MRO industrial product recommendation method based on knowledge graph

Abstract

The invention relates to the technical field of industrial product data processing, in particular to an MRO industrial product recommendation method based on a knowledge graph, and the method comprises the steps: collecting structured data and unstructured data, and carrying out the data transmission and integration of the collected multi-source data; performing data cleaning and standardized conversion on the integrated data, and performing knowledge extraction based on the converted data; obtaining an ontology and a relationship of the MRO field, and mapping the ontology and the relationship with the extracted entity and the relationship to construct a knowledge graph; according to the method, the current query demand of a user is obtained and analyzed, knowledge graph traversal and sorting are performed according to the initial entity node obtained through analysis, the recommendation list is generated, potential commodity association is mined through the semantic reasoning ability of the knowledge graph, the problem of data sparsity is effectively relieved, and the recommendation coverage rate of long-tail commodities and unpopular demands is increased.

Description

Technical Field

[0001] This invention relates to the field of industrial product data processing technology, and in particular to a knowledge graph-based MRO industrial product recommendation method. Background Technology

[0002] Traditional MRO (Maintenance, Repair, and Operation) industrial product recommendation systems typically rely on collaborative filtering or content filtering techniques, depending on user historical behavior data or product attributes for matching. However, MRO industrial products are characterized by their diverse categories, highly specialized specifications, and varied application scenarios, leading to the following limitations of traditional methods: First, data sparsity, particularly hindering the effective coverage of long-tail products and niche needs; second, a lack of semantic understanding, failing to capture deep functional, compatibility, or scenario-based relationships between products; and third, weak dynamic demand adaptation capabilities, making it difficult to combine multi-dimensional information such as repair scenarios, equipment types, and environmental factors for accurate recommendations. Existing technologies are largely limited to structured data queries, failing to fully utilize the advantages of knowledge graphs in semantic reasoning and relationship mining, resulting in low recommendation accuracy and coverage. Summary of the Invention

[0003] The purpose of this invention is to provide a knowledge graph-based recommendation method for MRO industrial products. By leveraging the semantic reasoning capabilities of knowledge graphs, it can uncover potential product associations, effectively alleviate the data sparsity problem, and improve recommendation accuracy and coverage.

[0004] To achieve the above objectives, this invention provides a knowledge graph-based method for recommending MRO industrial products, comprising the following steps: It collects structured and unstructured data, and transmits and integrates the collected multi-source data; The integrated data is cleaned and standardized, and knowledge is extracted based on the transformed data. Obtain the ontology and relationships in the MRO domain, and map them to the extracted entities and relationships to construct a knowledge graph; The system acquires and parses the user's current query request, performs knowledge graph traversal and sorting based on the initial entity nodes obtained from the parsing, and generates a recommendation list.

[0005] This includes collecting structured and unstructured data, and transmitting and integrating the collected multi-source data, including: Obtain structured and unstructured data, and determine the update frequency of the structured data; The unstructured data is transmitted using distributed object storage technology, with the corresponding transmission method matched to the update frequency of the structured data. The received multi-source data is encapsulated according to the set encapsulation format and cached in the data lake; After the data in the data lake is extracted, transformed, and loaded, it is stored in the data warehouse.

[0006] The method further includes: Data quality checkpoints are set up at key nodes in the transmission channel to verify data quality.

[0007] This includes data cleaning and standardization of the integrated data, followed by knowledge extraction based on the transformed data, including: The integrated structured data is cleaned and standardized, and the integrated unstructured data is processed into text data. Automatically identify and extract key entities from unstructured text using natural language processing techniques; A method combining rule-based matching and deep learning relation extraction models is used to extract triple relationships between entities and entity attributes from the structured and unstructured data.

[0008] Specifically, the automatic identification and extraction of key entities from the unstructured data using natural language processing technology includes: Using a Transformer-based pre-trained language model as a foundation, and based on a specific sequence labeling strategy, key entities and their attribute sets are automatically identified and extracted from the unstructured data.

[0009] Specifically, a method combining rule-based matching and deep learning relation extraction models is employed to extract triple relationships between entities and entity attributes from the structured and unstructured data, including: A method combining rule-based matching and deep learning relation extraction models is used to obtain multi-source evidence from the structured and unstructured data, and corresponding confidence scores are set. Link the key entities to the standard entity nodes of the knowledge graph; The multi-source evidence with multiple meanings is disambiguated and then stored in a temporary knowledge base.

[0010] This process involves acquiring and parsing the user's current query request, performing knowledge graph traversal and sorting based on the parsed initial entity nodes, and generating a recommendation list, including: Obtain the user's current query request and identify the query entities within the query request; By utilizing the contextual information in the knowledge graph, the query entity is disambiguated and mapped to a unique graph node to obtain the initial entity node; Based on the initial entity nodes, intelligent traversal and reasoning are performed to generate a candidate set; The candidate set is calculated based on multi-dimensional ranking factors, and a recommendation list is generated after dynamic weight adjustment.

[0011] The multi-dimensional ranking factors include graph correlation score, business rule score, and collaborative signal score; wherein the graph correlation score is the product of the sum of the weights of all inference paths between the candidate product node and the initial entity node and the path length penalty factor.

[0012] Specifically, the candidate set is calculated based on multi-dimensional ranking factors, and after dynamic weight adjustment, a recommendation list is generated, including: The weight values ​​corresponding to the candidate set are calculated based on the multi-dimensional ranking factors, and the similarity between the user vector and the product vector is calculated based on the user's historical interaction data. The weighted fusion method is used to calculate the comprehensive recommendation score corresponding to the weight value and the similarity, and the recommendation list is generated by sorting the comprehensive recommendation scores.

[0013] This invention discloses a knowledge graph-based MRO (Maintenance, Repair, and Overhaul) industrial product recommendation method. The method involves collecting structured and unstructured data, transmitting and integrating the collected multi-source data, cleaning and standardizing the integrated data, and extracting knowledge based on the transformed data. It then acquires the ontology and relationships within the MRO domain and maps them to the extracted entities and relationships to construct a knowledge graph. Finally, it acquires and parses the user's current query requirements, traverses and sorts the knowledge graph based on the parsed initial entity nodes, generates a recommendation list, and leverages the semantic reasoning capabilities of the knowledge graph to uncover potential product associations. This effectively alleviates data sparsity issues and improves the recommendation coverage of long-tail products and niche needs. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1 This is a schematic diagram illustrating the steps of a knowledge graph-based MRO industrial product recommendation method provided by the present invention.

[0016] Figure 2 This is a flowchart illustrating a knowledge graph-based MRO industrial product recommendation method provided by the present invention.

[0017] Figure 3 This is a flowchart illustrating the knowledge graph-based MRO industrial product recommendation method provided by the present invention. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0019] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0020] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0021] Please see Figures 1-3 This invention provides a knowledge graph-based method for recommending MRO industrial products, comprising the following steps: S101. Collect structured and unstructured data, and transmit and integrate the collected multi-source data.

[0022] Specifically, raw data is collected from multiple heterogeneous sources to lay the foundation for knowledge graph construction. The collected data includes structured and unstructured data. Structured data includes static product data and dynamic business data. Static product data: Structured data for MRO industrial products is obtained from supplier databases, product catalogs, and technical manuals, including SKUs, names, specifications (such as dimensions, materials, rated voltage), classification information, and compatibility lists. Dynamic business data: Historical transaction data, inventory status, purchase frequency, and supplier information are extracted from Enterprise Resource Planning (ERP), Warehouse Management System (WMS), and purchasing records. Unstructured text data: Text information describing equipment, faults, maintenance scenarios, and parts usage relationships is obtained from maintenance reports, equipment maintenance manuals, fault descriptions, engineer work logs, and online consultation content.

[0023] The collected multi-source data is then securely and efficiently transmitted and aggregated into a unified data processing platform.

[0024] During transmission, for unstructured data, distributed object storage (such as Amazon S3, Azure Blob Storage, or a self-built MinIO) is typically used as the core hub. The data source uploads files to a specific "bucket" in the object storage via API or client tools. The object storage returns a unique file access address. Subsequently, the system only needs to notify downstream processing systems of this address and file metadata (such as name, size, and type) through message queues or database records, greatly reducing the load on the core data transmission channel.

[0025] For structured data, it's necessary to determine its update frequency. This applies to static product data with large volumes but low update frequencies, such as complete product catalogs provided by suppliers, backlogged maintenance reports, and equipment technical manuals. This data is typically updated daily, weekly, or monthly. Mature ETL tools (such as Apache NiFi, Talend, and Kettle) or custom scripts are used. These tools extract data from source databases (such as SQL Server and Oracle) or file servers (FTP / SFTP), perform preliminary transformations (such as encoding conversion and simple cleaning) in memory or intermediate servers, and finally load the data into the target data lake. To improve efficiency, not all data is transferred each time. The system uses timestamps (such as last_modified_time) or version number fields in the data, or utilizes database change data capture (CDC) technology, to capture only newly added and updated data that has changed since the last synchronization for incremental synchronization.

[0026] For dynamic business data that needs to rapidly reflect changes in system status, such as real-time user search logs, clickstreams, generating maintenance work orders, and dynamic inventory changes (inbound / outbound), the value of this type of data diminishes rapidly over time, requiring low-latency processing. A high-throughput distributed message middleware (such as Apache Kafka, RabbitMQ, or RocketMQ) is used as the transmission hub. Data producers (such as front-end applications or ERP system log modules) act as producers, publishing data as messages to designated topics. The data processing platform acts as a consumer, subscribing to these topics and consuming messages in real time. This publish-subscribe (Pub-Sub) model decouples the data source from the processing system, providing powerful buffering and peak-shaving capabilities.

[0027] All incoming data, regardless of source or format, first enters a raw data layer. This layer faithfully preserves the original appearance of the data for auditing and traceability, using a schema registry (such as Confluent Schema Registry). When producers send data, it is encapsulated according to a predefined format (commonly using efficient serialization formats such as Avro and Protobuf) to ensure that downstream consumers can accurately parse the structure and meaning of the data. The data lake (built on HDFS, S3, etc.) receives and stores all types of raw data and data that has undergone initial cleaning. It stores massive amounts of raw data at low cost, maintaining maximum flexibility and allowing subsequent processing programs to parse and utilize it on demand. For some highly structured data that needs to support high-frequency queries and BI analysis (such as cleaned product master data and transaction data), it is further extracted, transformed, and loaded into a data warehouse (such as Amazon Redshift, Snowflake, BigQuery). The models in the data warehouse are optimized for analysis, facilitating rapid relational queries and providing high-quality structured data input for knowledge graph construction.

[0028] For highly structured data that requires support for high-frequency queries and BI analysis (such as cleaned product master data and transaction data), further extraction, transformation, and loading into a data warehouse (such as Amazon Redshift, Snowflake, or BigQuery) are performed from the data lake. Models in the data warehouse are optimized for analysis, facilitating rapid relational queries and providing high-quality structured data input for knowledge graph construction. All cross-network data transmissions use TLS / SSL encrypted channels to prevent data theft or tampering during transmission. Strict authentication (such as Kerberos, IAM) and authorization (such as RBAC permission models) are implemented for access to data sources, message queues, and data lakes / warehouses to ensure that only authorized systems and services can read and write data.

[0029] S102. Perform data cleaning and standardization transformation on the integrated data, and extract knowledge based on the transformed data.

[0030] Specifically, for the structured data, missing value handling is performed first: for missing fields such as product parameters and supplier information, a rule-based filling strategy is adopted. For example, for key parameters such as "rated voltage," the missing value is filled by querying common values ​​of similar products or higher-level categories; for non-key fields, they are marked as "unknown" instead of being simply deleted to avoid information loss. Then, outlier detection and correction are performed: outlier values ​​are identified using statistical methods (such as Z-Score) and business rules (such as the physical probability range: screw length cannot be negative). For obviously erroneous values, correction is performed by tracing the data source or based on historical versions.

[0031] Finally, standardization and normalization are implemented: Unit unification: All parameters are converted to international standard units (e.g., "inches" are unified to "millimeters," and "Fahrenheit" is converted to "Celsius"). Format unification: Date formats, string case, and symbol representations are standardized (e.g., "PTFE," "polytetrafluoroethylene," and "Teflon" are unified to the standard term "polytetrafluoroethylene"). Classification unification: Products are reclassified and labeled according to international standards (e.g., UNSPSC, eCl@ss) or the company's own standard classification system to eliminate classification ambiguities between different data sources.

[0032] For unstructured text data, preprocessing is first performed, including word segmentation, stop word removal, and lemmatization of maintenance reports, work order descriptions, etc., to transform them into standardized text units. Then, a domain dictionary is built and expanded: based on existing product libraries and standard manuals, a professional MRO dictionary (including equipment models, part numbers, fault types, and action terms) is constructed. An innovative active learning mechanism is introduced: the system automatically detects high-frequency new words not covered by the dictionary and submits them to domain experts for review and annotation, thereby continuously and dynamically expanding the dictionary and improving its ability to identify emerging technologies and new parts models.

[0033] Based on a Transformer-based pre-trained language model (such as BERT), a specific sequence labeling strategy is designed. The model not only labels entity boundaries and types (e.g., "bearing", "deep groove ball", "6305") in the text, but also labels the entity's attribute values ​​(e.g., "inner diameter: 25mm", "outer diameter: 62mm"). For example, from the sentence "Replace NSK brand model 6305ZZ deep groove ball bearing, with an inner diameter of 25mm and an outer diameter of 62mm", the model can extract the entity "bearing" in one go, along with its attribute set {"brand: NSK", "model: 6305ZZ", "type: deep groove ball", "inner diameter: 25mm", "outer diameter: 62mm"}. This greatly enriches the descriptive dimensions of entities, laying the foundation for subsequent accurate matching and recommendation.

[0034] Evidence supporting the relationship hypothesis (head entity H, relationship R, tail entity T) is obtained from three main dimensions: Structured source evidence: Explicitly existing (H, R, T) triples are directly extracted from structured data such as BOMs, product compatibility lists, and supplier directories. This is the most reliable source of evidence. Textual source evidence: Sentence fragments expressing the (H, R, T) relationship are extracted from unstructured text such as repair reports, technical forums, and user manuals using an attention-based relationship extraction model. For example, (sensor X, replaced, model Y) can be extracted from the sentence "When sensor X malfunctions, it can be replaced by model Y". Behavioral source evidence: The co-occurrence strength between entities is calculated by analyzing historical purchase orders, search logs, and browsing behavior. For example, similarity calculations in association rule mining (such as the Apriori algorithm) or collaborative filtering can reveal that parts E and F are frequently purchased or searched by the same engineer, thus providing potential evidence for the (E, related, F) relationship.

[0035] Before collecting evidence, it is crucial to ensure that the "same entity" from different data sources has been aligned to a single standard entity in the knowledge graph. This is the foundation for all subsequent operations.

[0036] For each piece of evidence supporting (H, R, T), the system transforms it into a set of quantifiable feature vectors for processing by the fusion model. Features include: Source authority: Where does the evidence originate? Product manuals have a much higher weight than forum posts. Frequency of occurrence: The number of times the relation is mentioned across different sources. Evidence strength: For textual evidence, the probability score given by the relation extraction model; for behavioral evidence, significant indicators of co-occurrence statistics (such as lift and confidence). Time freshness: The time when the evidence was generated; newer evidence typically has a higher weight.

[0037] The feature vector set of all evidence for (H, R, T) is input into a neural network for processing. The network outputs whether the relationship is true and a confidence score between 0 and 1, quantifying the credibility of the relationship. For example, (screw A, compatible, device B) has a confidence score of 0.95, while (lubricant C, used for, device D) has a confidence score of only 0.65. Relationships with confidence scores exceeding a high threshold (e.g., 0.9) are directly stored as high-quality knowledge in the knowledge graph; relationships with confidence scores below a low threshold (e.g., 0.4) are discarded; relationships with confidence scores in the middle range or conflicting (e.g., different sources provide completely contradictory evidence) are sent to a human review queue, along with all evidence sources and confidence calculation details, for domain experts to make a final decision. The experts' decisions are then fed back to the evidence fusion model for continuous optimization.

[0038] The MRO (Maintenance, Repair, and Operations) field contains a large number of long-tail products and new products with very little relevant data (few samples), making it difficult for traditional deep learning models to effectively extract relationships (the cold start problem). However, these products often need to be recommended. Therefore, a two-stage approach of "pre-training + few-sample learning" is adopted to discover implicit relationships.

[0039] First, a Transformer architecture (such as BERT) is used as the model. Pre-training is performed on large-scale general text corpora (such as industrial technical documents, Wikipedia, and patent documents). The task at this stage is typically masked language modeling (MLM), allowing the model to deeply understand language syntax and general semantics. Then, secondary pre-training is performed on public relation extraction datasets (such as FewRel), specifically adapting the model to the "relation classification" task paradigm. At this point, the model has become a powerful "relation recognition expert," possessing the strong potential to capture semantic relationships from sentences.

[0040] This paper employs a prototype network, a few-shot learning algorithm, to identify a new MRO relationship (such as a "substitution relationship"). First, a support set is constructed, which only requires a very small number of example sentences containing the substitution relationship, such as 3-5. For example: Support sample 1: "Part A can replace part B."; Support sample 2: "When model C is out of stock, we use model D as a substitute."; Support sample 3: "The new model E is an upgraded replacement for the old model F." Each sample sentence in the support set is input into a pre-trained model, and the sentence representation vector output by the model is extracted. The mean of the sentence vectors of all support samples for the same relation (in this example, "substitution") is calculated to obtain a representative vector, called the "prototype" of that relation. This "prototype" vector defines the core conceptual region of the "substitution relation" in the vector space. When a new unknown sentence (query sample) is encountered, it is also input into the model to obtain its sentence representation vector. The distance between this query vector and all previously calculated relation "prototype" vectors, such as Euclidean distance or cosine distance, is calculated. The relation type corresponding to the relation "prototype" with the closest distance is selected as the predicted relation for the query sample. For example, if the vector of a new sentence "The solution is to replace component H with component G" is closest to the "substitution relation prototype," the system determines that it expresses a substitution relation and extracts (component G, substitution, component H).

[0041] The extracted entity references are linked to standard entity nodes in the knowledge graph. For example, "6305ZZ bearing" and "NSK 6305ZZ" appearing in the text are correctly linked to the standard entity bearing_NSK_6305ZZ. Based on the context, relations with multiple meanings are disambiguated. For example, the word "connection" might mean "circuit connection" in an electrical context and "bolt connection" in a mechanical context; the system needs to select the correct relation type based on the context information. The processed high-quality triples are first stored in a temporary knowledge base, and after a round of quality verification or expert review, they are officially and batch-updated to the knowledge graph in the production environment to ensure the stability and accuracy of the graph.

[0042] S103. Obtain the ontology and relationships in the MRO domain, and map them to the extracted entities and relationships to construct a knowledge graph.

[0043] Specifically, the upper-level ontology of the MRO domain is first defined, clarifying core concepts (such as "product," "equipment," "fault," and "supplier") and their interrelationships (such as "compatible with," "used for," "replace," and "cause"), providing a semantic framework and logical constraints for the knowledge graph. The previously extracted entities and relationships are mapped to the ontology layer. Entity linking technology is used to eliminate different representations pointing to the same entity (such as "PTFE sealing ring" and "polytetrafluoroethylene sealing ring"), completing knowledge fusion. Finally, the processed triplet data is stored in a graph database, leveraging its powerful association query capabilities to efficiently store and manage the knowledge graph. Specifically, the triplet data uses knowledge graph embedding techniques, such as the TransE model, to learn low-dimensional vector representations of entities and relationships. The TransE model is chosen as the knowledge graph embedding model, randomly initializing a low-dimensional vector representation for each entity and relationship in the knowledge graph. Model training: Using the knowledge graph data, the TransE model is trained using optimization algorithms (such as stochastic gradient descent). During training, the model continuously adjusts the vector representations of entities and relationships to minimize the vector distance of triples that satisfy the translation invariance assumption. Model Evaluation and Saving: During training, the model's performance is evaluated periodically, such as by using a link prediction task to measure the model's accuracy in predicting missing triples. After training is complete, the model parameters are saved for later use.

[0044] S104. Obtain and parse the user's current query requirements, perform knowledge graph traversal and sorting based on the initial entity nodes obtained from the parsing, and generate a recommendation list.

[0045] Specifically, first, obtain the user's current query requirements and perform preprocessing. The query requirements include text type and image type. Receive the user's input text query and perform basic natural language processing (NLP), such as word segmentation,词性标注, error correction (correct misspelled words, such as correcting "pump" to "pump"). For the image type, extract the key visual features in the picture based on the CNN-based image classification model to identify possible device types, parts, or fault states.

[0046] Use the language model to identify entity mentions in the query text, such as "XX device", "oil leakage", "screw", and use the context information in the knowledge graph to disambiguate and accurately map them to a unique graph node, obtaining an initial entity node. For example, map "screw" to the node "standard parts - fasteners - hexagon bolt", rather than "electronic components". The model further analyzes the user's deep intention. For example, it identifies that "oil leakage" is a "fault" entity, and "the screw has been tightened but it's still not working" implies that the user's intention is "repair" rather than "purchase a new device", and the potential relationship is "fault: oil leakage - may be caused by... -> seal failure", and "repair plan - requires -> new seal". Integrate the image recognition results (such as identifying "O-ring wear") with the text parsing results. If the image and text both point to the same entity or fault, increase its confidence; if they are complementary to each other, jointly enrich the query intention, and finally output a structured initial query subgraph. This subgraph contains a set of core entity nodes extracted from the user query (such as: [device XX], [fault: oil leakage], [part: screw]) and the relationship paths to be verified or explored between them (such as: [fault: oil leakage] - (may be caused by...) -> [seal failure]).

[0047] After obtaining the initial query subgraph, the system takes it as the starting point and performs intelligent traversal and reasoning in the knowledge graph to discover the relevant candidate product set. Each relation edge in the knowledge graph is no longer a fixed value, but a dynamic weight determined by a basic weight and a context adjustment factor. The basic weight is preset based on the confidence and importance of the relation in historical data. Context adjustment factor: The system accesses multiple context signals in real time and converts them into weight adjustment coefficients: User profile: The user's industry and technical level. Increase the weights of "substitute relation" and "upgrade relation" for high-skill users. Real-time business data: Inventory status (assign a high weight to the "in-stock" relation), supplier rating, delivery timeliness. Environmental parameters: Equipment operating environment (such as in a "high temperature" environment, increase the weight of the "high temperature resistance" attribute). Urgency level: For high-urgency work orders, prioritize the "local inventory" path.

[0048] Taking the generated initial query subgraph as the starting point, perform游走guided by multiple goals: Solution objective: To find "repair solutions" and "parts" that can resolve the current fault.

[0049] Compatibility goal: Ensure that the path follows relationships such as "compatible" and "for", and the final recommended parts match the user's equipment.

[0050] Optimization goal: Explore relationships such as "substitution" and "performance superiority" to find better and more cost-effective solutions.

[0051] Algorithms (such as Monte Carlo random walks) make multiple walks through the graph. Each time a new path is chosen, edges with higher dynamic weights are selected, thus increasing the probability of reaching a node that fits the current context.

[0052] Through multiple random walks, the algorithm converges from different paths to one or more candidate product nodes, forming a diverse initial candidate set. These products include not only direct matches, but also functional substitutes, upgrade alternatives, and even solution suites.

[0053] For each candidate product in the candidate set, a comprehensive recommendation score is calculated. This score is a weighted sum of scores from multiple ranking factors, including graph relevance score, business rule score, and collaborative signal score. Graph relevance score measures the strength of the association between the product and the initial query subgraph. It is determined by the sum of path weights and path lengths accessed during the inference process (shorter paths, higher weights, and higher scores). Business rule score incorporates real-time business logic. Examples include: inventory status (in stock +100 points), supplier rating (5-star supplier +50 points), procurement cost (below budget +30 points), and delivery timeliness (delivery within 24 hours +70 points). This ensures the recommendation results are business-feasible. Collaborative signal score, as a supplement, calculates the historical popularity of the product in similar repair scenarios among similar users, avoiding overlooking potential collective intelligence.

[0054] Dynamic weight adjustment: The system uses a reinforcement learning mechanism to dynamically adjust the weights of the above-mentioned factors based on the user's final feedback on the recommendation results (clicks, purchases, ignores), thereby achieving continuous optimization of the personalized ranking strategy.

[0055] During the random walk, one or more key formation paths are recorded for each selected candidate product, such as: [Your Device] - [Fault] -> [Seal Leak] - [Solution] -> [Replace Seal Ring] - [Compatible Model] -> [Recommended Product A]; Using a lightweight NLG model, the above path is transformed into an easy-to-understand recommendation reason, such as: "We recommend [Seal Ring A] because it is an original factory compatible replacement part for your [XX Device] experiencing a [Seal Leak] fault, and it is [in stock in our local warehouse] for fast delivery." The system calculates weight values ​​for the candidate set based on multi-dimensional ranking factors, and calculates the similarity between user vectors and product vectors based on users' historical interaction data. A comprehensive recommendation score is calculated based on the weight values ​​and the similarity scores using a weighted fusion method. The system then sorts the products according to this comprehensive recommendation score to generate a recommendation list. Simultaneously, the system records and outputs the complete reasoning path (e.g., "Recommend B because it is a substitute for A and meets your high-temperature environment requirements") as the recommendation reason, greatly improving the credibility and interpretability of the recommendation results.

[0056] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0057] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A knowledge graph-based method for recommending MRO industrial products, characterized in that, Includes the following steps: It collects structured and unstructured data, and transmits and integrates the collected multi-source data; The integrated data is cleaned and standardized, and knowledge is extracted based on the transformed data. Obtain the ontology and relationships in the MRO domain, and map them to the extracted entities and relationships to construct a knowledge graph; The system acquires and parses the user's current query request, performs knowledge graph traversal and sorting based on the initial entity nodes obtained from the parsing, and generates a recommendation list.

2. The knowledge graph-based MRO industrial product recommendation method as described in claim 1, characterized in that, The process involves collecting structured and unstructured data, and transmitting and integrating the collected multi-source data, including: Obtain structured and unstructured data, and determine the update frequency of the structured data; The unstructured data is transmitted using distributed object storage technology, with the corresponding transmission method matched to the update frequency of the structured data. The received multi-source data is encapsulated according to the set encapsulation format and cached in the data lake; After the data in the data lake is extracted, transformed, and loaded, it is stored in the data warehouse.

3. The knowledge graph-based MRO industrial product recommendation method as described in claim 2, characterized in that, The method further includes: Data quality checkpoints are set up at key nodes in the transmission channel to verify data quality.

4. The knowledge graph-based MRO industrial product recommendation method as described in claim 1, characterized in that, The integrated data undergoes data cleaning and standardization transformation, and knowledge extraction is performed based on the transformed data, including: The integrated structured data is cleaned and standardized, and the integrated unstructured data is processed into text data. Automatically identify and extract key entities from unstructured text using natural language processing techniques; A method combining rule-based matching and deep learning relation extraction models is used to extract triple relationships between entities and entity attributes from the structured and unstructured data.

5. The knowledge graph-based MRO industrial product recommendation method as described in claim 4, characterized in that, Automatically identify and extract key entities from the unstructured data using natural language processing techniques, including: Using a Transformer-based pre-trained language model as a foundation, and based on a specific sequence labeling strategy, key entities and their attribute sets are automatically identified and extracted from the unstructured data.

6. The knowledge graph-based MRO industrial product recommendation method as described in claim 4, characterized in that, A method combining rule-based matching and deep learning relation extraction models is used to extract triple relationships between entities and entity attributes from the structured and unstructured data, including: A method combining rule-based matching and deep learning relation extraction models is used to obtain multi-source evidence from the structured and unstructured data, and corresponding confidence scores are set. Link the key entities to the standard entity nodes of the knowledge graph; The multi-source evidence with multiple meanings is disambiguated and then stored in a temporary knowledge base.

7. The knowledge graph-based MRO industrial product recommendation method as described in claim 1, characterized in that, The system acquires and parses the user's current query request, performs knowledge graph traversal and sorting based on the initial entity nodes obtained from the parsing, and generates a recommendation list, including: Obtain the user's current query request and identify the query entities within the query request; By utilizing the contextual information in the knowledge graph, the query entity is disambiguated and mapped to a unique graph node to obtain the initial entity node; Based on the initial entity nodes, intelligent traversal and reasoning are performed to generate a candidate set; The candidate set is calculated based on multi-dimensional ranking factors, and a recommendation list is generated after dynamic weight adjustment.

8. The knowledge graph-based MRO industrial product recommendation method as described in claim 7, characterized in that, The multi-dimensional ranking factors include graph correlation score, business rule score, and collaborative signal score; wherein, the graph correlation score is the product of the sum of the weights of all inference paths between the candidate product node and the initial entity node and the path length penalty factor.

9. The knowledge graph-based MRO industrial product recommendation method as described in claim 7, characterized in that, The candidate set is calculated based on multi-dimensional ranking factors, and after dynamic weight adjustment, a recommendation list is generated, including: The weight values ​​corresponding to the candidate set are calculated based on the multi-dimensional ranking factors, and the similarity between the user vector and the product vector is calculated based on the user's historical interaction data. The weighted fusion method is used to calculate the comprehensive recommendation score corresponding to the weight value and the similarity, and the recommendation list is generated by sorting the comprehensive recommendation scores.

Images