A method for retrieving industrial lubricant formulation data

By constructing a knowledge graph and multi-dimensional vector representation model in the field of industrial lubricants, and combining it with an intent recognition model based on the Transformer architecture, the problems of single retrieval dimensions, insufficient semantic understanding, and lack of compliance verification in existing technologies have been solved, thus achieving efficient and accurate retrieval of lubricant formula data.

CN123115383APending Publication Date: 2026-11-17EDSON (ANHUI) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611323712.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for retrieving industrial lubricant formulation data suffer from problems such as limited retrieval dimensions, insufficient semantic understanding capabilities, inaccurate formulation similarity calculations, lack of domain-specific knowledge support, and missing compliance verification.

Method used

A knowledge graph for the industrial lubricant industry is constructed, employing a multi-dimensional vector representation model and an intent recognition model based on the Transformer architecture. This is combined with the knowledge graph for semantic expansion and multi-dimensional similarity calculation, and dynamic weight adjustment and automatic compliance verification are implemented.

Benefits of technology

It enables multi-dimensional information retrieval, accurate semantic understanding, domain-specific professional knowledge support, and automatic compliance verification, thereby improving the relevance and usability of search results and reducing user workload and error risk.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115383A_ABST
    Figure CN123115383A_ABST
Patent Text Reader

Abstract

This invention discloses a method for retrieving industrial lubricating oil formula data, belonging to the field of industrial lubricating oil formula information retrieval technology. The method includes: constructing a knowledge graph in the industrial lubricating oil formula domain; establishing a multi-dimensional vector representation model for formulas; receiving user search requests; semantically expanding the extracted search entities based on the knowledge graph; performing multi-dimensional matching search in the formula database; sorting the search results in descending order based on the matching score according to the application scenario corresponding to the search intent; establishing a search result feedback learning mechanism; and periodically updating the parameters of the intent recognition model, the weights of entity relationships in the knowledge graph, and the dimension weight mapping table based on the feedback learning data. This invention solves the technical problems of single search dimensions, insufficient semantic understanding ability, and inaccurate formula similarity calculation in existing technologies, significantly improving the accuracy and efficiency of industrial lubricating oil formula retrieval.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial lubricant formulation information retrieval technology, and in particular to a method for retrieving industrial lubricant formulation data. Background Technology

[0002] Industrial lubricants are crucial for the normal operation of industrial equipment, and their formulation design involves multiple dimensions, including base oil selection, additive compatibility, performance index optimization, and application scenario adaptation. With the continuous development of industrial technology, lubricant formulation data is growing explosively, and traditional formulation retrieval methods can no longer meet the needs for efficient and accurate retrieval.

[0003] In existing technologies, lubricant formulation data retrieval mainly suffers from the following problems: First, the search dimensions are limited. Most existing search systems only support searches based on formula name, number, or simple keywords, and cannot perform comprehensive searches from multiple dimensions such as ingredient composition, performance indicators, process parameters, and application scenarios. Users need to query relevant information in different databases separately, resulting in low search efficiency.

[0004] Second, the semantic understanding capability is insufficient. Existing systems lack the ability to deeply understand natural language retrieval requests and cannot accurately identify the implicit needs and constraints in the user's search intent. For example, when a user enters "anti-wear lubricating oil for high-temperature and high-pressure gearboxes," the existing system cannot automatically parse out the temperature range and pressure conditions corresponding to "high-temperature and high-pressure," the equipment type corresponding to "gearbox," and the specific performance index requirements corresponding to "anti-wear."

[0005] Third, the calculation of formula similarity is inaccurate. Existing similarity calculation methods mostly use simple text matching or single-dimensional comparison, without considering the complex relationships between lubricant formulas in terms of component functional equivalence, performance index correlation, process parameter compatibility, and application scenario adaptability, resulting in inaccurate relevance ranking of search results.

[0006] Fourth, there is a lack of domain-specific expertise. Existing general-purpose search systems are not optimized for the specialized knowledge of the industrial lubricant field, and cannot utilize domain-specific knowledge such as compatibility rules, synergistic effects, and antagonistic effects to intelligently expand and recommend search results.

[0007] Fifth, compliance verification is lacking. Existing search systems cannot automatically verify whether the formulas in the search results comply with relevant national standards, industry standards, or enterprise internal control standards. Users need to manually verify these standards, which increases their workload and the risk of errors.

[0008] Therefore, there is an urgent need for an industrial lubricant formulation data retrieval method that can integrate multi-dimensional information, possess semantic understanding capabilities, be supported by domain knowledge graphs, and support intelligent similarity calculation and automatic compliance verification. Summary of the Invention

[0009] The purpose of this invention is to provide a method for retrieving industrial lubricating oil formula data, so as to solve the technical problems in the prior art, such as single retrieval dimension, insufficient semantic understanding ability, inaccurate formula similarity calculation, lack of domain professional knowledge support, and lack of compliance verification.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for retrieving industrial lubricant formulation data includes the following steps: S1: Constructing a knowledge graph for the field of industrial lubricant formulations; S2: Establish a multi-dimensional vector representation model for the formula, representing each lubricating oil formula as a component dimension vector, performance dimension vector, process dimension vector, and application dimension vector; S3: Receive user search requests, parse user search intent through an intent recognition model based on Transformer architecture, and extract search entities and search constraints; S4: Based on the knowledge graph, semantic expansion is performed on the extracted search entities, and expanded search conditions are generated by utilizing the synergistic effect relationship and substitution relationship between entities to construct the expanded search query graph; S5: Perform multi-dimensional matching search in the formula database, and calculate the similarity of candidate formulas with search conditions in terms of component dimension, performance dimension, process dimension, and application dimension respectively. S6: Based on the application scenario corresponding to the search intent, dynamically adjust the weight coefficients of the similarity of each dimension, and comprehensively calculate the matching score of each candidate recipe. The weight coefficients are determined by an adaptive weight allocation algorithm based on the importance of each dimension in the application scenario. S7: Sort the search results in descending order based on the matching score, automatically verify the compliance of each formula after sorting, remove formulas that do not meet the relevant standards and specifications, and output the final search results; S8: Establish a search result feedback learning mechanism to record users' browsing behavior, collection behavior, download behavior and evaluation information of search results. Based on the feedback learning data, regularly update the parameters of the intent recognition model, the weights of entity relationships in the knowledge graph and the dimension weight mapping table. When it is detected that a user has searched for similar topics multiple times, actively push related recipes and trend recipes. The trend recipes are dynamically generated based on recent industry technology developments and standard updates.

[0011] Preferably, the knowledge graph in step S1 includes base oil entities, additive entities, performance index entities, application scenario entities, standard specification entities, as well as compatibility relationships, synergistic effect relationships, antagonistic effect relationships, and substitution relationships between entities.

[0012] Preferably, the component dimension vector in step S2 is constructed based on the type and mass percentage of each component in the formulation; the performance dimension vector is constructed based on the kinematic viscosity, viscosity index, flash point, pour point, acid value, and anti-wear performance index of the formulation; the process dimension vector is constructed based on the preparation temperature, stirring speed, blending time, and filtration accuracy parameters of the formulation; and the application dimension vector is constructed based on the applicable equipment type, operating temperature range, load conditions, and rotation speed range parameters of the formulation.

[0013] Preferably, the construction of the knowledge graph in the field of industrial lubricant formulations in step S1 includes: S11: Extract entities from lubricating oil formulation databases, standard documents, technical manuals and patent documents. The entities include base oil type, additive type, performance index name, application scenario category and standard specification number. S12: Define the types of relationships between entities, including: compatibility relationships between base oils and additives, synergistic effects between additives, antagonistic effects between additives, limiting relationships between performance indicators and standard specifications, demand relationships between application scenarios and performance indicators, and substitution relationships between base oils. S13: Employ graph neural networks to learn the embeddings of entities and relations, generating entity vector representations and relation vector representations; S14: Based on the results of embedding learning, construct a knowledge graph containing entity nodes, relation edges and attribute features, and establish a dynamic update mechanism for the knowledge graph. When new recipe data is added or new compatibility rules are discovered, the entity relations and attribute weights in the graph are automatically updated.

[0014] Preferably, the establishment of the multi-dimensional vector representation model of the formula in step S2 includes: S21: Normalize the component dimension vector, convert the mass percentage of each component into a relative content vector, and introduce component functional labels, which include extreme pressure anti-wear agent, antioxidant, anti-corrosion agent, defoamer, and pour point depressant. S22: Standardize the performance dimension vector, map each performance index value to a unified dimension space, and introduce performance level labels, which are determined according to the industrial lubricant classification standard. S23: Discretize the process dimension vector, divide the continuous process parameters into standard process intervals, and introduce a process complexity index. S24: Encode the application dimension vector, convert the equipment type and operating conditions into a multi-dimensional one-hot encoded vector, and introduce environmental adaptability labels; S25: Establish a correlation mapping matrix between the four dimensional vectors to represent the intrinsic correlation between the component dimension and the performance dimension, and between the performance dimension and the application dimension.

[0015] Preferably, step S3, which involves parsing the user's retrieval intent using an intent recognition model based on the Transformer architecture, includes: S31: Perform word segmentation and part-of-speech tagging on user search requests, and identify numerical constraints, range constraints, and comparison constraints in the search requests; S32: Input the word segmentation results into the pre-trained domain language model, which is incrementally trained using technical literature in the field of lubricating oil based on a general corpus; S33: Capture long-distance dependencies in retrieval requests and identify implicit retrieval intents through a multi-head self-attention mechanism; S34: Output a structured search intent representation, including the search target type, core search entities, list of constraints, and priority order.

[0016] Preferably, the semantic expansion of the retrieved entities based on the knowledge graph in step S4 includes: S41: Map the extracted retrieval entities to the corresponding nodes in the knowledge graph to determine the position and relationship of each entity in the graph; S42: For categorized search entities, query related additives with synergistic effects along the compatibility relationship and query substitutable base oils or additives along the substitution relationship to generate an extended set of ingredients. S43: For performance-related search entities, query the application scenario's required range of performance indicators along the demand relationship, and query the standard specifications' prescribed values ​​of performance indicators along the limitation relationship to generate a performance extension range; S44: For application-type retrieval entities, query backwards along the demand relationship edge to satisfy the combination of performance indicators required for the application scenario, and generate an application-related performance set; S45: Merge the original search conditions with the extended search conditions to construct an extended search query graph, which includes the original search nodes, the extended search nodes, and the logical relationships between the nodes.

[0017] Preferably, the calculation of the multi-dimensional similarity between the candidate recipe and the search criteria in step S5 includes: S51: Calculate component-dimensional similarity using an improved cosine similarity algorithm, considering component type overlap, component content deviation, and component functional equivalence. The formula is:

[0018] in, The component dimension vector of the search criteria. Let be the component dimension vector of the candidate formulation. For functional equivalence functions, This is a content deviation function. , , For preset weighting coefficients and ; S52: Calculate the similarity of the performance dimension using a weighted Euclidean distance algorithm, assigning different weights to each performance indicator based on its importance in the application scenario; S53: Calculate the process dimension similarity and use the interval matching degree algorithm to determine whether the process parameters of the candidate formula fall within the process interval specified by the search conditions; S54: Calculate the application dimension similarity using a semantic matching degree algorithm. Based on the one-hot encoding of the application dimension vector, calculate the device type matching degree and operating condition compatibility.

[0019] Preferably, the step S6 of dynamically adjusting the weight coefficients of the similarity of each dimension according to the application scenario includes: S61: Establish an application scenario-dimensional weight mapping table. The mapping table is constructed based on historical retrieval logs and expert experience, and represents the importance weight of each dimension under different application scenarios. S62: When the search intent explicitly specifies the application scenario, directly query the corresponding weight coefficient from the mapping table; S63: When the search intent does not explicitly specify the application scenario, the application scenario category is inferred through the classification model based on the keywords and constraints in the search request, and then the corresponding weight coefficient is queried; S64: Introduce a user feedback mechanism to dynamically optimize the weight coefficients based on user click behavior and evaluation information of search results, and achieve adaptive adjustment of the weight coefficients; S65: Overall Match Score, calculated using the following formula: ,in , , , These are dynamic weighting coefficients for four dimensions: composition, performance, process, and application. .

[0020] Preferably, the automatic verification of the compliance of each formula after sorting in step S7 includes: S71: Extract the standard and specification requirements corresponding to each formula in the search results, including national standards, industry standards and enterprise internal control standards; S72: Compare the measured or predicted values ​​of each performance indicator of the formula with the limit values ​​in the standard specifications item by item; S73: For formulations that exceed the limits specified in the standard specifications, mark them as non-compliant and record the non-compliance items; S74: For formulations marked as non-compliant, recommend alternative components or adjustment schemes based on substitution relationships in the knowledge graph to make them meet compliance requirements; S75: Output a list of search results that includes compliance indicators, explanations of non-compliance items, and adjustment suggestions.

[0021] This invention also discloses an industrial lubricant formulation data retrieval system, comprising: The knowledge graph construction module is used to build a knowledge graph in the field of industrial lubricant formulation. The vector representation module is used to establish a multi-dimensional vector representation model of the recipe; The intent recognition module is used to receive user search requests and parse the search intent; The semantic extension module is used to semantically extend the retrieved entities based on the knowledge graph; The similarity calculation module is used to calculate the multi-dimensional similarity between candidate recipes and search criteria; The dynamic sorting module is used to dynamically adjust weights and calculate matching scores based on the application scenario. The compliance verification module is used to automatically verify the compliance of search results; The feedback learning module is used to record user behavior and optimize system parameters.

[0022] The beneficial effects of this invention are as follows: 1. This invention constructs a knowledge graph specifically for the industrial lubricating oil field, which includes unique entity types and relationship types (such as compatibility relationships, synergistic effect relationships, antagonistic effect relationships, and substitution relationships), and can provide in-depth domain professional knowledge support for formula retrieval, realizing intelligent semantic expansion and recommendation based on professional knowledge.

[0023] 2. This invention proposes a multi-dimensional vector representation model for formulas, which vectorizes lubricant formulas from four dimensions: composition, performance, process, and application. This breaks through the limitations of traditional single-dimensional retrieval and achieves more comprehensive and accurate formula similarity calculation.

[0024] 3. This invention adopts a domain intent recognition model based on the Transformer architecture, which can deeply understand the implicit intent and constraints in users' natural language search requests, and transform unstructured search requests into structured search intent representations, significantly improving semantic understanding capabilities.

[0025] 4. This invention proposes a dynamic weight adjustment mechanism, which automatically adjusts the weight coefficients of similarity in each dimension according to different application scenarios, making the search results more in line with actual application needs and improving the relevance and practicality of the search results.

[0026] 5. This invention integrates an automatic compliance verification function, which can automatically compare the search results with the requirements of relevant standards and specifications, mark non-compliant formulas and provide adjustment suggestions, thereby reducing the user's workload and error risk.

[0027] 6. This invention establishes a feedback learning mechanism that can continuously optimize system parameters based on user behavior, thereby achieving self-improvement in retrieval results and exhibiting good adaptability and scalability. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the steps of an industrial lubricating oil formulation data retrieval method proposed in an embodiment of the present invention. Figure 2 This is an architecture diagram of an industrial lubricating oil formula data retrieval system proposed in Embodiment 2 of the present invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] Example 1 like Figure 1 As shown, this embodiment provides a method for retrieving industrial lubricant formulation data, including the following steps: Step S1: Construct a knowledge graph for the industrial lubricant formulation domain. The knowledge graph for the industrial lubricant formulation domain constructed in this embodiment includes the following entity types: Base oil entities include mineral base oils (such as Group I, Group II, and Group III base oils), synthetic base oils (such as PAO, ester oils, polyethers, and silicone oils), and bio-based base oils. Each base oil entity contains viscosity grade, viscosity index, flash point, pour point, and oxidation stability properties.

[0031] Additive entities include detergents and dispersants, antioxidants and corrosion inhibitors, extreme pressure anti-wear agents, friction modifiers, antifoaming agents, pour point depressants, viscosity index improvers, etc. Each additive entity contains chemical composition, recommended dosage range, functional mechanism, and compatibility characteristics.

[0032] Physical performance indicators include kinematic viscosity (40℃, 100℃), viscosity index, flash point (open / closed cup), pour point, acid value, carbon residue, ash content, abrasion resistance (four-ball test results), oxidation stability (rotating oxygen bomb value), and demulsibility.

[0033] Application scenario entities include gear transmission, hydraulic system, steam turbine, compressor, bearing lubrication, metal processing, heat treatment, etc. Each application scenario entity contains typical operating parameters (temperature, pressure, speed, load).

[0034] Standard and specification entities include GB standards (such as GB 5903, GB 11120), industry standards (such as SH / T, NB / SH / T), international standards (such as ISO, ASTM, DIN) and enterprise internal control standards. Each standard and specification entity contains applicable lubricant types, performance index limits, and test methods.

[0035] Specifically, the types of relationships between the aforementioned entities include: Compatibility relationship: Characterizes the compatibility between base oil and additives, such as "a certain type of base oil can be blended with a certain type of additive". The relationship attributes include compatibility level (excellent, good, average, poor) and compatibility conditions (temperature, time requirements).

[0036] Synergistic effect relationship: Characterizes the positive interaction between additives, such as "antioxidant A and anti-wear agent B have a synergistic effect", and the relationship attributes include the strength of the synergistic effect and the mechanism of action.

[0037] Antagonistic effect relationship: Characterizes the negative interaction between additives, such as "a certain defoamer and a certain dispersant have an antagonistic effect", and the relationship attributes include the strength of the antagonistic effect and avoidance measures.

[0038] Substitution relationship: Characterizes the substitutability between base oils or additives, such as "PAO 6 can replace PAO 4 to improve low-temperature performance". Relationship attributes include substitution conditions and performance change prediction.

[0039] Limitation relationship: Characterizes the performance requirements of standards and specifications, such as "GB 5903-2011 specifies that the kinematic viscosity range of L-CKD 150 is 135-165 mmHg". 2 / s (40℃)".

[0040] Demand Relationship: Characterizes the performance requirements of the application scenario, such as "heavy-duty gear transmission requires extreme pressure anti-wear performance PB value ≥ 800N".

[0041] In the knowledge graph construction process, graph neural networks (GNNs) are used to embed entities and relations, generating 128-dimensional entity vector representations and 64-dimensional relation vector representations. The embedding learning adopts an improved algorithm based on the TransE model, which learns high-quality vector representations by minimizing the energy function of triples (head entity, relation, and tail entity) in the knowledge graph.

[0042] A dynamic update mechanism for the knowledge graph is established. When the system receives new formula data or literature, it automatically extracts entities and identifies relationships to update the graph content. The update frequency is set to a combination of weekly batch updates and real-time incremental updates.

[0043] Step S2: Establish a multi-dimensional vector representation model for the formulation. Each lubricant formulation is represented as a four-dimensional vector: Component dimension vector: The number of dimensions is the total number of predefined component types. For each formulation, each element of the component dimension vector represents the mass percentage of the corresponding component (components not included are marked as 0). Additionally, a functional tag vector is attached to each component. These functional tags are uniquely coded, and the tag categories include: extreme pressure anti-wear agents, antioxidants, corrosion inhibitors, defoamers, pour point depressants, viscosity index improvers, detergent-dispersants, and friction modifiers—a total of eight categories.

[0044] Performance dimension vector: The number of dimensions is the total number of predefined performance metrics. The study included eight core indicators: kinematic viscosity (40℃), kinematic viscosity (100℃), viscosity index, flash point, pour point, acid value, anti-wear performance (wear scar diameter), and oxidation stability (rotary bomb oxidation value). Each indicator value was Z-score standardized and mapped to a standard normal distribution space with a mean of 0 and a standard deviation of 1.

[0045] Process dimension vector: The number of dimensions is the total number of predefined process parameters. The process parameters include six parameters: preparation temperature (°C), stirring speed (rpm), mixing time (min), filtration accuracy (μm), dehydration vacuum degree (Pa), and cooling rate (°C / min). The process parameters adopt interval discretization coding, dividing continuous parameter values ​​into three intervals: low, medium, and high, and representing them using 3D unique thermal coding.

[0046] Application dimension vector: The number of dimensions is the total number of predefined application parameters. It includes 17 dimensions, including applicable equipment types (6 categories such as gearboxes, hydraulic systems, steam turbines, compressors, bearings, and metal processing equipment, with 6-dimensional unique thermal coding), operating temperature range (low temperature <-20℃, normal temperature -20-60℃, medium temperature 60-120℃, high temperature >120℃, with 4-dimensional unique thermal coding), load conditions (light load, medium load, heavy load, and extremely heavy load, with 4-dimensional unique thermal coding), and speed range (low speed <500rpm, medium speed 500-3000rpm, and high speed >3000rpm, with 3-dimensional unique thermal coding).

[0047] Establish the association mapping matrix between the four-dimensional vectors: Ingredient-performance correlation matrix Dimensions are Matrix elements Indicates the first The first component on the first The influence coefficients of each performance indicator are represented by positive values ​​indicating positive influence and negative values ​​indicating negative influence, with the absolute value indicating the intensity of the influence. This matrix is ​​obtained through multiple regression analysis based on historical formula data and experimental results.

[0048] Performance-Application Correlation Matrix Dimensions are Matrix elements Indicates the first The performance index for the first The importance weights of each application parameter are calculated. This matrix is ​​constructed based on expert experience and application scenario requirements analysis.

[0049] Step S3: Receive the user's search request and parse the search intent. This embodiment uses an intent recognition model based on the Transformer architecture. The model input is the user's search request text, which undergoes the following processing flow: Word segmentation and tagging: A domain-customized word segmentation dictionary is used to segment and tag the search requests. The domain-customized word segmentation dictionary contains more than 5,000 professional terms in the lubricating oil field, covering base oil names, additive names, performance index names, equipment type names, standard numbers, etc. It also recognizes numerical constraints (e.g., viscosity greater than 100), range constraints (e.g., temperature between 80 and 120 degrees Celsius), and comparison constraints (e.g., flash point not lower than 200 degrees Celsius).

[0050] Domain-specific language model encoding: The word segmentation results are input into a pre-trained domain-specific BERT model. Based on the general Chinese BERT-base model, this model undergoes incremental training using 500,000 sentences from lubricant-related technical documents, standard texts, and patent documents, with 5 training epochs and a learning rate of 2e-5. Incremental training improves the model's semantic understanding accuracy of domain-specific terms by approximately 23%.

[0051] Multi-head self-attention encoding: A 12-layer Transformer encoder is used to capture long-distance dependencies in retrieval requests using a multi-head self-attention mechanism. For example, in the retrieval request "finding a synthetic gear oil formulation for wind turbine gearboxes that can maintain fluidity at -30°C", the model can establish semantic associations between "wind turbine gearbox" and "synthetic gear oil", and between "-30°C low temperature" and "maintain fluidity".

[0052] Intent Classification and Entity Extraction: The model output layer includes an intent classifier and an entity extractor. The intent classifier categorizes search intents into four types: exact search, similar search, recommended search, and comparative search. The entity extractor uses the BIO annotation scheme to extract search entities and constraints, outputting a structured search intent representation in the following format: { "intent_type": "Similar Search", "target": "recipe", "core_entities": ["synthetic gear oil", "wind turbine gearbox"], "constraints": [ {"type": "performance", "indicator": "pour point", "operator": "<=", "value": "-30℃"}, {"type": "Ingredients", "component": "Base oils", "category": "Synthetic base oils"} ], "priority": ["Low-temperature performance", "Extreme pressure anti-wear performance"] } Step S4: Semantic expansion based on the knowledge graph. The semantic expansion process is as follows: Entity mapping: Mapping the extracted retrieval entities to the corresponding nodes in the knowledge graph. For example, "synthetic gear oil" is mapped to the "gear oil - synthetic" node, and "wind turbine gearbox" is mapped to the "application scenario - wind turbine gear transmission" node.

[0053] Composition expansion: For the "synthetic base oil" entity, query the substitutable base oil types along the substitution relationship to obtain the expanded set {PAO, ester oil, polyether}; query the synergistic combination with commonly used gear oil additives along the compatibility relationship to obtain the expanded set {extreme pressure anti-wear agent - sulfur-phosphorus type, antioxidant - phenol-amine complex}.

[0054] Performance expansion: For the "pour point <= -30℃" constraint, consult relevant standards (such as GB 5903) for the pour point requirements of low-temperature gear oils along the limit relationship, and confirm that the expansion range is pour point ≤ -33℃; consult the demand relationship for the demand indicators of low-temperature start-up performance in the "wind power gearbox" application scenario, including related indicators such as low-temperature dynamic viscosity and low-temperature pumpability.

[0055] Application Extension: For the "wind turbine gearbox" application scenario, the complete set of performance indicators required for the scenario is queried in reverse along the demand relationship edge, including extreme pressure performance, anti-micropitting performance, anti-emulsification performance, filtration performance, etc., to generate an application-related performance set.

[0056] Expanded Query Graph Construction: The original search nodes (blue nodes) and expanded search nodes (green nodes) are merged to construct an expanded query graph. The relationships between nodes in the graph include AND relationships (both must be satisfied) and OR relationships (either one is sufficient). For example, the original search condition "base oil is synthetic" and the expanded condition "base oil is PAO or ester oil" have an OR relationship; the original search condition "pour point ≤ -30℃" and the expanded condition "low-temperature dynamic viscosity ≤ 150000 mPa·s" have an AND relationship.

[0057] Step S5: Perform multi-dimensional matching retrieval and calculate multi-dimensional similarity.

[0058] Component-dimensional similarity calculation: An improved cosine similarity algorithm is used, and the formula is as follows:

[0059] in, For traditional cosine similarity, calculate the cosine of the angle between the two ingredient vectors of the formulation; The functional equivalence function is used to determine the degree of overlap between two formulations at the functional label level. For example, if formulation A contains sulfur and phosphorus-type extreme pressure agent and formulation B contains borate-type extreme pressure agent, and both of their functional labels are "extreme pressure anti-wear agent", then the functional equivalence is 1. This is the content deviation function; calculate the normalized average of the content differences of the corresponding components.

[0060] Performance dimension similarity calculation: The weighted Euclidean distance algorithm is used, and the formula is as follows:

[0061] in, For the first The weights of each performance indicator are dynamically determined based on the application scenario. For wind turbine gearbox applications, the weights are: extreme pressure performance 0.25, low temperature performance 0.20, oxidation resistance 0.15, demulsification performance 0.15, viscosity characteristics 0.15, and other performance indicators 0.10.

[0062] Process dimension similarity calculation: An interval matching algorithm is used, and the formula is:

[0063] The Match function determines whether the process parameters of the candidate formulation fall within the process range specified by the search conditions. If they do, the matching degree is 1; otherwise, it decreases linearly according to the degree of deviation.

[0064] Application-dimensional similarity calculation: A semantic matching algorithm is used, and the formula is as follows:

[0065] in, Device type match score (1 for exact match, 0.5 for same category match, 0 for different category match); The compatibility of operating conditions is calculated based on the overlap of operating temperature range, the compatibility of load conditions, and the matching degree of speed range.

[0066] Step S6: Dynamically adjust the weights and calculate the overall matching score.

[0067] The application scenario-dimension weight mapping table established in this embodiment is shown in the following table:

[0068] When the search intent explicitly specifies an application scenario (e.g., wind turbine gearbox), the weight coefficients can be obtained directly by querying the mapping table. .

[0069] When the search intent does not explicitly specify an application scenario, the application scenario category is inferred based on the keywords in the search request using a pre-trained application scenario classification model (fine-tuned based on BERT). The classification model was trained on 10,000 labeled search requests, achieving an accuracy of 92.3%.

[0070] Introducing a user feedback mechanism: Recording the number of clicks, favorites, downloads, and ratings of user search results. When users frequently click on recipes with high ingredient similarity for a certain type of search request, the ingredient weight of that type of request is automatically increased. When users frequently click on high-performance similarity recipes, the performance weight is automatically increased. The weight adjustment uses exponential smoothing, and the update formula is as follows: .

[0071] Overall matching score calculation:

[0072] Step S7: Result sorting and compliance verification.

[0073] The search result processing flow is as follows: Results sorting: All candidate recipes were sorted in descending order based on the comprehensive matching score, and the top 100 were taken as the preliminary search results.

[0074] Standard and specification extraction: For each search result, the corresponding standard and specification are automatically associated based on the formula type and application scenario. For example, gear oil formulas are associated with GB 5903 "Industrial Closed Gear Oils", and hydraulic oil formulas are associated with GB 11118.1 "Hydraulic Oils (L-HL, L-HM, L-HV, L-HS, L-HG)".

[0075] Compliance verification item by item: The performance indicators of the formulation are compared with the limits specified in the standard specifications. Taking L-CKD220 gear oil as an example, the comparison items include: kinematic viscosity (40℃) 198-242 mm. 2 / s, viscosity index ≥90, flash point (open cup) ≥200℃, pour point ≤-9℃, mechanical impurities ≤0.02%, moisture ≤trace, copper strip corrosion (100℃, 3h) ≤1 grade, no rust in liquid phase corrosion test, demulsibility (82℃) ≤30min, Timken OK load ≥200N (or PB value of four-ball machine ≥800N).

[0076] Non-compliance labeling and adjustment suggestions: For formulations exceeding the limits of standard specifications, mark them as "non-compliant" and record the specific non-compliance items. For example, if a candidate formulation has a pour point of -6℃, exceeding the L-CKD 220 requirement of ≤-9℃, then mark it as "pour point non-compliant". Based on substitution relationships in the knowledge graph, the system recommends the following adjustment: It is recommended to replace the base oil from Group II mineral oil (pour point -12℃) to Group III mineral oil (pour point -15℃) or PAO (pour point -40℃), and add 0.5-1.0% of a pour point depressant (such as polymethyl methacrylate).

[0077] Output results: Output the final list of search results. Each result includes: recipe number, recipe name, matching score, similarity in each dimension, compliance indicator, explanation of non-compliance items (if any), adjustment suggestions (if any), and summary of standard and specification compliance.

[0078] S8: Establish a feedback learning mechanism for search results.

[0079] Record user browsing, collection, download and evaluation information of search results. Based on feedback learning data, regularly update the parameters of the intent recognition model, the weights of entity relationships in the knowledge graph and the dimension weight mapping table. When it is detected that a user has searched for similar topics multiple times, proactively push related recipes and trend recipes. Trend recipes are dynamically generated based on recent industry technology developments and standard updates.

[0080] Example 2 like Figure 2 As shown: This embodiment provides an industrial lubricating oil formula data retrieval system, including: Knowledge graph construction module: responsible for extracting entities and relationships from multi-source data, constructing a knowledge graph in the field of industrial lubricant formulation, and maintaining the dynamic updates of the graph.

[0081] Vector Representation Module: Responsible for converting each recipe in the recipe database into a multi-dimensional vector representation and maintaining the association mapping matrix between the dimensional vectors.

[0082] Intent recognition module: responsible for receiving user search requests, parsing the search intent through the domain BERT model, and outputting a structured search intent representation.

[0083] Semantic extension module: Responsible for semantic extension of retrieved entities based on knowledge graph, generating extended search conditions, and constructing extended search query graph.

[0084] Similarity calculation module: responsible for calculating the similarity between candidate formulas and search criteria in four dimensions: ingredients, performance, process, and application.

[0085] Dynamic sorting module: Responsible for dynamically adjusting the weights of each dimension according to the application scenario, calculating the comprehensive matching score, and sorting the search results.

[0086] Compliance verification module: Responsible for automatically verifying the standard and specification compliance of search results, marking non-compliant items and providing adjustment suggestions.

[0087] Feedback and Learning Module: Responsible for recording user behavior data, regularly updating system parameters, and continuously optimizing search results.

[0088] User interaction module: Provides a search interface, a results display interface, and a feedback collection interface, and supports natural language search input and multi-dimensional filtering conditions.

[0089] The system is deployed using a microservice architecture, with modules communicating via RESTful APIs. Knowledge graph storage uses Neo4j graph database, recipe vector data storage uses Milvus vector database, raw recipe data storage uses MySQL relational database, and the caching layer uses Redis. The system supports concurrent retrieval request processing, with a single-node QPS exceeding 500 and an average retrieval response time of less than 200ms.

[0090] Example 3 This embodiment uses a real-world retrieval scenario as an example to illustrate the application effect of the present invention.

[0091] Search scenario: A lubricant R&D engineer needs to develop a low-temperature synthetic gear oil for offshore wind turbine gearboxes, requiring a pour point no higher than -30℃, while also possessing good extreme pressure anti-wear properties and anti-micropitting properties.

[0092] User search request: Low-temperature synthetic gear oil for offshore wind turbine gearboxes, pour point below -30 degrees Celsius, with good extreme pressure performance.

[0093] Traditional search systems can only match formulas containing keywords such as wind power, gear oil, and synthetic materials. The results are sorted by name matching degree and cannot identify specific requirements such as "pour point below -30 degrees" and "good extreme pressure performance". The results contain a large number of formulas that do not meet the low temperature and extreme pressure requirements, and users need to filter them one by one.

[0094] The system of this invention processes: Intent recognition: The search intent type is parsed as "recommended search", the core entities are "synthetic gear oil" and "offshore wind power gearbox", the constraints are "pour point ≤ -30℃" and "excellent extreme pressure performance", and the priorities are "low temperature performance", "extreme pressure performance" and "corrosion resistance".

[0095] Semantic expansion: Based on the knowledge graph, the following can be expanded: the base oil can be PAO, ester oil or polyether; the additives must include sulfur-phosphorus type extreme pressure anti-wear agent and anti-micropitting agent; the associated performance indicators include low temperature dynamic viscosity, Timken OK load, and FZG failure load level.

[0096] Multi-dimensional search: Candidate recipes that meet the expanded criteria are retrieved from the recipe database, and the similarity across each dimension is calculated. The similarity across each dimension for a candidate recipe (Recipe A) is as follows: , , , .

[0097] Dynamic weights: Based on the application scenario of "offshore wind turbine gearbox", the weight mapping table is queried to obtain... (The marine environment places higher demands on performance, hence the increased weighting of performance.) Overall Score .

[0098] Compliance Verification: Automatically linked to GB 5903 standard, verifying various indicators of formulation A. Results show that kinematic viscosity, viscosity index, flash point, and extreme pressure performance all meet the L-CKD standard, but the pour point is -28℃, failing to meet the ≤-30℃ requirement, and is marked as "pour point non-compliant." The system recommends the following adjustment: "Increase the proportion of PAO 40 in the base oil from 60% to 70%, and add 0.8% polymethyl methacrylate pour point depressant; the pour point is expected to drop to -35℃."

[0099] Search results output: Sorted by overall score, the top 10 results include 8 compliant recipes and 2 non-compliant recipes (with adjustment suggestions). Users can click on the 3rd result to view details and save the 1st and 5th results. The system records user behavior and automatically increases the weight of the performance dimension in this scenario in subsequent similar searches.

[0100] Comparative results: The relevance of the results returned by the system of this invention is improved by about 65% compared with the traditional system, the user screening time is shortened by about 80%, and the accuracy rate of formula compliance verification reaches more than 98%.

[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for retrieving industrial lubricant formulation data, characterized in that, Includes the following steps: S1: Constructing a knowledge graph for the field of industrial lubricant formulations; S2: Establish a multi-dimensional vector representation model for the formula, representing each lubricating oil formula as a component dimension vector, performance dimension vector, process dimension vector, and application dimension vector; S3: Receive user search requests, parse user search intent through an intent recognition model based on Transformer architecture, and extract search entities and search constraints; S4: Based on the knowledge graph, semantic expansion is performed on the extracted search entities, and expanded search conditions are generated by utilizing the synergistic effect relationship and substitution relationship between entities to construct the expanded search query graph; S5: Perform multi-dimensional matching search in the formula database, and calculate the similarity of candidate formulas with search conditions in terms of component dimension, performance dimension, process dimension, and application dimension respectively. S6: Based on the application scenario corresponding to the search intent, dynamically adjust the weight coefficients of the similarity of each dimension, and comprehensively calculate the matching score of each candidate recipe. The weight coefficients are determined by an adaptive weight allocation algorithm based on the importance of each dimension in the application scenario. S7: Sort the search results in descending order based on the matching score, automatically verify the compliance of each formula after sorting, remove formulas that do not meet the relevant standards and specifications, and output the final search results; S8: Establish a search result feedback learning mechanism to record users' browsing behavior, collection behavior, download behavior and evaluation information of search results. Based on the feedback learning data, regularly update the parameters of the intent recognition model, the weights of entity relationships in the knowledge graph and the dimension weight mapping table. When it is detected that a user has searched for similar topics multiple times, actively push related recipes and trend recipes. The trend recipes are dynamically generated based on recent industry technology developments and standard updates.

2. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, The knowledge graph mentioned in step S1 includes base oil entities, additive entities, performance index entities, application scenario entities, standard specification entities, as well as compatibility relationships, synergistic effect relationships, antagonistic effect relationships, and substitution relationships between entities.

3. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, The component dimension vector in step S2 is constructed based on the type and mass percentage of each component in the formulation. The performance dimension vector is constructed based on the kinematic viscosity, viscosity index, flash point, pour point, acid value, and anti-wear performance index of the formulation. The process dimension vector is constructed based on the preparation temperature, stirring speed, blending time, and filtration accuracy parameters of the formulation. The application dimension vector is constructed based on the applicable equipment type, operating temperature range, load conditions, and rotation speed range parameters of the formulation.

4. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, The construction of the knowledge graph in the field of industrial lubricant formulations mentioned in step S1 includes: S11: Extract entities from lubricating oil formulation databases, standard documents, technical manuals and patent documents. The entities include base oil type, additive type, performance index name, application scenario category and standard specification number. S12: Define the types of relationships between entities, including: compatibility relationships between base oils and additives, synergistic effects between additives, antagonistic effects between additives, limiting relationships between performance indicators and standard specifications, demand relationships between application scenarios and performance indicators, and substitution relationships between base oils. S13: Employ graph neural networks to learn the embeddings of entities and relations, generating entity vector representations and relation vector representations; S14: Based on the results of embedding learning, construct a knowledge graph containing entity nodes, relation edges and attribute features, and establish a dynamic update mechanism for the knowledge graph. When new recipe data is added or new compatibility rules are discovered, the entity relations and attribute weights in the graph are automatically updated.

5. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, Step S2, which involves establishing a multi-dimensional vector representation model of the formula, includes: S21: Normalize the component dimension vector, convert the mass percentage of each component into a relative content vector, and introduce component functional labels, which include extreme pressure anti-wear agent, antioxidant, anti-corrosion agent, defoamer, and pour point depressant. S22: Standardize the performance dimension vector, map each performance index value to a unified dimension space, and introduce performance level labels, which are determined according to the industrial lubricant classification standard. S23: Discretize the process dimension vector, divide the continuous process parameters into standard process intervals, and introduce a process complexity index. S24: Encode the application dimension vector, convert the equipment type and operating conditions into a multi-dimensional one-hot encoded vector, and introduce environmental adaptability labels; S25: Establish a correlation mapping matrix between the four dimensional vectors to represent the intrinsic correlation between the component dimension and the performance dimension, and between the performance dimension and the application dimension.

6. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, Step S3, which involves parsing the user's retrieval intent using an intent recognition model based on the Transformer architecture, includes: S31: Perform word segmentation and part-of-speech tagging on user search requests, and identify numerical constraints, range constraints, and comparison constraints in the search requests; S32: Input the word segmentation results into the pre-trained domain language model, which is incrementally trained using technical literature in the field of lubricating oil based on a general corpus; S33: Capture long-distance dependencies in retrieval requests and identify implicit retrieval intents through a multi-head self-attention mechanism; S34: Output a structured search intent representation, including the search target type, core search entities, list of constraints, and priority order.

7. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, Step S4, which involves semantic expansion of the retrieved entities based on the knowledge graph, includes: S41: Map the extracted retrieval entities to the corresponding nodes in the knowledge graph to determine the position and relationship of each entity in the graph; S42: For categorized search entities, query related additives with synergistic effects along the compatibility relationship and query substitutable base oils or additives along the substitution relationship to generate an extended set of ingredients. S43: For performance-related search entities, query the application scenario's required range of performance indicators along the demand relationship, and query the standard specifications' prescribed values ​​of performance indicators along the limitation relationship to generate a performance extension range; S44: For application-type retrieval entities, query backwards along the demand relationship edge to satisfy the combination of performance indicators required for the application scenario, and generate an application-related performance set; S45: Merge the original search conditions with the extended search conditions to construct an extended search query graph, which includes the original search nodes, the extended search nodes, and the logical relationships between the nodes.

8. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, Step S5, which involves calculating the multi-dimensional similarity between candidate recipes and search criteria, includes: S51: Calculate the component dimension similarity using an improved cosine similarity algorithm, taking into account the overlap of component types, the deviation of component content, and the functional equivalence of components. S52: Calculate the similarity of the performance dimension using a weighted Euclidean distance algorithm, assigning different weights to each performance indicator based on its importance in the application scenario; S53: Calculate the process dimension similarity and use the interval matching degree algorithm to determine whether the process parameters of the candidate formula fall within the process interval specified by the search conditions; S54: Calculate the application dimension similarity using a semantic matching degree algorithm. Based on the one-hot encoding of the application dimension vector, calculate the device type matching degree and operating condition compatibility.

9. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, The step S6, which involves dynamically adjusting the weight coefficients of the similarity across different dimensions based on the application scenario, includes: S61: Establish an application scenario-dimensional weight mapping table. The mapping table is constructed based on historical retrieval logs and expert experience, and represents the importance weight of each dimension under different application scenarios. S62: When the search intent explicitly specifies the application scenario, the corresponding weight coefficient is directly queried from the mapping table; S63: When the search intent does not explicitly specify the application scenario, the application scenario category is inferred through the classification model based on the keywords and constraints in the search request, and then the corresponding weight coefficient is queried; S64: Introduce a user feedback mechanism to dynamically optimize the weight coefficients based on user click behavior and evaluation information of search results, and achieve adaptive adjustment of the weight coefficients; S65: Overall Matching Score.

10. The method for retrieving industrial lubricating oil formulation data according to claim 1, characterized in that, The automatic verification of the compliance of each formula after sorting in step S7 includes: S71: Extract the standard and specification requirements corresponding to each formula in the search results, including national standards, industry standards and enterprise internal control standards; S72: Compare the measured or predicted values ​​of each performance indicator of the formula with the limit values ​​in the standard specifications item by item; S73: For formulations that exceed the limits specified in the standard specifications, mark them as non-compliant and record the non-compliance items; S74: For formulations marked as non-compliant, recommend alternative components or adjustment schemes based on substitution relationships in the knowledge graph to make them meet compliance requirements; S75: Output a list of search results that includes compliance indicators, explanations of non-compliance items, and adjustment suggestions.