least one program is loaded and executed by a processor to implement the aforementioned mechanical manufacturing knowledge graph interpretation method.

[0009] In a fourth aspect of this application, an electronic device is provided, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0010] This application has at least the following beneficial effects: The knowledge graph interpretation method for the machinery manufacturing industry provided in this application first responds to the interpretation request and obtains the corresponding request parameters, ensuring the relevance and clarity of the interpretation task and avoiding inefficiency caused by ambiguity in the interpretation scope; when the interpretation request comes from the model output, the Top-K key fields are selected from the model feature field pool based on the core influencing factors, which not only ensures the focus of key information but also avoids interference from redundant fields, thus improving interpretation efficiency; the key fields are mapped to pre-constructed manufacturing knowledge graph entities to form a set of key entities. By leveraging the standardized semantic mapping capability of the knowledge graph for specialized terms and industry terms in the machinery manufacturing field, a stable association between fields and real process links, equipment components, and quality indicators is established, effectively solving the problems of semantic alignment deficiency and difficulty in reusing model results across systems in existing platforms; with the target quality indicator entity as the center, the shortest path and the highest weight path that conform to the preset process mechanism pattern are obtained through neighborhood retrieval and constraint path retrieval, breaking the limitation that existing platforms can only output descriptive data. The limitations of the results were addressed by providing a structured framework for revealing the technological mechanisms behind quality anomalies or changes in indicators. Key entities were extracted from multimodal raw manufacturing data and packaged into a raw evidence package, integrating multi-source evidence such as structured time-series data, text documents, and images. This provided solid support from actual production data for the interpretation, ensuring the verifiability of the results. Historical evidence packages were obtained through similarity searches in historical cases, leveraging mature causal chains and conclusions to further enhance the credibility and reference value of the interpretation. Finally, the key entity set, technological mechanism path, multi-source raw evidence package, and historical evidence package were encapsulated as the interpretation content. This not only constrained the interpretation boundaries of the large model through a knowledge graph, avoiding the problem of arbitrary interpretations being disconnected from actual processes, but also unified the interpretation style through a structured and linked output format. This achieved an organic combination of anomaly attribution and technological mechanism inference, significantly reducing engineers' reliance on experience and the cost of secondary interpretation, truly achieving the goal of enabling the model to understand the process and making the interpretation practical and usable in manufacturing scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of a knowledge graph interpretation method for the machinery manufacturing industry provided in this application embodiment; Figure 2 A schematic diagram of the overall architecture of the multimodal knowledge graph-enhanced large model interpretation system for the machinery manufacturing industry provided in this application embodiment; Figure 3 The structural block diagram of the knowledge graph interpretation device for the machinery manufacturing industry provided in this application embodiment. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] It should be noted that the following description covers various aspects of embodiments within the scope of the appended claims. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0016] Please refer to Figure 1 As shown, an embodiment of this application provides a method for interpreting a knowledge graph in the machinery manufacturing industry. The method includes the following steps: S100, in response to receiving an explanation request, the request parameters corresponding to the explanation request are obtained.

[0017] Specifically, in response to receiving an explanation request from the upper-level modeling or analysis platform (e.g., anomalies in quality indicators, model prediction deviations, etc.), the request parameters corresponding to the explanation request are obtained, including the target quality indicator, analysis type, list of key fields, field importance or anomaly score, time range, and equipment / production line / material identifier, and standardized to form an ExplanationRequest object.

[0018] S200: When the explanation request comes from the model output, the Top-K key fields are filtered in the feature field pool corresponding to the model to obtain several key fields.

[0019] Specifically, when an explanation request comes from the model output, the Top-K key fields are selected from the feature field pool corresponding to the model based on feature importance, SHAP value, etc.

[0020] S300 maps several key fields to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities.

[0021] Specifically, the construction process of the pre-built manufacturing knowledge graph includes: S310 collects multimodal raw manufacturing data.

[0022] Here, the multi-source data from manufacturing enterprises, such as MES, QMS, ERP, SCADA / PLC, equipment logs, and industrial cameras, are first accessed by the multi-source manufacturing data access module (1). Specifically, this includes: structured access: extracting structured data such as orders, process parameters, and quality records from relational databases and Excel / CSV files through JDBC, ODBC, etc., and realizing historical full import and incremental synchronization through CDC technology; unstructured access: accessing documents such as process specifications, inspection specifications, and equipment manuals, as well as unstructured data such as industrial images and videos, performing text parsing and paragraph segmentation on PDF and Word documents, and processing JPEG and PNG documents. Image decoding and basic preprocessing are performed, preserving metadata such as time and device information. Real-time streaming access: SCADA / PLC sensor data streams, device alarm streams, and log event streams are integrated. A time-window-based micro-batch processing algorithm (fixed window, sliding window) is used for segmentation, disordered rearrangement, and deduplication to achieve low-latency, high-throughput acquisition. Metadata acquisition: Table structure, field names, Chinese field descriptions, units, enumeration code tables, primary keys, time fields, etc., are automatically extracted to form a unified metadata catalog, providing basic information for subsequent field semantic parsing, graph construction, and data lineage. After step S310, a multimodal raw dataset with data source identifiers and metadata descriptions is obtained, providing input for subsequent preprocessing and knowledge modeling.

[0023] S320, the multimodal raw manufacturing data is preprocessed and standardized to obtain a characteristic manufacturing dataset; wherein the characteristic dataset has a statistical summary and a unified multimodal semantic representation.

[0024] Here, multi-source data enters the data preprocessing and standardization process, specifically including: data cleaning: retaining one record for primary key conflicts or duplicate records based on timestamps or priority rules; identifying outliers using statistical methods (such as the 3σ principle) and Isolation Forest, and marking or correcting outliers; imputing missing values ​​using mean / median / linear interpolation or models based on similar samples; performing denoising, unified encoding, and basic word segmentation on the text; format conversion and standardization: unifying field naming based on a pre-defined data dictionary (e.g., unifying "Temp1" and "temperature 1" to "preheating zone outlet temperature"), performing unit conversion and normalization on physical quantities such as temperature and pressure, converting data from different source tables into a unified columnar storage or key-value pair structure, and adding data source and time labels. After the above processing, a "cleaned dataset" that is aligned in semantics, units, and format is obtained. Then, statistical summaries and multimodal features are generated. Based on cleaning and standardization, data preprocessing and standardization further generate statistical summaries and multimodal features, specifically including: Statistical summaries and distribution drift processing: calculating statistics such as mean, variance, quantiles, skewness, and kurtosis by field and time window, and calculating distribution drift indices such as KS test statistic and PSI to form quantitative evidence that can be directly used for interpretation; Multimodal feature extraction and cross-modal alignment: jointly encoding text and image data: where BERT or ERNIE is used on the text side. Pre-trained language models are used to encode context-dependent text feature vectors. On the image side, object detection models (such as the YOLO series) and image caption generation models (such as BLIP) are combined to extract visual semantic feature vectors. Based on this, a dual-tower cross-modal coding network based on contrastive learning is constructed. One tower processes text features, and the other processes image features. Each tower consists of several fully connected layers and normalization layers, outputting a unified-dimensional embedding vector. During training, a contrastive learning loss is used to maximize the similarity of positive sample text-image pairs and minimize the similarity of negative sample pairs within the same batch, thereby achieving cross-modal semantic alignment within a unified semantic space and obtaining multimodal feature vectors that can be used for subsequent reasoning and interpretation. Finally, a feature-rich dataset with statistical summaries and unified multimodal semantic representations is obtained based on the cleaned data.

[0025] S330 constructs a manufacturing knowledge graph based on a characteristic manufacturing dataset.

[0026] Here, the preprocessed and characterized data, along with metadata, flows into the manufacturing knowledge graph construction and storage service unit to complete knowledge modeling and persistence. This includes: Ontology and Lexicon Management: defining ontology concepts such as process segments, process parameters, equipment components, material properties, quality indicators, fault types, and mechanism states, and their hierarchical relationships; maintaining domain dictionaries and alias tables to form a unified industry semantic benchmark; Entity Extraction: extracting entities such as process parameters, quality indicators, equipment, materials, and defects from process procedures, specification documents, metadata descriptions, and log text, using BERT / ERNIE+CRF. The text is labeled with entities using a pointer network-based named entity recognition model. The multimodal alignment results obtained above are used to map the identified defect regions and component names in the image to text entities for cross-modal completion. Relation extraction: Relation candidates are constructed for entity pairs. Relationships such as "belong to," "be located at," "cause," and "measured by" are identified using semantic patterns (e.g., "due to…", "affects…indicators") and supervised learning models (multilayer perceptron, BiLSTM+attention, etc.). A relation classification model is trained using cross-entropy loss to output relation types and confidence levels. Graph storage and query processing: Attribute graph models are used to store nodes and edges. Nodes record attributes such as entity type, alias, unit, and embedding vector. Edges record relation types, confidence levels, effective time, and evidence sources, supporting point queries, edge queries, K-hop neighborhood queries, and constraint-based path searches. Graph embedding and indexing processing: Node2vec and GraphSAGE are used. The graph embedding algorithm learns embedding vectors for nodes in the graph and constructs an Approximate Nearest Neighbor (ANN) index for subsequent entity similarity retrieval and similar case retrieval. In summary, the original business data and document knowledge are transformed into a structured, searchable, and reasonable manufacturing knowledge graph.

[0027] This embodiment performs semantic parsing and cross-system schema alignment on several key fields to obtain a set of key entities. The semantic parsing and cross-system schema alignment include: field description construction, field semantic embedding generation, candidate entity recall and multi-factor scoring, constraint verification and confirmation, and mapping version management.

[0028] Furthermore, the field description construction process constructs a FieldDescriptor for each field, combining information such as table name, field name, Chinese description, unit, enumeration value, sample statistical summary, system and process section into field description text; the field semantic embedding generation process uses a sentence vector model (such as BERT-based Sentence Embedding) to encode the field description text into a field vector, and encodes the unit, value range, system type, etc. into auxiliary features, which are then concatenated with the text vector and mapped to a unified semantic space; the candidate entity recall and multi-factor scoring process uses the field vector as the query vector, retrieves candidate entities in the graph embedding and indexing module (23), and supplements the recall using an alias dictionary and string similarity, and scores the candidate entities comprehensively; the constraint verification and confirmation process removes candidates with inconsistent units or obviously inconsistent value ranges by rules, automatically confirms high-confidence mappings, and triggers manual review for gray-scale interval mappings; the mapping version management process stores the field-to-entity mapping relationship in a versioned manner, records change information, and references the specific mapping version number when interpreting, ensuring that the results are traceable and replayable.

[0029] S400 takes the target quality indicator entity among the key entities as the center, and performs neighborhood search and constraint path search based on the pre-built manufacturing knowledge graph to obtain the shortest path and the highest weight path that conforms to the preset pattern.

[0030] Specifically, Top-K key fields are selected based on feature importance, SHAP value, etc., and these key fields are mapped to knowledge graph entities to obtain a set of key entities. The context subgraph retrieval module (31) is used to perform K-hop neighborhood retrieval and constraint path search with the target indicator entity as the center, retaining only the path that conforms to the process parameters to mechanism state to quality indicator pattern; it enumerates the shortest path and the highest weight path between key entities and target indicator entities to provide a skeleton for subsequent interpretation.

[0031] S500 extracts relevant original evidence corresponding to key entity sets from the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph and packages it to obtain the original evidence package.

[0032] Specifically, step S500 includes: S510 retrieves procedures, standards, and historical reports from the text and document-side data of the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph to obtain original evidence of document fragments related to the key entity set.

[0033] S520: For the structured and time-series data of the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph, the statistical summary and drift index original evidence within the time window corresponding to the interpretation request are read from the preprocessing module. The drift index is used to measure whether there has been a significant change in the data distribution.

[0034] S530 retrieves image defect information and visual features from the image-side data of the multimodal raw manufacturing data corresponding to the pre-constructed manufacturing knowledge graph, for the same time period, the same equipment, or the same workstation, in order to obtain original image evidence.

[0035] S540, package the original evidence of document fragments, original evidence of statistical summaries and drift indicators, and original evidence of images related to the set of key entities into an original evidence package.

[0036] Here, for text and document-side data, BM25 or similar probabilistic ranking functions are used to retrieve procedures, standards, and historical reports to obtain document fragments related to target indicators and key entities; for structured / time-series data, statistical summaries and drift indicators within the corresponding time window are read from the preprocessing module; for image-side data, image defect information and visual features are retrieved from the same time period and the same equipment / workstation; the above evidence is packaged into an EvidencePackage, including evidence type, source, time range, related entities, and brief summary.

[0037] S600, perform a similar search in several historical cases to obtain a package of historical evidence.

[0038] Specifically, step S600 also includes: S610, In several historical cases, similarity retrieval is performed based on the target quality indicator entity, the root cause entity combination, and the graph embedding vector to obtain several similar historical cases; wherein, the root cause entity combination is the set of root cause entities in the key entity set that have a causal relationship with the target quality indicator entity, and the graph embedding vector is the vector obtained by graph embedding calculation on the context subgraph composed of the target indicator entity, the root cause entity, and their relationship.

[0039] S620, encapsulate and package the key causal chain and conclusion summary of each similar historical case in several similar historical cases to obtain a historical evidence package.

[0040] Here, in the historical interpretation results and case library, similarity retrieval is performed by target indicators, root cause entity combinations and graph embedding vectors to select several cases with the highest similarity, and their key causal chains and conclusion summaries are extracted as reference case evidence.

[0041] S700 encapsulates the key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package to obtain the explanation content corresponding to the explanation request.

[0042] In one exemplary embodiment of this application, the method further includes, after step S700: S800, under the constraints of a pre-built manufacturing knowledge graph, original evidence package, and historical evidence package, generates a structured explanation and outputs the explanation content corresponding to the explanation request.

[0043] Specifically, the layered prompt template library handles the following: It pre-configures multiple layered prompt templates, including at least: a domain knowledge layer: injecting basic physical mechanisms, process rules, and quality control logic; an evidence constraint layer: listing the evidence list in the current ContextPackage, requiring the explanation to reference several evidence IDs; and a structural constraint layer: specifying the output JSON Schema (such as fields like summary, root_causes, causal_chains, and recommendations) and restricting the output format. The explanation input construction process involves filling the ExplanationRequest and ContextPackage with corresponding placeholders according to the template, forming a structured prompt. The large model call and routing process selects appropriate large model instances (different sizes / dedicated fine-tuning models, etc.) based on task type and complexity, controlling parameters such as temperature and maximum generation length to ensure output stability and controllability. It automatically retryes or rolls back to a simplified template in case of timeout or parsing failure. The multi-stage reasoning process, through chained or multi-round prompts, breaks down the reasoning process into stages such as "phenomenon induction to mechanism hypothesis to causal chain fitting to parameter adjustment suggestions." Self-checking instructions are added between stages, requiring the model to verify the consistency of conclusions before and after each stage, reducing the risk of "illusions." The output self-repair and structured constraint processing is used to perform JSON schema validation, field integrity validation, and evidence citation validation on the large model output. If a missing field or a non-existent evidence ID is found, an error correction prompt is constructed to guide the model to perform one or more self-repairs until the structured requirements are met or the retry limit is reached.

[0044] In one exemplary embodiment of this application, the method further includes, after step S800: The S900 performs closed-loop updates to field mapping, knowledge graph edge weights, and inference strategies based on user feedback on structured interpretations.

[0045] Specifically, the feedback collection and processing involves providing "confirm / reject / modify" entry points for each root cause, causal chain, and suggestion on the front-end interface. User actions are recorded as FeedbackRecords, including: object type (link / rule / mapping), operation type, expected result after correction, user role, and timestamp. The mapping and terminology update processing involves updating the field-entity mapping table and alias dictionary based on user modifications to field mappings, forming a mapping difference package (MappingDelta), and updating the mapping version. The graph incremental update processing involves converting user-added or modified relationships and mechanism nodes into graph update patches (GraphPatch), writing them to the graph database in a transactional manner, and performing weight decay or deletion on edges with low confidence or those frequently rejected. The rule weight adjustment and forgetting mechanism processing treats rules as "actions" in a policy, constructs a reward function based on user feedback, and updates rule weights using reinforcement learning methods such as policy gradients. Rules that have not been triggered for a long time or have received multiple negative feedbacks are subject to weight decay until they reach a deletion threshold and are removed from the rule base, thus achieving rule forgetting. The prompt template and model evaluation process are used to continuously monitor indicators such as structured output pass rate, evidence hit rate, and interpretation consistency. A / B experiments are conducted on different versions of the prompt template / model, and the version with better performance is given higher weight or set as the default.

[0046] like Figure 2 As shown, the knowledge graph interpretation method for the machinery manufacturing industry provided in this application relies on a multimodal knowledge graph-enhanced large-model interpretation system for machinery manufacturing data. This system includes a multi-source manufacturing data access module, a data preprocessing and standardization module, a manufacturing knowledge graph construction and storage service module, a field semantic parsing and alignment module, an interpretation request management and context retrieval module, a large-model interpretation engine and reasoning orchestration module, an anomaly cause attribution and causal chain generation module, an interpretation result structuring and multi-role display module, an incremental learning optimization and knowledge evolution module, and a system support and security governance module.

[0047] The system comprises the following modules: a multi-source manufacturing data access module connected to a data preprocessing and standardization module; a data preprocessing and standardization module connected to a manufacturing knowledge graph construction and storage service module and a field semantic parsing and alignment module; a manufacturing knowledge graph construction and storage service module connected to a field semantic parsing and alignment module and an explanation request management and context retrieval module; an explanation request management and context retrieval module connected to a large model explanation engine and inference orchestration module; and a large model explanation engine and inference orchestration module connected to anomaly attribution and causal chain generation module and an explanation result structuring and multi-role display module. An incremental learning optimization and knowledge evolution module is connected to the manufacturing knowledge graph construction and storage service module, the field semantic parsing and alignment module, and the large model explanation engine and inference orchestration module. A system support and security governance module is connected to all the above modules, providing unified access control, log auditing, version management, and fault tolerance recovery capabilities.

[0048] Among them, such as Figure 2 As shown, the multi-source manufacturing data access module is used to access multi-source data from within the machinery manufacturing enterprise, such as MES, QMS, ERP, SCADA, PLC, equipment logs, and industrial cameras, forming a multimodal raw dataset. This includes a structured access module, an unstructured access module, a real-time streaming access module, and a metadata acquisition module. The structured access module extracts structured data such as orders, process parameters, and quality records from relational databases via JDBC, ODBC, etc.; it uses Change Data Capture (CDC) technology to monitor database logs (such as Binlog) for full historical data import and incremental synchronization of new / updated records; and it supports batch extraction by timestamp, primary key range, version number, and other dimensions. The unstructured access module is used to access documents such as process specifications, inspection standards, and equipment manuals, as well as industrial images and videos; it uses a document parsing engine to extract text and segment paragraphs from PDFs and Word documents; and it decodes and performs basic preprocessing (scaling, normalization, etc.) on JPEG and PNG images, retaining metadata such as shooting time and equipment number. The real-time streaming access module is used to interface with SCADA / PLC sensor data streams, device alarm streams, and log event streams. It employs time-window-based micro-batch processing algorithms, such as fixed windows and sliding windows, to segment, reorder, and deduplicate the data streams. It utilizes a window aggregation approach similar to streaming processing frameworks to achieve low-latency, high-throughput data access. The metadata acquisition module automatically extracts table structures, field names, Chinese field descriptions, units, enumeration code tables, primary keys, time fields, etc., to form a metadata catalog. This provides foundational information for subsequent field semantic parsing, graph construction, and data lineage determination.

[0049] The data preprocessing and standardization module is used to clean, transform, and standardize the incoming data, forming statistical summaries and multimodal features that can be used for interpretation and evidence citation. This includes a data cleaning module, a format conversion and standardization module, a statistical summarization and distribution drift module, and a multimodal feature extraction and cross-modal alignment module. The data cleaning module retains one record for primary key conflicts or duplicate records based on timestamps or priority rules. For numerical fields, it identifies outliers using statistical methods (such as the 3σ principle) and algorithms like Isolation Forest, and outliers can be marked or replaced with missing values. Missing values ​​are filled using mean, median, linear interpolation, or regression / imputation models based on similar samples. The text is denoised (removing HTML tags and anomalous symbols), uniformly encoded, and simply segmented. The format conversion and standardization module unifies field naming based on a preset data dictionary (e.g., unifying "Temp1" and "Temp 1" into "preheating zone outlet temperature"); it performs unit conversion and normalization for physical quantities such as temperature and pressure; it converts table data from different sources into a unified columnar storage or key-value pair structure, and adds tags such as data source and time range. The statistical summary and distribution drift module calculates statistics such as mean, variance, quantiles, skewness, and kurtosis by field and time window for subsequent interpretation; it calculates distribution difference indicators between different periods, such as the KS test statistic and PSI (Population Stability Index), to monitor changes in operating conditions or data distribution. The multimodal feature extraction and cross-modal alignment module performs feature extraction and semantic alignment on text and image data. For text, it uses pre-trained language models such as BERT or ERNIE to encode the text, obtaining context-dependent text feature vectors. For image-side processing, a combination of object detection models (such as the YOLO series) and image description generation models (such as BLIP) is used to extract visual semantic feature vectors. Building upon this, a dual-tower cross-modal coding network based on contrastive learning is constructed. One tower processes text features, while the other processes image features. Each tower consists of several fully connected layers and normalization layers, outputting a uniform-dimensional embedding vector. , During training, a contrastive learning loss is used to maximize the similarity of positive sample image-text pairs and minimize the similarity of negative sample pairs within the same batch, thereby achieving cross-modal semantic alignment in a unified semantic space and obtaining multimodal feature vectors that can be used for subsequent reasoning and interpretation.

[0050] The manufacturing knowledge graph construction and storage service module is used to construct a knowledge graph covering process flow, equipment, materials, quality indicators, quality rules, and industry terminology, and provides graph query and subgraph retrieval services. It includes an ontology and vocabulary management module, an entity extraction module, a relation extraction module, a graph storage and query module, and a graph embedding and indexing module. The ontology and vocabulary management module defines the ontology of the mechanical manufacturing field, including concepts such as process segments, process parameters, equipment components, material properties, quality indicators, fault types, and mechanism states, and their hierarchical relationships; it maintains the domain dictionary and alias table for subsequent entity recognition and field alignment. The entity extraction module extracts entities from process specifications, standard documents, metadata descriptions, and log text; it uses BERT / ERNIE+CRF or pointer network named entity recognition models to label text sequences as entity boundaries and categories; and it utilizes the aforementioned multimodal alignment results to map defect areas and component names identified in images to text entities for cross-modal completion. The relation extraction module constructs relation candidates for entity pairs, identifying relations such as "belongs to," "is located at," "causes," and "affects the index of..." through semantic patterns ("due to...," "affects the index of...", etc.) and supervised learning models (multilayer perceptron, BiLSTM+attention, etc.). It trains a relation classification model using cross-entropy loss, outputting relation type and confidence level. The graph storage and query module uses a property graph model to store nodes and edges. Nodes record attributes such as entity type, alias, unit, and embedding vector, while edges record relation type, confidence level, and effective time. It supports point query, edge query, K-hop neighborhood query, and constraint-based path search, providing subgraph retrieval capabilities for the interpretation stage. The graph embedding and indexing module uses biased random walk-based graph embedding algorithms such as node2vec to randomly sample nodes in the graph, then learns node vectors using the word2vec approach to capture local and global structural information. It uses graph neural networks such as GraphSAGE to sample and aggregate local neighborhoods of nodes, learning inductive node representations to facilitate online embedding of new nodes or subgraphs. The node embedding vectors are stored in an Approximate Nearest Neighbor (ANN) index structure (such as HNSW) for subsequent entity similarity retrieval and similar case retrieval.

[0051] The field semantic parsing and alignment module maps multi-source system fields to standard industry entities in the knowledge graph, achieving schema semantic alignment and forming a unified semantic coordinate system for the interpretation task. It includes a field description construction module, a field semantic embedding generation module, a candidate entity recall and multi-factor scoring module, a constraint verification and confirmation module, and a mapping version management module. The field description construction module constructs a FieldDescriptor for each field, including: table name, field name, Chinese description, unit, enumerated value, sample statistical summary (mean, variance, distribution), system and process section, etc. The field semantic embedding generation module uses a sentence vector model (such as BERT-based SentenceEmbedding) to encode the field description into a vector; it encodes the unit, value range, system type, etc., into auxiliary features, concatenates them with the text vector, and maps them to the unified semantic space through a fully connected layer. The candidate entity recall and multi-factor scoring module uses the field embedding as the query vector to retrieve several candidate entities in the graph embedding index; it also performs candidate recall based on an alias dictionary and string similarity (edit distance, Jaccard); and calculates a comprehensive score for each candidate entity. ,in For cosine similarity, UnitMatch represents unit consistency, StageMatch represents process segment consistency, RangeMatch represents value range consistency, and AliasHit represents alias hit. Top-N candidates are ranked by score. The constraint verification and confirmation module removes candidates with high scores but inconsistent units or significantly mismatched value ranges using rules. Mappings with confidence scores above a threshold are automatically confirmed, while those below the threshold are submitted for manual review and confirmed by domain experts before taking effect. The mapping version management module is used to version-store field-to-entity mapping relationships, recording change time, changer, and differences. The specific mapping version number is referenced during interpretation to ensure traceability and replayability of results.

[0052] The explanation request management and context retrieval module receives explanation requests from the upper-level modeling / analysis platform, converts data results into a format that can be processed by the knowledge graph and retrieval system, and organizes context subgraphs and evidence packages. This includes an explanation request parsing module, a key factor filtering module, a context subgraph retrieval module, an evidence retrieval and fusion module, a similar case retrieval module, and a context packaging module. The explanation request parsing module receives explanation request parameters, including target quality indicators, analysis type (anomaly explanation / model explanation), a list of key fields, field importance or anomaly score, time range, equipment / production line / material identifiers, etc., and standardizes them into ExplanationRequest objects. The key factor filtering module, when the explanation request originates from model output, selects Top-K key fields based on feature importance, SHAP values, etc., and maps these key fields to knowledge graph entities to obtain a set of key entities. The context subgraph retrieval module performs K-hop neighborhood retrieval and constraint path search centered on the target indicator entity, retaining only paths conforming to the pattern "process parameters → mechanism state → quality indicator". It enumerates the shortest and highest-weighted paths between key entities and the target indicator entity, providing a framework for subsequent interpretation. The evidence retrieval and fusion module uses BM25 or similar probability ranking functions to search procedures, standards, and historical reports for text and document-side data, obtaining document fragments related to the target indicator and key entities. For structured / time-series data, it reads statistical summaries and drift indicators within the corresponding time window from the preprocessing module. For image-side data, it retrieves image defect information and visual features from the same time period and the same equipment / workstation. The above evidence is packaged into an EvidencePackage, including evidence type, source, time range, associated entities, and a brief summary. The similar case retrieval module performs similarity searches in historical interpretation results and the case library using target indicators, root cause entity combinations, and graph embedding vectors, selecting several cases with the highest similarity and extracting their key causal chains and conclusion summaries as "reference case evidence". The context packaging module encapsulates the context subgraph, key entities, statistical summaries, multimodal evidence packages, similar case summaries, and output structure constraints into a ContextPackage, which is then provided to the large model interpretation engine.

[0053] The large model explanation engine and inference orchestration module are used to generate business-level and mechanism-level explanation outputs under the constraints of knowledge graphs and evidence. Under the constraints of ContextPackage, it calls the large language model to generate structured explanations and ensures the output structure is stable and parsable. This includes a layered prompt template library module, an explanation input construction module, a large model invocation and routing module, a multi-stage inference module, and an output self-repair and structured constraint module. The layered prompt template library module pre-sets multiple layered prompt templates, including at least: a domain knowledge layer: injecting basic physical mechanisms, process rules, and quality control logic; an evidence constraint layer: listing the evidence list in the current ContextPackage, requiring the explanation to reference several evidence IDs; and a structure constraint layer: specifying the output JSON schema (such as fields like summary, root_causes, causal_chains, and recommendations) and restricting the output format. The explanation input construction module fills the ExplanationRequest and ContextPackage into the corresponding placeholders according to the template, forming a structured prompt (StructuredPrompt). The large model invocation and routing module selects appropriate large model instances (different sizes / dedicated fine-tuning models, etc.) based on task type and complexity, controlling parameters such as temperature and maximum generation length to ensure output stability and controllability. It automatically retryes or rolls back to a simplified template in case of timeout or parsing failure. The multi-stage inference module breaks down the inference process into stages such as "phenomenon induction → mechanism hypothesis → causal chain fitting → parameter adjustment suggestions" through chained or multi-round prompts. Self-checking instructions are added between stages, requiring the model to verify the consistency of conclusions before and after each stage, reducing the risk of "illusions." The output self-repair and structured constraint module performs JSON schema validation, field integrity validation, and evidence citation validation on the large model output. If a missing field or a non-existent evidence ID is found, an error correction prompt is constructed, guiding the model to perform one or more self-repairs until the structured requirements are met or the retry limit is reached.

[0054] The anomaly attribution and causal chain generation module transforms the natural language output of the large model into standardized causal chains and root cause lists, and performs logical verification and confidence quantification evaluation on the links. This includes an entity linking and standardization module, a causal chain extraction module, a logical verification and conflict resolution module, a confidence evaluation module, and an explanation subgraph generation and write-back module. The entity linking and standardization module matches entity names in the explanation text to knowledge graph node IDs using a field mapping table, alias table, and embedding similarity. It disambiguates ambiguities of entities with the same name on different process sections or equipment by combining contextual time ranges and equipment numbers. The causal chain extraction module extracts the "cause node—relationship—result node" triple sequence from the structured fields output by the large model, maps it to a directed path in the graph, and completes the missing intermediate mechanism nodes in the path using the graph structure. The logic verification and conflict resolution module checks whether the causal chain conforms to a predefined path pattern (e.g., "process parameters → mechanism state → quality indicators"). Links that conflict with the domain common sense base or existing high-confidence rules are marked and downweighted. When multiple links conflict, strategies such as rule priority and link support are used to resolve the conflict. The confidence assessment module integrates graph support (average weight of edge edges on the path), data evidence support (statistical deviation significance, multimodal evidence strength), historical case support, and interpretability consistency into a link confidence score: Conf = α·GraphSupport + β·DataEvidence + γ·CaseSupport + δ·Consistency, where GraphSupport is the graph path edge weight / frequency support, DataEvidence is the deviation and time consistency evidence, CaseSupport is the historical case support, and Consistency is the degree of logical verification pass rate. Root causes are sorted, and a root cause list aggregated by process segment, equipment, material, and other dimensions is output. The explanatory subgraph generation and write-back module converts the verified causal chain into an explanatory subgraph, associates it with the nodes and edges in the original graph, and increases the edge weight or promotes the link with high confidence and multiple confirmations to rule candidates for use by the subsequent rule generation and optimization module.

[0055] The structured explanation result and multi-role display module outputs a unified structured explanation result and generates differentiated display views based on different user roles. This includes a structured result output module, a multi-role view adaptation module, a visualization module, and an interface and report output module. The structured result output module outputs results in a unified `ExplanationResult` structure, including fields such as `summary`, `root_causes`, `causal_chains`, `evidence`, `recommendations`, and `trace`. The `trace` field contains information such as data version, graph version, field mapping version, and rule version. The multi-role view adaptation module performs granular trimming and language style adaptation on the structured results. For process engineers, it highlights process segments, parameter offset ranges, and process windows; for quality engineers, it highlights quality indicator trends, fluctuation periods, historical comparisons, and risk assessments; and for management personnel, it outputs brief explanations, impact scope, and priority handling suggestions. The visualization module displays causal chains in a node-side graph format, allowing users to click on nodes to view corresponding evidence. It also displays key indicators in time series graphs, box plots, etc., to aid in understanding the relationship between indicator changes and root causes. The interface and report output module provide interpretation results to external systems via REST API or message queue, and support exporting the interpretation results as PDF / HTML reports for use in quality meetings and process optimization reviews.

[0056] The incremental learning optimization and knowledge evolution module is used to collect user feedback and perform closed-loop updates on field mapping, knowledge graph edge weights, and inference strategies, thereby achieving continuous optimization of the knowledge graph, rules, and large model prompts. This includes a feedback collection module, a mapping and terminology update module, a graph incremental update module, a rule weight adjustment and forgetting mechanism module, and a prompt template and model evaluation module. The feedback collection module provides "confirm / reject / modify" entries for each root cause, causal chain, and suggestion on the front-end interface, recording user operations as FeedbackRecords, including: object type (link / rule / mapping), operation type, expected result after correction, user role, timestamp, etc. The mapping and terminology update module updates the field-entity mapping table and alias dictionary based on user modifications to field mappings, forming a mapping difference package (MappingDelta), and updating the mapping version. The graph incremental update module converts user-added or modified relationships, mechanism nodes, etc., into graph update patches (GraphPatch), writes them to the graph database in a transactional manner, and performs weight decay or deletion on edges with low confidence or those frequently rejected. The rule weight adjustment and forgetting mechanism module treats rules as "actions" in a policy, constructs a reward function based on user feedback, and updates rule weights through reinforcement learning methods such as policy gradients. Rules that have not been triggered for a long time or have received negative feedback multiple times are subject to weight decay until they are reduced to the deletion threshold and removed from the rule base, thus achieving rule "forgetting". The prompt template and model evaluation module are used to continuously monitor indicators such as structured output pass rate, evidence hit rate, and interpretation consistency. A / B experiments are conducted on different versions of prompt templates / models, and the version with better performance is given a higher weight or set as the default.

[0057] The system support and security governance module provides capabilities such as access control, log auditing, fault tolerance and recovery, version management, and privacy masking. It offers operational management and security governance for the entire system to ensure stable operation and compliant use. This module includes access control and authentication, logging and auditing, version management and fault tolerance and recovery, and data security and privacy protection. The access control and authentication module employs role-based access control (RBAC) to configure data access and function operation permissions for different roles, and combines single sign-on (SSO) or OAuth mechanisms to achieve unified identity authentication. The logging and auditing module records detailed logs for each interpretation request, data access, rule change, and graph update, supporting log queries by time, user, and object type for security auditing and compliance checks. The version management and fault tolerance and recovery module manages versions of important assets such as the knowledge graph, rule base, field mapping, and prompt templates, supporting rollback to historical versions in the event of abnormal or erroneous updates. The data security and privacy protection module employs data desensitization strategies (such as masking and generalization) and encrypted storage (such as using symmetric encryption algorithms AES and hash digest algorithms SHA series) for sensitive fields, and uses encrypted channels such as TLS in network transmission to avoid transmitting sensitive data in plaintext.

[0058] The following is a specific application example of the knowledge graph-enhanced large model interpretation system of this application in the scenario of abnormal yield diagnosis in injection molding production lines, to illustrate the application process of the above-described specific implementation method in actual production. This example takes the injection molding workshop of a machinery manufacturing company producing automotive interior parts as an example, with the target quality indicator being the part yield.

[0059] Example: Explanation and Application of Sudden Drop in Injection Molding Yield I. Business Background and Abnormal Phenomena During a certain shift, the workshop's MES / QMS system detected that the yield rate of a certain injection molding production line dropped from approximately 98% to 92% between 9:00 and 11:00 AM, accompanied by a significant increase in the proportion of "flash" and "porosity" defects in scrapped parts. Based on this, the upper-level quality analysis platform initiated an explanation request to the system of this invention, requesting clarification on "the main reasons for the sudden drop in yield rate and which process parameters and equipment parts should be prioritized for inspection."

[0060] II. Data Access and Preprocessing Stage First, the relevant data is accessed by the multi-source manufacturing data access module: The structured access module extracts process parameter curves such as injection speed, mold temperature, holding pressure, and cooling time from the MES system for the production line from 6:00 to 14:00 on the same day, as well as yield statistics and defect distribution data from the QMS system. The unstructured access module retrieves standard documents such as "Guide to Injection Molding Process Defect Investigation" and "Mold Maintenance Specification" from the document library, and accesses part appearance images (multiple defect samples labeled "flash" and "pores") from the vision inspection system within the same time period. The real-time stream access module connects to the SCADA / PLC to record real-time injection pressure, screw position curves, and equipment alarm logs (including "low flow rate in mold cooling circuit" alarm). The metadata collection module synchronously extracts information such as field structure, unit, and field description from relevant data tables.

[0061] Then, the incoming data enters the data preprocessing and standardization module: The data cleaning module removes duplicate records and rearranges out-of-order timestamps, repairs missing sensor readings using linear interpolation, and identifies and removes obvious erroneous values ​​(such as mold temperature jumping to 0 instantly) using isolated forest. The format conversion and standardization module unifies the naming of fields such as "mold temperature" and "MOLD_TEMP" in different systems as "mold temperature", and unifies the temperature unit to ℃ and the pressure unit to bar, converting various data sources into a unified time series and event record format; The statistical summary and distribution drift module calculated the mean, variance, and distribution drift index (PSI, KS) of fields such as mold temperature, injection speed, and holding pressure for two time windows: 6:00–9:00 and 9:00–11:00. It was found that the average mold temperature in the 9:00–11:00 time period was about 8°C lower than the historical average, with increased fluctuations, and the flash and porosity defect rates increased significantly. The multimodal feature extraction and cross-modal alignment module encodes the quality inspection report text, defect images, and alarm logs into multimodal features: the text side is encoded by BERT as... On the image side, semantic vectors of defect regions are extracted using YOLO+BLIP. The data is mapped to a unified semantic space via a dual-tower contrastive learning network, resulting in comparable semantic spaces. , Vectors are used for subsequent defect pattern recognition and evidence alignment.

[0062] III. Knowledge Graph Construction and Field Alignment The manufacturing knowledge graph construction and storage service module has already built a graph for the injection molding field based on enterprise process specifications and historical cases. This embodiment will use some of the knowledge from this graph. The entity and term management module has defined the following entities: process parameters (such as "mold temperature", "first stage injection speed", "holding pressure"), quality indicators (yield), defect modes (flash, porosity), mechanism states (filling resistance, melt viscosity, chilling layer thickness), equipment components (cooling circuit, nozzle), etc. The entity extraction module has extracted entities such as "injection speed" and "flash" from the "Guideline for Injection Molding Process Defect Investigation" and mapped the "flash" defect entity with the defect image output by the vision inspection system through multimodal alignment results; The relationship extraction module extracts the following relationship examples from the specifications and process documents: (mold temperature, NEGATIVELY_CORRELATED, chilling layer thickness), (chilling layer thickness, POSIVITELY_CORRELATED, filling resistance), (filling resistance, CAUSES, air entrapment), (air entrapment, CAUSES, porosity defects), (first-stage injection speed, CAUSES, flash).

[0063] The segment semantic parsing and alignment module aligns the data fields involved in this process: The field description construction module constructs FieldDescriptors for fields such as "MOLD_TEMP_ZONE1" and "INJ_SPD_V1" in MES, which include Chinese descriptions such as "mold temperature (front mold)" and "first-stage injection speed", units, value ranges, and sample statistical summaries. The field semantic embedding generation module encodes field descriptions into field embeddings. and embedding with map entities Compare; The candidate entity recall and multi-factor scoring module calculates the score, resulting in, for example, "MOLD_TEMP_ZONE1". The entity "Mold Temperature (Cavity Surface)" (Score=0.93); "INJ_SPD_V1" Entity "first-stage injection rate" (Score=0.91); After the constraint verification and confirmation module confirms that the unit and value range match, it automatically establishes a field → entity mapping relationship, and the mapping version management module records the version number.

[0064] IV. Interpreting Requests and Context Building When QMS detects a yield drop event, the upper-layer platform submits an explanation request to this system, and the explanation request management and context retrieval module begins to work: The ExplanationRequest parsing module receives the following request parameters: 1. Target quality indicator: Yield (QualityMetric entity); 2. Analysis type: Anomaly explanation; 3. List of key fields and anomaly scores: mold temperature (high anomaly score), first-stage injection speed, holding pressure, etc.; 4. Time range: 9:00–11:00 on the same day; 5. Equipment identifier: Injection molding machine A1. After parsing, an ExplanationRequest object is generated.

[0065] The key factor screening module selects mold temperature and first-stage injection speed as Top-K key factors based on the feature importance and anomaly indicators of the upstream model, and uses the mapping results of the module to obtain the corresponding graph entity set.

[0066] The context subgraph retrieval module takes the entity "yield" as the endpoint and entities such as "mold temperature," "first-stage injection speed," and "cooling circuit" as the starting points. It performs K-hop neighborhood retrieval and constrained path search in the knowledge graph, resulting in paths such as: Path 1: Mold temperature ↓ → Cooling layer thickness ↑ → Filling resistance ↑ → Air entrapment ↑ → Porosity defects ↑ → Yield ↓; Path 2: First-stage injection speed ↑ → Impact of the filling front on the mold cavity → Flash defects ↑ → Yield ↓.

[0067] The evidence retrieval and fusion module retrieves statistical deviations of mold temperature and first-stage injection speed within the 9:00–11:00 window from the statistical summary and distribution drift module; it retrieves multiple "low cooling circuit flow" alarms from the equipment log; it retrieves images of "flash and porosity" defects during this period from the visual inspection system and aligns them with the defect entities using multimodal features; and it extracts paragraphs related to mold temperature, air entrapment, and porosity from the procedure document. All of this evidence is packaged into an EvidencePackage, which includes evidence type, time range, associated entities, and a brief summary.

[0068] The similar case retrieval module searches the historical case database for historical events similar to the pattern of "decreased yield + porosity + low mold temperature". It finds that many past cases are related to "cooling circuit blockage" and "abnormal cooling water temperature" and extracts their typical repair measures as a reference.

[0069] The context packaging module encapsulates the context subgraph, key entities, statistical summaries, multimodal evidence packages, similar case summaries, and output structure constraints into a ContextPackage for subsequent use by larger models.

[0070] V. Large-scale model explanation and causal chain generation The large model explanation engine and inference orchestration module generate explanations based on ContextPackage: The layered prompt template library module selects the "Quality Anomaly Explanation" template, injecting "the influence mechanism of mold temperature on melt viscosity, air entrapment and porosity" and "the influence law of first-stage injection speed on flash defects" into the domain knowledge layer; listing the statistical deviation, image evidence, alarm records and similar historical cases of this event in the evidence constraint layer; and specifying the output fields summary, root_causes, causal_chains, recommendations, etc. in the structural constraint layer. The interpreter constructs the ExplanationRequest and ContextPackage into a Structured Prompt based on a template. The large model calls the corresponding domain-specific large language model, and generates preliminary explanation results based on the prompts. The multi-stage reasoning module first summarizes the phenomena (decreased yield, changes in defect types), then proposes mechanism hypotheses (low mold temperature leads to air entrapment, high injection speed leads to flash), then combines subgraphs to fit the causal chain, and finally gives suggestions for equipment and process adjustments. The output self-repair and structured constraint module performs JSON schema and evidence reference verification on the output, performs a self-repair for missing fields, and finally outputs a structured interpretation result.

[0071] The structured results are then fed into the module for attributing anomalies and generating causal chains. The entity linking and standardization module links entities such as "low mold temperature", "insufficient cooling", "air entrapment", and "air hole" in the explanatory text to the graph nodes, and eliminates ambiguity by combining the time range and equipment number; The causal chain extraction module extracts and constructs two main chains from the structured output: Chain A: Abnormal cooling circuit flow → Unstable mold temperature control, low average temperature → Increased chilled layer thickness → Increased filling resistance → Increased gas retention (air entrapment) → Increased porosity defect ratio → Decreased yield; Chain B: First-stage injection speed setting is too high → High-speed melt impacts the cavity parting surface → Increased flash defect ratio → Decreased yield. The logic verification and conflict resolution module performs logical verification on the above links based on the path patterns in the graph and the domain common sense base, confirming that they are consistent with the existing high-confidence rules and have no obvious conflicts. The confidence assessment module integrates the edge weights of the graph path, the significance of statistical bias (the mold temperature deviates significantly from the historical distribution), multimodal evidence (images of pores and flash), and the support of similar cases to calculate the link confidence. For example, the confidence of link A is 0.88, and the confidence of link B is 0.79. The Explanation Subgraph Generation and Write-back module writes links A and B back to the knowledge graph in the form of explanation subgraphs. Due to the support of multiple historical cases and the "confirmation" of this feedback, the weight of the relevant edges is increased, providing stronger knowledge support for similar scenarios in the future.

[0072] VI. Results Presentation and Feedback Closed Loop The structured interpretation and multi-role display module outputs differentiated views based on different user roles: The structured output module outputs two main causes in the root_causes section: 1) Abnormal cooling circuit flow leads to lower mold temperature and increased fluctuations, resulting in more air entrapment and porosity defects, which is the main reason for the decrease in yield; 2) The first-stage injection speed is set too high, which aggravates flash defects and is a secondary cause.

[0073] The multi-role view adaptation module is geared towards process engineers, highlighting the deviation range of the mold temperature curve, the corresponding changes in porosity defect rate, and parameter adjustment suggestions; towards equipment maintenance personnel, it emphasizes the "cooling circuit flow alarm" and recommended equipment components for inspection; and towards management personnel, it provides brief conclusions and expected improvement margins.

[0074] The visualization module displays links A and B in a node-edge cause-effect graph format, and overlays the mold temperature and defect ratio curves on the yield trend graph; The interface and report output module pushes the interpretation results to the enterprise quality dashboard via REST API and generates PDF reports for quality meeting discussions.

[0075] In actual use, process engineers and equipment maintenance personnel cleaned and repaired the cooling circuit, adjusted the cooling water temperature settings, and appropriately reduced the first-stage injection speed based on the interpretation results. Subsequent monitoring data from several shifts showed a significant decrease in porosity and flash defects, with the yield recovering to approximately 98%. The user marked the interpretation result as "confirmed as valid" on the front-end interface.

[0076] The incremental learning optimization and knowledge evolution module received this confirmation feedback: The feedback collection module records this FeedbackRecord and marks links A and B as "accepted"; The mapping and terminology update module updates the alias dictionary based on expressions such as "cooling circuit blockage" and "abnormal cooling water temperature" added by engineers; The graph incremental update module increases the edge weights of the path "abnormal cooling circuit flow → unstable mold temperature → porosity defects → decreased yield". The rule weight adjustment and forgetting mechanism module increases the priority of this type of causal pattern in subsequent explanations based on this positive feedback; The prompt template and model evaluation module records the pass rate and user satisfaction of the structured output in this explanation task, which will be used for subsequent prompt template optimization.

[0077] Throughout the process, the system support and security governance module, through the permission and authentication module, log and audit module, version management and fault tolerance recovery module, and data security and privacy protection module, conducts full-process control over data access, operation behavior, and version changes, ensuring the security, traceability, and recoverability of the interpretation task in this embodiment.

[0078] As can be seen from this embodiment, the system of this application can organically combine multi-source heterogeneous data, multimodal evidence, manufacturing knowledge graphs, and large model reasoning capabilities in complex injection molding production environments, automatically provide interpretable, traceable, and executable yield anomaly cause analysis and rectification suggestions, and continuously optimize the knowledge graph and interpretation strategy through feedback loop.

[0079] Please refer to Figure 3 As shown, an embodiment of this application provides a knowledge graph interpretation device 100 for the machinery manufacturing industry, the device comprising: The acquisition unit 110 is used to acquire the request parameters corresponding to the interpretation request in response to receiving the interpretation request; The filtering unit 120 is used to filter the Top-K key fields in the feature field pool corresponding to the model when the explanation request comes from the model output, so as to obtain a number of key fields. Mapping unit 130 is used to map several key fields to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities; The retrieval unit 140 is used to perform neighborhood retrieval and constraint path retrieval based on the pre-built manufacturing knowledge graph, with the target quality indicator entity in the key entity as the center, so as to obtain the shortest path and the highest weight path that conforms to the preset pattern. The original evidence acquisition unit 150 is used to extract relevant original evidence corresponding to the key entity set from the multimodal original manufacturing data corresponding to the pre-built manufacturing knowledge graph and package it to obtain the original evidence package. The historical evidence acquisition unit 160 is used to perform similar searches in several historical cases to obtain a historical evidence package; The explanation generation unit 170 is used to encapsulate the key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package to obtain the explanation content corresponding to the explanation request.

[0080] Embodiments of this application also provide a computer program product including program code that, when the program product is run on an electronic device, causes the electronic device to perform the steps of the methods described above according to various exemplary embodiments of this application.

[0081] Furthermore, although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0082] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0083] In an exemplary embodiment of this application, an electronic device capable of implementing the above-described method is also provided.

[0084] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0085] An electronic device according to this embodiment of the present application. The electronic device is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of this application.

[0086] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and buses connecting different system components (including memory and processor).

[0087] The memory stores program code that can be executed by a processor, causing the processor to perform the steps described in the "Exemplary Methods" section above, according to various exemplary embodiments of this application.

[0088] The storage may include readable media in the form of volatile storage, such as random access memory (RAM) and / or cache memory, and may further include read-only memory (ROM).

[0089] The storage may also include programs / utilities having a set (at least one) of program modules, including but not limited to: an operating system, one or more applications, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0090] A bus can represent one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus that uses any of the various bus architectures.

[0091] The electronic device can also communicate with one or more external devices (such as keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (such as routers, modems, etc.). This communication can be performed via input / output (I / O) interfaces. Furthermore, the electronic device can communicate with one or more networks (such as local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. As shown in the figure, the network adapter communicates with other modules of the electronic device via a bus. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0092] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this application.

[0093] In exemplary embodiments of this application, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible implementations, various aspects of this application may also be implemented as a program product including program code, which, when the program product is run on a terminal device, causes the terminal device to perform the steps of the various exemplary embodiments of this application described in the "Exemplary Methods" section above.

[0094] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0095] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0096] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0097] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0098] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of this application, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0099] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0100] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for interpreting knowledge graphs in the machinery manufacturing industry, characterized in that the method includes: In response to receiving an explanation request, the request parameters corresponding to the explanation request are obtained; When the explanation request comes from the model output, the Top-K key fields are filtered in the feature field pool corresponding to the model to obtain a number of key fields; Several key fields are mapped to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities; Centered on the target quality indicator entity among the key entities, neighborhood search and constraint path search are performed based on the pre-constructed manufacturing knowledge graph to obtain the shortest path and the highest weight path that conforms to the preset pattern. Extract relevant original evidence corresponding to key entity sets from the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph and package them to obtain the original evidence package; Similar searches are conducted in several historical cases to obtain a package of historical evidence; The key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package are encapsulated to obtain the explanation content corresponding to the explanation request.

2. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 1, characterized in that, The construction process of the pre-built manufacturing knowledge graph includes: Collect multimodal raw manufacturing data; The multimodal raw manufacturing data is preprocessed and standardized to obtain a characteristic manufacturing dataset; wherein the characteristic dataset has a statistical summary and a unified multimodal semantic representation; A manufacturing knowledge graph is constructed based on a feature-based manufacturing dataset.

3. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 1, characterized in that, The process of mapping several key fields to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities includes: Several key fields are subjected to semantic parsing and cross-system schema alignment to obtain a set of key entities. The semantic parsing and cross-system schema alignment include: field description construction, field semantic embedding generation, candidate entity recall and multi-factor scoring, constraint verification and confirmation, and mapping version management.

4. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 1, characterized in that, Extract relevant original evidence corresponding to key entity sets from the multimodal raw manufacturing data corresponding to the pre-constructed manufacturing knowledge graph and package it to obtain the original evidence package, including: In the text and document-side data of the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph, procedures, standards and historical reports are retrieved to obtain original evidence of document fragments related to the key entity set; For the structured and time-series data of the multimodal raw manufacturing data corresponding to the pre-built manufacturing knowledge graph, the statistical summary and drift index raw evidence within the time window corresponding to the interpretation request are read from the preprocessing module. The drift index is used to measure whether the data distribution has changed significantly. In the image-side data of multimodal raw manufacturing data corresponding to the pre-constructed manufacturing knowledge graph, image defect information and visual features of the same time period, the same equipment or the same workstation are retrieved to obtain original image evidence. The original evidence package consists of document fragments related to the key entity set, statistical summaries and drift indicators, and images.

5. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 1, characterized in that, The process of performing similar searches on several historical cases to obtain a package of historical evidence includes: In several historical cases, similarity retrieval is performed based on the target quality indicator entity, the root cause entity combination, and the graph embedding vector to obtain several similar historical cases. Among them, the root cause entity combination is the set of root cause entities in the key entity set that have a causal relationship with the target quality indicator entity, and the graph embedding vector is the vector obtained by graph embedding calculation on the context subgraph composed of the target indicator entity, the root cause entity, and their relationship. The key causal chains and conclusion summaries of each of several similar historical cases are encapsulated and packaged to obtain a historical evidence package.

6. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 1, characterized in that... After encapsulating the key entity set, the shortest path and the highest weight path conforming to a preset pattern, the original evidence package and the historical evidence package to obtain the explanation content corresponding to the explanation request, the method further includes: Under the constraints of a pre-built manufacturing knowledge graph, original evidence package, and historical evidence package, the explanation content corresponding to the explanation request is generated into a structured explanation and output.

7. The method for interpreting knowledge graphs in the machinery manufacturing industry according to claim 6, characterized in that... After generating a structured explanation output corresponding to the explanation request, under the constraints of a pre-constructed manufacturing knowledge graph, original evidence package, and historical evidence package, the method further includes: Based on user feedback on structured interpretation, the field mapping, knowledge graph edge weights, and reasoning strategies are updated in a closed loop.

8. A knowledge graph interpretation device for the machinery manufacturing industry, characterized in that the device comprises: The acquisition unit is used to acquire the request parameters corresponding to the interpretation request in response to receiving the interpretation request; The filtering unit is used to filter the Top-K key fields in the feature field pool corresponding to the model when the explanation request comes from the model output, so as to obtain a number of key fields. The mapping unit is used to map several key fields to entities in a pre-built manufacturing knowledge graph to obtain a set of key entities. The retrieval unit is used to perform neighborhood retrieval and constraint path retrieval based on the pre-built manufacturing knowledge graph, with the target quality indicator entity among the key entities as the center, in order to obtain the shortest path and the highest weight path that conforms to the preset pattern. The original evidence acquisition unit is used to extract relevant original evidence corresponding to key entity sets from the multimodal original manufacturing data corresponding to the pre-built manufacturing knowledge graph and package it to obtain the original evidence package. The historical evidence acquisition unit is used to perform similar searches in several historical cases to obtain a historical evidence package; The explanation generation unit is used to encapsulate the key entity set, the shortest path and the highest weight path that conform to the preset pattern, the original evidence package and the historical evidence package to obtain the explanation content corresponding to the explanation request.

9. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 9.

Images