A hydraulic cylinder fault diagnosis and repair method fusing a knowledge graph and a large model thinking chain

Abstract

The application discloses a hydraulic cylinder fault diagnosis and repair method fusing a knowledge graph and a large model thinking chain, and belongs to the field of intelligent operation and maintenance of industrial hydraulic equipment. The method establishes a two-dimensional knowledge graph ontology framework based on the multi-field coupling failure mechanism of the core components of the hydraulic cylinder, integrates multi-source fault data to construct the knowledge graph, quantifies the coupling failure correlation degree between the components, marks the core correlation groups, executes generative completion by using a large language model, and stores the result in the database after multidimensional verification, dynamically corrects the fault triggering threshold and the upper limit value of the verification standard in the safe direction according to the working condition parameters, finally executes multi-round progressive reasoning positioning of the fault root cause along the fault correlation dimension, and outputs a whole-process maintenance decision document in the linkage maintenance execution dimension. The fault positioning accuracy and the maintenance scheme adaptation rate are improved, the outsourcing part adaptation failure rate is reduced, and the maintenance cycle is shortened.

Description

Technical Field

[0001] This invention relates to the field of industrial hydraulic equipment operation and maintenance technology, specifically to a method for diagnosing and repairing hydraulic cylinder faults that integrates knowledge graphs and large model thinking chains. Background Technology

[0002] Hydraulic cylinders are key linear actuators in heavy-duty equipment in core industrial sectors such as construction machinery, mining and metallurgy, and heavy machine tools. Their operational reliability directly determines the continuous operation capability of the production line, equipment efficiency, and on-site production safety. Industry statistics show that hydraulic cylinder failures account for over 35% of hydraulic system failures in heavy-duty equipment. Among these, failures caused by multi-field coupling failures of core components account for 62% of all hydraulic cylinder failures. Over 70% of hydraulic cylinder failures originate from the failure of core purchased components such as seals and guide rings. Incorrect compatibility of purchased components leads to a 28% rework rate, and the recurrence rate of failures within 30 days after repair under harsh operating conditions exceeds 30%.

[0003] Current technology suffers from four core deficiencies. First, it lacks fundamental research on the multi-component, multi-field coupled failure mechanisms of hydraulic cylinders, leaving a fundamental gap in coupled fault diagnosis and knowledge completion, thus failing to provide a precise and complete domain knowledge foundation for large-scale industrial models. Second, it lacks a full lifecycle data system and knowledge completion mechanism for core outsourced components, resulting in a persistently high error rate in component selection and adaptation. Third, theoretical maintenance standards are based on ideal operating conditions, which significantly conflict with actual field scenarios, leading to insufficient on-site adaptability of maintenance solutions under harsh conditions. Fourth, the dual lack of domain knowledge completion and a full-process reasoning mechanism results in low accuracy of scenario reasoning in large-scale models, hindering the formation of a safe closed loop for diagnosis, maintenance, acceptance, and traceability. These deficiencies, when combined, lead to problems such as an incomplete domain knowledge system, inaccurate fault diagnosis, and a high fault recurrence rate. Summary of the Invention

[0004] This invention proposes a hydraulic cylinder fault diagnosis and repair method integrating knowledge graphs and large-scale model thinking chains, aiming to solve the following technical problems: First, the associated failures of core components of hydraulic cylinders under multi-field coupling cannot be accurately identified and located. Existing technologies lack the underlying physical mechanism support for coupled failures, making coupled fault diagnosis impossible. Second, general industrial large-scale models in the hydraulic cylinder field suffer from fragmented knowledge, low accuracy in completion, and the generation of false knowledge, lacking a domain-specific knowledge graph ontology framework and knowledge completion verification mechanism. Third, theoretical maintenance standards are based on ideal working conditions, which have systematic adaptation conflicts with harsh on-site operation scenarios, resulting in insufficient on-site adaptability of maintenance solutions. Fourth, core parameters such as fault thresholds, coupling coefficients, and scenario weights lack scientific quantitative basis and verification systems, resulting in poor engineering repeatability. Fifth, the entire process from knowledge completion to fault diagnosis, maintenance execution, and acceptance traceability lacks a safety closed-loop control mechanism, leading to a persistently high fault recurrence rate.

[0005] To address the aforementioned technical problems, this invention provides a hydraulic cylinder fault diagnosis and repair method that integrates knowledge graphs and large-scale model thinking chains, comprising the following steps: Step S1: Based on the multi-field coupling failure mechanism of the core components of the hydraulic cylinder, establish an ontology framework of a two-dimensional knowledge graph including the fault association dimension and the maintenance execution dimension; Step S2: Integrate multi-source fault domain data, identify entities and relationships through a knowledge extraction model, and construct a knowledge graph according to the ontology framework; Step S3: Based on the fault records in the knowledge graph, quantify the degree of coupling failure association between core component pairs, and mark the component pairs whose association degree reaches a preset threshold as the core association group of coupling failure. Step S4: Generative completion is performed on the missing knowledge in the knowledge graph using a large language model. The completion results are then updated after passing multi-dimensional verification. Step S5: Calculate the scenario basic weight based on the actual working parameters of the hydraulic cylinder, and dynamically correct the fault trigger threshold in the fault association dimension and the upper limit of the verification standard in the maintenance execution dimension based on the scenario basic weight. Step S6: Based on the input fault phenomenon information, perform multiple rounds of progressive reasoning along the fault association dimension to converge to the root cause of the fault layer by layer. During the reasoning process, pruning protection is applied to the core association group of the coupled fault. After the root cause is located, it is associated with the maintenance execution dimension, and the full-process maintenance decision document is output.

[0006] Preferably, in step S1, the fault association dimension adopts an eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-thight-two-eight ... The maintenance execution dimension adopts a seven-element link, which includes seven types of entity nodes: fault phenomenon, root cause, coupled related components, maintenance steps, safety specifications, spare parts and verification standards. Among them, the spare parts node distinguishes between self-made parts and purchased parts and marks their attributes respectively. The fault association dimension and the maintenance execution dimension are integrated through a dual-chain fusion via shared root cause nodes and coupled associated component nodes; In step S2, the multi-source fault domain data includes fault case data, maintenance practice data, and purchased parts archive data.

[0007] Preferably, the degree of coupling failure correlation in step S3 is determined by the inherent coupling failure strength. The representation, expressed as: ; In the formula The number of times a target component co-occurs with a fault in the knowledge graph. This represents the total number of hydraulic cylinder failure records. The operating condition influence coefficient is set according to the severity of the operating conditions. The coupling failure mechanism factor is determined based on the physical model of hydraulic cylinder failure and large sample statistics; When the value exceeds a preset threshold, the component pair is determined to be the core associated group of the coupling fault.

[0008] Preferably, the actual operating condition parameters in step S5 include start-stop frequency, environmental severity, and operational standardization, and the scenario basic weights... The calculation expression is: ; In the formula , These are weighted sub-items set according to the start-stop frequency. These are the weighted sub-items set according to the severity of the environment. These are weighted sub-items set according to the level of operational standardization. The dynamic correction method is as follows: As a correction factor, the fault triggering threshold is multiplied by The corrected actual trigger threshold is obtained; As a correction factor, the upper limit of the verification standard is multiplied by... The revised actual acceptance limit value is obtained, where Safety margin factors set according to operating condition levels and .

[0009] Preferably, the knowledge extraction model in step S2 is a pre-trained language model that uses fault reports, maintenance manuals, industry standards and enterprise operation records in the field of hydraulic cylinders as fine-tuned corpus. The knowledge extraction model is configured with purchased component rules and coupled fault rules. The purchased component rules extract model specifications and differential attributes through regular expressions and material feature lexicon. The coupled fault rules extract the coupled fault causal chain between components through a coupled association lexicon.

[0010] Preferably, the generative completion in step S4 includes: comparing the constraints of the ontology framework using a graph traversal algorithm to identify entries in the knowledge graph that have missing entity attributes, broken associations, or incomplete parameters of purchased components, forming a completion requirement list; using a domain-fine-tuned large language model as the generation engine, injecting semantically similar knowledge fragments retrieved from the vector database into the context to generate completion content item by item; writing the completion knowledge that has passed the multi-dimensional verification into the knowledge graph after standardization according to the ontology framework, and transferring the content that has not passed the verification to manual review.

[0011] Preferably, the multi-dimensional verification includes mechanism compliance verification, standard compliance verification, and data support verification; The compliance verification of the mechanism involves matching the causal link of the supplemented content with a preset coupling failure rule base to verify whether it conforms to the multi-field coupling failure mechanism. The standard compliance verification compares the numerical parameters in the completed content with a preset parameter-standard mapping table to verify whether they fall within the range allowed by industry standards. The data support verification involves retrieving semantically similar historical test records from the vector database to verify whether the completed content is supported by actual test data.

[0012] Preferably, the knowledge graph is stored collaboratively using a graph database and a vector database: the graph database carries entity nodes and associated edges to support the step-by-step retrieval of associated entities along the fault association dimension in step S6; the vector database carries feature vectors encoded by the knowledge extraction model to support semantic similarity retrieval in step S4; each entity node in the graph database carries a unique identifier, and the corresponding vector in the vector database is indexed using the unique identifier to achieve bidirectional linkage query from entity to vector and from vector to entity.

[0013] Preferably, step S6 includes: The multi-round progressive reasoning adopts a four-level targeted inquiry to converge layer by layer: the operating condition layer confirms the start-stop frequency and environmental parameters, the symptom layer locks in the core manifestations of leakage, jamming or abnormal noise, the coupling fault layer investigates the accompanying failure phenomena of related components, and the detail layer confirms the fault location and occurrence sequence. The pruning protection method includes: measuring the statistical correlation between component faults and target faults using mutual information entropy. : ; In the formula , These are component failure events and target failure events, respectively. The joint probability of component failure and target failure. This represents the probability of component failure. The probability of the target failure occurring; Components below a preset pruning threshold are removed from the causal chain, but components belonging to the core association group of the coupling fault are retained.

[0014] Preferably, in step S2, the following adaptation processing is also performed on the purchased parts file data: a dedicated data subset for purchased parts is established, with fields including model code, material composition, rated parameters, suitable working conditions, shelf life, repair records, alternative models, and on-site verification pass rate; A three-dimensional matching matrix is ​​constructed for the models of purchased components, the specifications of self-made components, and the operating parameters. Cosine similarity is used to measure the degree of matching between purchased components. ; In the formula , Let the feature vectors of the two purchased components be... , They are vectors , The Each component; selects models whose on-site verification pass rate reaches a preset standard from candidate models whose similarity reaches a preset similarity threshold, and associates the selection result with the spare parts node of the maintenance execution dimension.

[0015] The beneficial effects of the present invention include at least the following: First, by establishing a multi-field coupling failure mechanism system and quantifying the intensity of coupling failure, this invention elevates fault diagnosis from single-component investigation to a systematic tracing of coupled related groups. Combined with a multi-round progressive reasoning mechanism that applies pruning protection to the core related groups of coupled faults, the accuracy of fault location is improved.

[0016] Second, by establishing a dedicated subset of data for purchased parts and a three-dimensional adaptation matrix, candidate models are retrieved using cosine similarity and the best model is selected based on the pass rate of on-site verification, thereby reducing the failure rate of purchased parts adaptation and shortening the response time for alternative model recommendations.

[0017] Third, by dynamically adjusting the fault trigger threshold and the upper limit of the verification standard through the basic weight of the scenario, the fault warning is triggered in advance under harsh working conditions and the maintenance acceptance standard is automatically tightened, thereby improving the adaptability of the maintenance plan and the stability of the scenario adaptation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the fault repair double-chain structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the federated data acquisition system process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the thought chain reasoning framework of an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0020] This invention takes the multi-component, multi-field coupling failure mechanism of hydraulic cylinders as its underlying core, uses a knowledge graph in the hydraulic cylinder field as its knowledge carrier, and employs a large industrial model as the core for knowledge completion and reasoning. It constructs a six-step, full-chain technical system: Step S1 establishes a two-dimensional ontology framework based on the coupling mechanism; Step S2 integrates multi-source data to construct a knowledge graph; Step S3 quantifies the coupling failure intensity and identifies core association groups; Step S4 completes missing knowledge using a large language model; Step S5 dynamically adjusts thresholds and standards according to operating conditions; and Step S6 executes multi-round progressive reasoning to locate the fault and outputs a maintenance decision document. Each step is progressive and mutually supportive, forming a closed-loop process of "knowledge completion—diagnosis and decision-making—acceptance and traceability," as follows: Figure 1 As shown.

[0021] Step S1: Based on the multi-field coupling failure mechanism of the core components of the hydraulic cylinder, establish an ontology framework of a two-dimensional knowledge graph including the fault association dimension and the maintenance execution dimension.

[0022] The dual-dimensional ontology construction is the core framework of the knowledge graph in this embodiment. It deeply integrates the multi-field coupling failure mechanism of the core components of the hydraulic cylinder, achieving a systematic and standardized representation of hydraulic cylinder fault knowledge. Simultaneously, it provides a standardized knowledge structure and constraint boundaries for knowledge completion in the large industrial model, clarifying the entity types, relationships, and attribute specifications for knowledge completion. This ensures that the knowledge completed in the large model conforms to the physical mechanisms and engineering specifications of the hydraulic cylinder field, preventing the generation of unfounded false knowledge. It includes fault association dimensions and maintenance execution dimensions, such as... Figure 2 As shown.

[0023] The fault correlation dimension targets the coupled failure characteristics of hydraulic cylinders, constructing an eight-eight-element link consisting of core components, coupled failure pairs, assembly level, material property system, typical working condition type, core performance parameters, coupled failure strength, and fault trigger threshold. Starting with the five core components of cylinder barrel, piston, piston rod, seal ring, and guide ring as anchor points, the corresponding elements are correlated level by level, and finally the fault trigger threshold range corresponding to each performance parameter is anchored.

[0024] The fault trigger threshold is calculated based on a weighted fitting mechanism of measured data and theoretically calculated values, and the expression is: ; In the formula These are measured values, and the data comes from the measured records of more than 2,000 historical fault cases of the enterprise, with a weight of 0.6. These are theoretical calculation values, with data cited from the *Hydraulic Cylinder Design Manual*, GB / T 3766-2015, and ISO 10771-1:2015, with a weight of 0.4. The weight allocation is based on the principle in GB / T 3766-2015 that "field measured data takes precedence over theoretical calculation values," combined with statistical analysis of 1500 hydraulic cylinder failure cases. The measured data demonstrates a 22.7% higher accuracy in characterizing the failures than the theoretical values, meeting the requirement in reliability design specifications that field data should have a weight of no less than 0.5. Orthogonal experiments determined that 0.6 / 0.4 is the optimal weight combination.

[0025] The maintenance execution dimension constructs a seven-element full-process link consisting of fault phenomena, root causes, coupled related components, maintenance steps, safety specifications, spare parts, and calibration standards, supplementing the practical and safety knowledge system for fault handling. The core optimization is a dual-attribute labeling system: component attribute labeling clearly distinguishes between self-made and purchased parts; self-made parts are labeled with production batch and internal control standards, while purchased parts are labeled with supplier code, full lifecycle parameters, alternative models, and on-site verification pass rate; safety level labeling marks the safety operation level and high-risk operation warnings for maintenance steps, conforming to the hydraulic system safety specifications of GB / T 3766-2015.

[0026] The fault correlation dimension and maintenance execution dimension are integrated through a dual-chain fusion of shared root cause nodes and coupled component nodes: the fault causal chain is rigidly linked to the maintenance process chain through root cause nodes, and flexibly linked through component and operating condition nodes, so that the fault diagnosis results can be directly indexed to the corresponding maintenance plan, safety specifications and verification standards.

[0027] Step S2: Integrate multi-source fault domain data, identify entities and relationships using a knowledge extraction model, and construct a knowledge graph according to the ontology framework. The cross-domain data integration process focuses on data security and comprehensiveness, breaking down barriers to cross-enterprise data sharing. It provides data support for coupling mechanism research, large-scale model fine-tuning, and model training. Simultaneously, it provides high-quality labeled datasets for knowledge extraction and completion of large industrial models in the hydraulic cylinder field, and offers fundamental data support throughout the entire lifecycle for knowledge completion of purchased components. Specifically, for example… Figure 3 As shown.

[0028] In this embodiment of the invention, a horizontal federated learning architecture is used to integrate a differential privacy protection mechanism to achieve cross-enterprise data security integration, and privacy budget parameters are configured. (Based on GB / T 35273-2020), three types of core data are securely integrated while ensuring that the enterprise's original data does not leave its local area.

[0029] The three core data categories include failure case data, maintenance practice data, and purchased component archive data. Failure case data includes failure symptoms, operating condition information, coupled component groups, failure analysis, and handling results; maintenance practice data includes maintenance procedures, safe operation records, tool lists, spare parts consumption, and calibration results; and purchased component archive data includes model parameters, material properties, suitable operating conditions, warranty period, repair records, replacement models, and on-site verification pass rate.

[0030] For core outsourced components such as hydraulic cylinder seals and guide rings, this invention constructs a dedicated data subset. The core fields cover model code, material composition, rated parameters, compatible components, compatible operating conditions, supplier information, warranty period, repair records, alternative models, and on-site verification pass rate, thus solving the pain points of missing parameters and difficulty in selecting outsourced components.

[0031] The knowledge extraction model of this invention is a pre-trained language model using hydraulic cylinder-specific text data as its fine-tuning corpus. The fine-tuning corpus includes 45,000 fault reports, 30,000 maintenance manuals, 20,000 industry standards, and 25,000 enterprise practical records, totaling 120,000 texts, of which at least 12,000 are coupled with specific fault-related case studies. The model optimization parameter is a learning rate of 5e. -5 The training consisted of 10 rounds with a batch size of 32. The AdamW optimizer and cross-entropy loss function were used. After testing, the accuracy rate for domain terminology recognition was no less than 93.5%, and the accuracy rate for extracting causal chains of coupled faults was no less than 90.8%.

[0032] Meanwhile, the knowledge extraction model in this embodiment of the invention is equipped with two sets of dedicated extraction rules. The rules for purchased components accurately identify different specifications and models within the same series using regular expressions, and extract differentiated attributes through a material feature lexicon. The rules for coupled faults construct a hydraulic cylinder coupled association lexicon, accurately extracting coupled fault causal chains such as "seal ring wear—piston rod scratch," ensuring the complete extraction of coupled failure knowledge. Furthermore, the parameterized extraction tool, which integrates regular expression matching and semantic similarity matching mechanisms, has been tested to achieve a key information extraction accuracy of no less than 91.2%.

[0033] In this embodiment of the invention, the knowledge graph is stored collaboratively using a graph database and a vector database. The graph database is Neo4j, which stores entity nodes such as faults, operating conditions, components, coupled component groups, maintenance steps, safety specifications, and verification standards, as well as related edges such as coupling relationships between components and causal relationships between faults. This supports subsequent retrieval of related entities level by level along the fault-related dimension, with a query response time of no more than 500ms. The vector database is Milvus, which stores 768-dimensional feature vectors encoded by the Knowledge Extraction Model (BERT) from coupled fault features, similar fault text, purchased component feature vectors, and knowledge completion feature vectors. This supports subsequent semantic similarity retrieval, with a retrieval response time of no more than 100ms and a default similarity threshold of no less than 90%.

[0034] Each entity node in the graph database is assigned a unique identifier, and the corresponding vector in the vector database is indexed using this unique identifier. This enables bidirectional linkage between retrieving entity nodes from feature vectors and querying related features from entity nodes to entity nodes.

[0035] In this embodiment of the invention, a dual-drive mechanism combining automatic updates and manual feedback is also established to ensure the timeliness, accuracy, and traceability of the knowledge graph, and to achieve closed-loop iteration of knowledge completion.

[0036] In terms of automatic updates, the associated enterprise ERP system completes the update of spare parts inventory and specifications, the associated MES system completes the update of maintenance records and fault cases, and the associated supplier system completes the update of the full life cycle information of purchased parts. The timed scanning interval does not exceed 24 hours. When data changes are detected, the graph nodes and associations are automatically updated, and the knowledge completion process is triggered simultaneously.

[0037] Regarding human feedback, we have opened feedback channels on both the web and mobile platforms, allowing maintenance personnel to input new coupled fault cases and suggestions for optimizing maintenance solutions. We employ a review mechanism involving three hydraulic cylinder industry experts, and content with a pass rate of no less than 80% is standardized by knowledge engineers and updated to the atlas.

[0038] The version rollback mechanism fully preserves the historical version information of each node, supports querying historical status by version number and rolling back to any historical version with administrator authorization, ensuring the traceability of knowledge updates.

[0039] Step S3: Based on the fault records in the knowledge graph, quantify the degree of coupling failure association between core component pairs, and mark the component pairs with the degree of association reaching the preset threshold as the core association group of coupling failure.

[0040] This step builds upon the fault records accumulated in the knowledge graph constructed in step S2, and utilizes the inherent coupling failure strength. The degree of coupling failure correlation between core component pairs is expressed as: ; In the formula This represents the number of times a target component's faults co-occur in the knowledge graph. The total number of hydraulic cylinder failure records is calculated based on statistics from over 2,000 failure cases. According to GB / T 3766-2015 and ISO 10771-1:2015 classification settings: normal operating conditions (10℃ to 30℃, no more than 100 start-stop cycles per day, pressure fluctuation no more than 5%) Moderately harsh operating conditions (30℃ to 45℃ or below 10℃, 100 to 300 start-stop cycles per day, pressure fluctuation of 5% to 15%) Extremely harsh operating conditions (above 45℃ or below -20℃, more than 300 start-stop cycles per day, pressure fluctuations exceeding 15%) For cross-interval scenarios, the coefficients are determined using weighted interpolation.

[0041] The coupling failure mechanism factor is set based on the hydraulic cylinder failure physical model and large sample statistical analysis, and the core values ​​are shown in Table 1.

[0042] Table 1 Specifically, the seal ring and piston rod assembly A value of 1.4 corresponds to tribological fluid-structure interaction failure, and this pair accounts for 42.3% of all coupled failure cases; the guide ring and piston rod pair... Taking 1.3 corresponds to motion precision coupling failure, accounting for 31.5%; piston and cylinder assembly pair Taking 1.2 corresponds to gap flow-structural deformation coupling failure, accounting for 22.7%; the remaining component combinations... If we take 1.0, the percentage is less than 5%.

[0043] The values ​​were assigned using an arithmetic progression method (step size 0.1), with each coupled component assigned a value from high to low according to its proportion in the total coupled faults. Sensitivity analysis verified that the accuracy of coupled fault identification fluctuated by less than 2.5% within a value variation range of ±0.1.

[0044] In this embodiment, The component pair was then identified as a core associated group of coupled faults. This threshold was determined through ROC curve analysis. Among 1500 fault cases, the Youden index of this threshold was 0.83, with a sensitivity of 92.1% and a specificity of 90.3%, making it the optimal classification threshold.

[0045] Step S4: Generative completion is performed on the missing knowledge in the knowledge graph using a large language model. The completion results are then updated after passing multi-dimensional verification.

[0046] In this embodiment, based on the rule constraints of the two-dimensional knowledge graph ontology framework, a graph traversal algorithm is used to identify problems such as missing entity attributes, broken relationships, gaps in operational condition adaptation knowledge, and incomplete parameters of purchased parts in the graph, generating a knowledge completion requirement list. Simultaneously, based on newly added fault cases and maintenance data, incremental knowledge gaps such as novel coupled failure modes, novel maintenance processes, and novel purchased part models not yet included in the graph are identified.

[0047] Subsequently, using a domain-fine-tuned large language model as the generation engine, a retrieval-enhanced generation mechanism is adopted: before generating the complete content, existing knowledge fragments that are semantically closest to the complete requirements are retrieved from a vector database that includes the underlying mechanism of coupling failure, industry standard specifications, and federated integration as reference context injection prompts. Complete content is generated item by item for the complete requirements list, including core content such as entity attribute completion, relational reasoning, alternative model adaptation knowledge completion, working condition fault association knowledge completion, and maintenance process standardization completion.

[0048] Finally, the verification is carried out through a three-level verification mechanism: mechanism compliance verification, standard compliance verification, and data support verification.

[0049] Mechanism compliance verification involves matching the causal links of the supplementary content with a pre-defined coupled failure rule base. This rule base includes coupled failure rules, such as "aging of the sealing ring leads to an increase in the sealing gap, which in turn leads to an increase in leakage." These rules are written by experts in the field of hydraulic cylinders and peer-reviewed to verify whether the supplementary knowledge conforms to the physical causal logic of multi-field coupled failure mechanisms.

[0050] The standard compliance verification compares the numerical parameters in the supplementary content with a preset parameter-standard mapping table. This mapping table covers the core parameters in the hydraulic cylinder field and their corresponding allowable ranges in standards such as GB / T 3766-2015 and ISO 10771-1:2015, verifying whether the numerical parameters fall within the allowable range of industry standards.

[0051] Data support verification involves retrieving historical test records in the vector database that are semantically similar to the completed content to verify whether the completed content is supported by actual test data.

[0052] The completed content must pass the above three levels of verification to be deemed qualified. The completed knowledge that passes the verification is standardized according to the two-dimensional ontology framework and then synchronously updated to the graph database and vector database. If any level fails, the completed content will be transferred to the manual review process.

[0053] As a preferred implementation, the output of the three-level verification is further quantified to calculate the overall confidence level: the first level uses the association ratio of the matching rule base in the causal link as the score for this level; the second level uses the ratio of parameters that meet the standard range as the score for this level; and the third level uses the mean cosine similarity of the retrieved similar measured records as the score for this level. The three levels of scores are weighted and summed to obtain the overall confidence level. The completed content with an overall confidence level of not less than 90% is directly entered into the database, the annotation with 70% to 90% is subject to review, and the content with less than 70% is rejected.

[0054] Step S5: Calculate the scenario basic weight based on the actual working parameters of the hydraulic cylinder, and dynamically correct the fault trigger threshold in the fault association dimension and the upper limit of the verification standard in the maintenance execution dimension based on the scenario basic weight.

[0055] In this embodiment, the scenario base weights are calculated based on the actual operating parameters of the hydraulic cylinder. The expression is: ; In the formula Default value , , . These are weighted sub-items set according to the start-stop frequency. These are the weighted sub-items set according to the severity of the environment. The weighted sub-items are set according to the level of operational standardization. The three are based on a value of 1 under normal working conditions, and the value is larger for more severe working conditions.

[0056] Based on the cycle grading settings for hydraulic cylinder durability testing in ISO 107711: high-frequency start-stop scenario (more than 100 start-stop cycles per day). Typical scenarios . According to the environmental classification of hydraulic systems in GB / T3766-2015: harsh environments such as high temperature, low temperature or dusty environments. normal environment . Based on the safety operating procedures for hydraulic systems in GB / T3766-2015, the following standardized operating scenarios are defined: Non-standard operating scenarios (overload / uneven load / improper start / stop) .

[0057] Weighting coefficient , , The basis for this analysis is as follows: Based on FMEA analysis of 2000 hydraulic cylinder failure cases, the statistical results show that the contribution of start-stop frequency to failure risk is 42.3%, environmental severity is 30.5%, and operational standardization is 27.2%, with normalized weights of 0.4 / 0.3 / 0.3. Simultaneously, a four-factor, three-level orthogonal experiment was conducted to verify this, using the maintenance plan adaptability rate as the experimental index. Range analysis showed... It has the greatest impact on the results and is completely consistent with the FMEA statistical results.

[0058] In another embodiment of the present invention, it is also based on scene-based weights. Independent correction coefficients are designed for the fault warning and maintenance acceptance stages, respectively. Both stages strictly adhere to the principle of safety first, addressing the original issues. The shortcomings of relaxing standards at times.

[0059] Specifically, with To correct the coefficient, multiply the fault triggering threshold in the fault association dimension by... The corrected actual trigger threshold is obtained. Under harsh operating conditions. , The corrected threshold is lower than the theoretical threshold, thus triggering an early warning before the performance parameters deteriorate to the theoretical failure point.

[0060] Taking a typical scenario as an example: the theoretical threshold for triggering a sealing gap fault is 0.05mm, under high-frequency start-stop and moderately harsh environments. , The actual fault trigger threshold is This aligns with the safety principle of early warning for severe working conditions.

[0061] For the upper limit of the verification standard in the maintenance execution dimension, To correct the coefficient, multiply the upper limit of the verification standard in the maintenance execution dimension by... The revised upper limit of actual acceptance value is obtained. Safety margin factors set according to operating condition levels and : Normal operating conditions Moderately harsh working conditions Extremely harsh working conditions It is set in accordance with the "Specification for Reliability Design of Hydraulic Components".

[0062] Following the previous example: the theoretical upper limit for the sealing gap acceptance is 0.05mm. The actual acceptance limit is mm, which is about 20% tighter than the theoretical standard, significantly improving the maintenance safety margin under harsh working conditions.

[0063] The core logic of the above-mentioned dual-dimensional correction mechanism is: the more severe the working conditions, the greater the basic weight of the scenario, the lower the fault trigger threshold (early warning), and the stricter the acceptance standards (tightening the upper limit), which fully conforms to the safety margin design principle of hydraulic systems.

[0064] Step S6: Based on the input fault phenomenon information, perform multiple rounds of progressive reasoning along the fault association dimension and converge to the root cause of the fault layer by layer. During the reasoning process, pruning protection is applied to the core association group of coupled faults. After the root cause is located, it is associated with the maintenance execution dimension, and the full-process maintenance decision document is output.

[0065] In this step, optimization is performed on the characteristics of hydraulic cylinder coupling faults. A multi-round progressive reasoning mechanism is constructed, linking with maintenance safety specifications and real-time linkage with the knowledge completion module, to achieve accurate fault location and safe output of maintenance solutions. When a knowledge gap is identified during the reasoning process, the large-scale model knowledge completion process is triggered in real time. The completed knowledge is then used simultaneously for reasoning and decision optimization, achieving real-time linkage between knowledge completion and accurate decision-making. Specifically, as follows... Figure 4 As shown.

[0066] Specifically, based on the input fault phenomenon information, multiple rounds of progressive reasoning are performed along the causal chain along the fault correlation dimension, converging layer by layer to the root cause of the fault. The targeted inquiry level is divided into four levels: the operating condition level confirms operating condition information such as operating condition type, start-stop frequency, ambient temperature and pressure fluctuations; the symptom level confirms core fault symptoms such as leakage, jamming, abnormal noise, and insufficient pressure; the coupled fault level investigates accompanying fault phenomena of related components, such as leakage accompanied by piston rod scratches, jamming accompanied by guide ring wear, etc.; the detail level confirms detailed information such as leakage location, jamming stroke, and fault occurrence sequence. Tests show that the fault phenomenon convergence efficiency of this mechanism is no less than 88%, and after four levels of inquiry, the fault investigation scope can be reduced to within 12% of the original scope.

[0067] In another embodiment of the present invention, mutual information entropy is used to measure the statistical correlation between the faults of each component and the target fault during the inference process: ; In the formula , These are component failure events and target failure events, respectively. The joint probability of component failure and target failure. This represents the probability of component failure. The probability of the target failure occurring. When the value falls below the preset pruning threshold of 0.1, the corresponding component is removed from the causal chain. The pruning threshold of 0.1 is verified through information gain analysis: at this threshold, the redundant information removal rate reaches 92%, while the retention rate of key fault information is not less than 95%. However, components belonging to the core association group of coupled faults identified in step S3 are not removed regardless of their mutual information entropy value, to avoid key association components being mistakenly pruned.

[0068] After root cause localization is completed, the root cause node and coupled component nodes shared by the fault association dimension and maintenance execution dimension are linked to the maintenance execution dimension, automatically outputting four core documents conforming to the GB / T 3766-2015 standard: a fault location report, recording the fault phenomenon, root cause, associated coupled components, and operational condition impact analysis; a step-by-step maintenance guide, including the entire process from shutdown and depressurization to pressure testing, and safety warnings for each step; a tool and spare parts list, specifying the required special tools and spare parts models; and a verification standard, providing the corrected acceptance upper limit values ​​and test conditions for various performance parameters.

[0069] In another embodiment of the present invention, adaptation processing is also performed on the purchased component archive data. After establishing a dedicated subset of purchased component data, a three-dimensional adaptation matrix is ​​constructed for purchased component models, self-made component specifications, and operating parameters. When it is necessary to select a replacement purchased component, candidate models are retrieved from the dedicated data subset using cosine similarity. ; In the formula Let be the feature vector of the purchased part to be replaced. The feature vector of the candidate purchased component. , They are vectors , The Each component is selected. From candidate models whose similarity reaches a preset similarity threshold, models with a field verification pass rate meeting a preset standard are selected. Testing shows that the differentiation accuracy between similar models does not exceed 0.03 feature distance, and the recommended response time for alternative models does not exceed 10 seconds. The selection results are linked to the spare parts node in the maintenance execution dimension.

[0070] Example 1 This example illustrates a hydraulic cylinder sealing piston rod coupling leakage failure under moderately harsh operating conditions with high-frequency start-stop. Basic scenario information: An engineering machinery manufacturing company, 280 start-stop cycles per day (high-frequency start-stop), ambient temperature 38℃ (moderately harsh operating conditions), pressure fluctuation 8%, the fault symptom is leakage at the cylinder head during hydraulic cylinder extension accompanied by slight jamming of the piston rod movement.

[0071] Step S1: Determine the fault correlation link: Starting from the sealing ring, through the sealing ring-piston rod coupling assembly, sealing assembly, fluororubber material, high-frequency start-stop plus 38℃ environmental conditions, and sealing gap performance parameters, it is correlated to the theoretical fault triggering threshold of 0.081mm. ).

[0072] Step S3: Calculate the coupling failure strength: Total Fault Record Number of times the seal ring and piston rod failures co-occur. Second-rate, , , Confirm that the sealing ring-piston rod is the core associated group of the coupling fault, and set up pruning protection.

[0073] Step S4 performs knowledge completion: By traversing the knowledge graph, it identifies the missing parameters of the fluororubber seal ring adaptation under high-frequency start-stop conditions for this type of hydraulic cylinder and the knowledge gap in the scenario-based acceptance standards for piston rod scratch repair process; the large language model completes the compression threshold and aging rate parameters of the seal ring of this specification under high-frequency start-stop and 38℃ conditions, as well as the roughness and hardness acceptance standards after piston rod scratch repair; the completed content passes three-level verification and is synchronously updated to the knowledge graph.

[0074] Step S5: Calculate scene weights and correction thresholds. The fault warning threshold is corrected to The measured sealing gap of 0.12mm far exceeds the threshold, meeting the fault triggering conditions; the theoretical upper limit for sealing gap acceptance is 0.05mm, and the maintenance acceptance standard is revised to... mm.

[0075] Step S6 executes multi-round reasoning: Level 4 follow-up questions confirm high-frequency start-stop conditions, cylinder head leakage accompanied by jamming, scratches on the piston rod surface, and increased leakage with increasing start-stop frequency, narrowing the investigation scope to 10% of the original scope. Mutual information entropy of the sealing rings during causal chain pruning. Piston rod All are retained, cylinders Pruning was performed to focus on the root causes of coupled faults. The purchased component was a fluororubber sealing ring, model S008-O25-35, with a temperature range of 20℃ to 120℃, a similarity of 0.92, and a 95% pass rate in field verification. Four documents were ultimately output: Fault Location Report: The fault symptoms are cylinder head leakage and piston rod jamming under high-frequency start-stop conditions; the root cause is the coupling failure of the seal ring wear and piston rod surface scratches, with the sealing gap of 0.12mm far exceeding the fault trigger threshold of 0.0686mm; the associated coupling components are the seal ring and piston rod; the operating condition impact is that high-frequency start-stop exacerbates seal wear and piston rod scratches, forming a positive feedback of coupling failure.

[0076] Step-by-step maintenance guide: Includes the entire process of shutdown and pressure relief, cylinder head disassembly, removal of old seals, piston rod surface polishing and repair, seal groove cleaning, new seal installation, cylinder head assembly, pressure testing, and no-load trial operation, with safety operation warnings for each step.

[0077] Tool and spare parts list: special puller, new sealing ring, cleaning agent, torque wrench, 2000-grit metallographic sandpaper, etc.

[0078] Verification standards: Test pressure 30MPa, hold pressure for 5 minutes without leakage; sealing gap ≤0.040mm; piston rod surface scratch depth ≤0.02mm, roughness Ra≤0.4μm; no jamming or abnormal noise during full stroke no-load operation.

[0079] Implementation Results: This project completed the knowledge supplementation for 12 core parameters and 8 related relationships. The coupled fault location results were verified to be correct during maintenance. The maintenance cycle was shortened from 72 hours to 40 hours. There were zero errors in the adaptation of purchased parts. The safety margin was increased by 24% after maintenance, and there was no recurrence of faults within 30 days, verifying the accuracy and safety of this solution. It should be noted that the components and algorithms in the embodiments of the present invention can be replaced with equivalent ones according to actual engineering conditions without affecting the overall implementation of the technical solution of the present invention.

[0080] For example, in terms of database selection, graph databases other than Neo4j can be replaced by ArangoDB or NebulaGraph, and vector databases other than Milvus can be replaced by FAISS or Chroma, as long as they meet the basic capabilities of graph topology query and vector approximate retrieval. Regarding pre-trained models, the pre-trained language models used in the knowledge extraction stage can be replaced by Chinese pre-trained language models such as RoBERTa, ALBERT, or ERNIE, in addition to BERT. The model selection is primarily evaluated based on the accuracy of domain terminology recognition. In terms of similarity calculation, the cosine similarity used in the purchased component adaptation stage can be replaced by Euclidean distance or Pearson correlation coefficient. After replacement, the similarity threshold needs to be recalibrated to ensure adaptation accuracy. Regarding coupling failure mechanism factors, in this embodiment... The values ​​are set based on failure statistics data of hydraulic cylinders in engineering machinery. For different types of hydraulic cylinders such as piston cylinders, telescopic cylinders, and swing cylinders, the values ​​can be adjusted according to their respective failure statistics data. The value of . Regarding differential privacy parameters, the privacy budget parameter. The value can be adjusted within the range of 0.8 to 1.2 according to the enterprise's data security level requirements. The smaller the value, the stronger the privacy protection, but the lower the model training efficiency. Regarding the knowledge verification mechanism, the three-level progressive verification mechanism can be replaced by a verification mechanism combining large model scoring and manual review, or a consistency verification algorithm based on knowledge graph embedding. When replacing, the verification coverage should be no less than the original scheme.

[0081] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0082] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for diagnosing and repairing hydraulic cylinder faults that integrates knowledge graphs and large-scale model thinking chains, characterized in that, Includes the following steps: Step S1: Based on the multi-field coupling failure mechanism of the core components of the hydraulic cylinder, establish an ontology framework of a two-dimensional knowledge graph including the fault association dimension and the maintenance execution dimension; Step S2: Integrate multi-source fault domain data, identify entities and relationships through a knowledge extraction model, and construct a knowledge graph according to the ontology framework; Step S3: Based on the fault records in the knowledge graph, quantify the degree of coupling failure association between core component pairs, and mark the component pairs whose association degree reaches a preset threshold as the core association group of coupling failure. Step S4: Generative completion is performed on the missing knowledge in the knowledge graph using a large language model. The completion results are then updated after passing multi-dimensional verification. Step S5: Calculate the scenario basic weight based on the actual working parameters of the hydraulic cylinder, and dynamically correct the fault trigger threshold in the fault association dimension and the upper limit of the verification standard in the maintenance execution dimension based on the scenario basic weight. Step S6: Based on the input fault phenomenon information, perform multiple rounds of progressive reasoning along the fault association dimension to converge to the root cause of the fault layer by layer. During the reasoning process, pruning protection is applied to the core association group of the coupled fault. After the root cause is located, it is associated with the maintenance execution dimension, and the full-process maintenance decision document is output.

2. The method according to claim 1, characterized in that: In step S1, the fault association dimension adopts an eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-eight-thight-two-eight ... The maintenance execution dimension adopts a seven-element link, which includes seven types of entity nodes: fault phenomenon, root cause, coupled related components, maintenance steps, safety specifications, spare parts and verification standards. Among them, the spare parts node distinguishes between self-made parts and purchased parts and marks their attributes respectively. The fault association dimension and the maintenance execution dimension are integrated through a dual-chain fusion via shared root cause nodes and coupled associated component nodes; In step S2, the multi-source fault domain data includes fault case data, maintenance practice data, and purchased parts archive data.

3. The method according to claim 1, characterized in that, The degree of coupling failure correlation mentioned in step S3 is determined by the inherent coupling failure strength. The representation, expressed as: ； In the formula The number of times a target component co-occurs with a fault in the knowledge graph. This represents the total number of hydraulic cylinder failure records. The operating condition influence coefficient is set according to the severity of the operating conditions. The coupling failure mechanism factor is determined based on the physical model of hydraulic cylinder failure and large sample statistics; When the value exceeds a preset threshold, the component pair is determined to be the core associated group of the coupling fault.

4. The method according to claim 1, characterized in that, The actual operating parameters mentioned in step S5 include start-stop frequency, environmental severity, and operational standardization, and the scenario basic weights. The calculation expression is: ； In the formula , These are weighted sub-items set according to the start-stop frequency. These are the weighted sub-items set according to the severity of the environment. These are weighted sub-items set according to the level of operational standardization. The dynamic correction method is as follows: As a correction factor, the fault triggering threshold is multiplied by The corrected actual trigger threshold is obtained; by As a correction factor, the upper limit of the verification standard is multiplied by... The revised actual acceptance limit value is obtained, where Safety margin factors set according to operating condition levels and .

5. The method according to claim 1, characterized in that, The knowledge extraction model mentioned in step S2 is a pre-trained language model that uses fault reports, maintenance manuals, industry standards and enterprise operation records in the field of hydraulic cylinders as fine-tuning corpus. The knowledge extraction model is configured with purchased component rules and coupled fault rules. The purchased component rules extract model specifications and differential attributes through regular expressions and material feature lexicon. The coupled fault rules extract the coupled fault causal chain between components through a coupled association lexicon.

6. The method according to claim 1, characterized in that, The generative completion in step S4 includes: comparing the constraints of the ontology framework using a graph traversal algorithm to identify entries in the knowledge graph that have missing entity attributes, broken associations, or incomplete parameters of purchased components, forming a completion requirement list; using a domain-fine-tuned large language model as the generation engine, injecting semantically similar knowledge fragments retrieved from the vector database into the context to generate completion content item by item; writing the completion knowledge that has passed the multi-dimensional verification into the knowledge graph after standardization according to the ontology framework, and transferring the content that has not passed the verification to manual review.

7. The method according to claim 6, characterized in that, The multi-dimensional verification includes mechanism compliance verification, standard compliance verification, and data support verification; The compliance verification of the mechanism involves matching the causal link of the supplemented content with a preset coupling failure rule base to verify whether it conforms to the multi-field coupling failure mechanism. The standard compliance verification compares the numerical parameters in the completed content with a preset parameter-standard mapping table to verify whether they fall within the range allowed by industry standards. The data support verification involves retrieving semantically similar historical test records from the vector database to verify whether the completed content is supported by actual test data.

8. The method according to claim 1, characterized in that, The knowledge graph is stored collaboratively using a graph database and a vector database: the graph database carries entity nodes and associated edges to support the step-by-step retrieval of associated entities along the fault association dimension in step S6; the vector database carries feature vectors encoded by the knowledge extraction model to support semantic similarity retrieval in step S4; each entity node in the graph database carries a unique identifier, and the corresponding vector in the vector database is indexed using the unique identifier to achieve bidirectional linkage query from entity to vector and from vector to entity.

9. The method according to claim 1, characterized in that, Step S6 includes: The multi-round progressive reasoning adopts a four-level targeted inquiry to converge layer by layer: the operating condition layer confirms the start-stop frequency and environmental parameters, the symptom layer locks in the core manifestations of leakage, jamming or abnormal noise, the coupling fault layer investigates the accompanying failure phenomena of related components, and the detail layer confirms the fault location and occurrence sequence. The pruning protection method includes: measuring the statistical correlation between component faults and target faults using mutual information entropy. : ； In the formula , These are component failure events and target failure events, respectively. The joint probability of component failure and target failure. This represents the probability of component failure. The probability of the target failure occurring; Components below a preset pruning threshold are removed from the causal chain, but components belonging to the core association group of the coupling fault are retained.

10. The method according to claim 1, characterized in that, In step S2, the following adaptation processing is also performed on the purchased parts file data: a dedicated data subset for purchased parts is established, with fields including model code, material composition, rated parameters, suitable working conditions, shelf life, repair records, alternative models and on-site verification pass rate; A three-dimensional matching matrix is ​​constructed for the models of purchased components, the specifications of self-made components, and the operating parameters. Cosine similarity is used to measure the degree of matching between purchased components. ； In the formula , Let the feature vectors of the two purchased components be... , They are vectors , The Each component; selects models whose on-site verification pass rate reaches the preset standard from candidate models whose similarity reaches the preset similarity threshold, and associates the selection results with the spare parts node of the maintenance execution dimension.

Images