A method and system for constructing a large vertical shaft through-flow pump multi-dimensional intelligent operation and maintenance based on a large language model, and a storage medium
By constructing a dedicated knowledge graph for large vertical shaft axial flow pumps and using a small-sample fine-tuning method, the large language model was optimized, solving the problems of asynchronous time series of multi-source heterogeneous data and difficulty in identifying early fault characteristics. This enabled fully closed-loop intelligent operation and maintenance, improving the accuracy of fault diagnosis and the efficiency of operation and maintenance.
Patent Information
- Application Number
- CN202611125472.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
The operation and maintenance of large vertical shaft axial flow pumps suffers from problems such as asynchronous timing of multi-source heterogeneous data, weak early fault characteristics that are difficult to identify, severe illusion of professional knowledge in hydraulic engineering in large models, and lack of closed-loop iteration in the operation and maintenance process, which makes it difficult to meet the requirements for long-term safe and stable operation.
A multi-dimensional knowledge base for large vertical shaft axial pumps is constructed, deeply integrating mechanism, knowledge, and data. By designing a dedicated knowledge graph ontology architecture and embedding a hydraulic mechanism model, combined with small-sample enhancement fine-tuning and time-series awareness attention mechanism, the large language model is optimized to achieve a fully closed-loop intelligent operation and maintenance system that supports natural language interaction.
It enables accurate identification and timely warning of early faults, improves the reliability and efficiency of operation and maintenance decisions, forms a continuously self-learning operation and maintenance system, and lowers the threshold for operation and maintenance.
Smart Images

Figure CN123070911A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method, system, and storage medium for constructing multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model, belonging to the fields of smart water conservancy and artificial intelligence technology. Background Technology
[0002] Currently, the operation and maintenance management of water conservancy pumping stations mainly relies on the following methods:
[0003] Manual periodic inspections and recording: Maintenance personnel conduct on-site inspections of the water pump units according to fixed cycles (such as daily or weekly), judging the equipment's operating status through methods such as "listening, touching, looking, and smelling," and manually recording parameters such as voltage, current, temperature, and vibration. This method relies heavily on personal experience, is highly subjective, and cannot achieve real-time monitoring. It often only discovers problems after they occur, easily leading to downtime and water waste.
[0004] Threshold alarms based on SCADA systems: Most modern pumping stations deploy data acquisition and monitoring control systems. This system collects pump operating data in real time through sensors and compares it with preset thresholds (such as bearing temperature exceeding 80°C). An alarm is triggered if the threshold is exceeded. The disadvantages of this method are that alarms are isolated and delayed, it cannot analyze the correlation between parameters, and it frequently produces false alarms or missed alarms, making it difficult to diagnose the root cause of the fault.
[0005] Traditional fault diagnosis expert systems: Some advanced pumping stations have attempted to introduce rule-based expert systems. These systems summarize the experience of domain experts into an "if...then..." rule base to assist in fault diagnosis. However, the rule base construction and maintenance of traditional expert systems are costly and time-consuming, and they lack self-learning capabilities. When faced with complex and coupled fault modes, their reasoning ability is limited, and they cannot flexibly cope with unseen operating conditions.
[0006] Based on the aforementioned shortcomings in existing operation and maintenance management, it is currently impossible to solve the four core technical problems in the operation and maintenance of large vertical shaft axial flow pumps: asynchronous timing of multi-source heterogeneous data, weak early fault characteristics that are difficult to identify, severe illusion of professional knowledge in hydraulic engineering in large models, and lack of closed-loop iteration in operation and maintenance processes. As a result, it is difficult to meet the operation and maintenance requirements for the long-term safe and stable operation of large water conservancy pumping stations. Summary of the Invention
[0007] This invention provides a method and system for constructing multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model. It integrates multi-source heterogeneous time-series data, mechanistic models and large language models, and solves technical problems in the operation and maintenance of large vertical shaft axial flow pumps, such as asynchronous time-series data of multi-source heterogeneous data, weak early fault characteristics that are difficult to identify, serious illusion of hydraulic professional knowledge in large models, and lack of closed-loop iteration in operation and maintenance processes. It can meet the operation and maintenance needs of long-term safe and stable operation of large water conservancy pumping stations.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0009] A method for constructing multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model includes the following steps:
[0010] S1 Constructs a multi-dimensional knowledge base for large vertical shaft axial flow pumps that deeply integrates mechanism, knowledge, and data: It performs spatiotemporal synchronization and preprocessing of multi-source heterogeneous data, designs a knowledge graph ontology architecture for large vertical shaft axial flow pumps, embeds hydraulic and fault evolution mechanism models, and completes the three-dimensional deep integration of mechanism, knowledge, and time-series data.
[0011] S2 Shaft Axial Flow Pump Operation and Maintenance Large Language Model Small Sample Enhancement Fine-tuning and Illusion Suppression: Construct a multi-turn dialogue-based small sample fine-tuning dataset, adopt a parameter-efficient fine-tuning framework combined with contrastive learning to train the model, optimize the temporal awareness attention mechanism, and add an illusion suppression module with dual constraints of mechanism and knowledge graph to obtain a vertical domain-specific large operation and maintenance model;
[0012] S3 constructs a fully closed-loop intelligent operation and maintenance system: based on a dedicated operation and maintenance big model and a multi-dimensional knowledge base, it realizes closed-loop operation and maintenance of large vertical shaft axial flow pumps throughout the entire life cycle, including early warning, diagnosis, handling, simulation, feedback and iteration, and supports interactive operation and maintenance management in all scenarios using natural language.
[0013] Among them, the large vertical shaft axial flow pump is a vertical shaft axial flow pump unit with a single unit rated flow rate ≥25m³ / s, a single unit power ≥1000kW, and an impeller diameter ≥2000mm.
[0014] This invention targets large vertical shaft axial flow pumps, which, according to GB 50265-2022 "Pump Station Design Standard", refer to core units with a single unit flow rate ≥25m³ / s, power ≥1000kW, and impeller diameter ≥2000mm, used for flood control, drainage, and inter-basin water transfer. This type of equipment has technical characteristics such as multi-source heterogeneous sensors, large sampling frequency range, strong fault coupling, weak early characteristics, and high requirements for operation and maintenance safety.
[0015] In step S1 above, this step breaks through the limitations of simple knowledge extraction in existing technologies, designs a knowledge graph ontology architecture dedicated to vertical shaft axial flow pumps, integrates hydraulic mechanism models and fault evolution models, realizes spatiotemporal synchronization and deep fusion of multi-source data, and provides a professional knowledge base without illusions for model reasoning.
[0016] In step S1 above, the specific method for spatiotemporal synchronization and preprocessing of multi-source heterogeneous data is as follows: Addressing the issues of temporal asynchrony and feature coupling distortion caused by large differences in sampling frequencies among multiple sensors in a vertical shaft axial flow pump, a spatiotemporal alignment algorithm combining cubic spline interpolation resampling with cross-correlation delay estimation is employed to unify the data from vibration sensors, temperature sensors, pressure sensors, electrical sensors, hydraulic sensors, equipment status sensors, and environmental auxiliary sensors to a 1ms time base, achieving a temporal synchronization error ≤0.5ms. An adaptive wavelet denoising algorithm is used to filter operating noise and retain weak early fault features, completing data cleaning, normalization, and feature extraction to generate a standardized time-series dataset.
[0017] In step S1 above, the multi-source heterogeneous data includes structured operation data and unstructured operation and maintenance data; the structured operation data includes vibration acceleration, vibration velocity, vibration displacement, bearing temperature, winding temperature, cooling water temperature, inlet and outlet pressure, flow rate, head, voltage, current, power, speed and upstream and downstream water levels; the unstructured operation and maintenance data includes equipment design drawings, operation manuals, maintenance records, fault cases, expert experience and water conservancy industry standards and specifications.
[0018] The knowledge graph ontology architecture for vertical shaft axial flow pumps in step S1 above breaks through the limitations of general knowledge graphs. It adopts a 6-level entity architecture, 12 proprietary semantic relationships and 28 engineering attribute dimensions, specifically for the operation and maintenance characteristics of vertical shaft axial flow pumps.
[0019] The 6-layer entity architecture consists of: system layer, component layer, failure mode layer, symptom and characteristic layer, root cause layer, and remedial measures layer.
[0020] The 12 proprietary semantic relationships include 6 basic association relationships and 6 core relationships specific to vertical shaft axial flow pumps, which fully depict the dynamic evolution of faults. The 6 basic association relationships include: composition relationship, fault occurrence, fault manifestation, cause, treatment measures, and execution specifications. The 6 core relationships specific to vertical shaft axial flow pumps include: evolution sequence, fault coupling relationship, mechanism constraint, treatment priority, operating condition adaptability, and degradation trend.
[0021] The 28 engineering attribute dimensions include 28 exclusive attributes for each entity, such as design operating condition parameters, operating condition boundaries, fault evolution time thresholds, fault confidence, operation and maintenance history records, and handling effect evaluation, to achieve accurate matching between knowledge and engineering practice.
[0022] The above 6-level entity architecture includes the following dedicated entity levels: System layer (vertical shaft axial flow pump unit) → Component layer (impeller, main shaft, thrust bearing, guide bearing, synchronous motor, inlet and outlet water channels, coupling, cooling and lubrication system, mechanical seal) → Failure mode layer (blade cavitation, rotor misalignment, rotor imbalance, bearing wear, winding overheating, seal leakage, flow channel blockage, etc.) → Symptom characteristic layer (excessive vibration, abnormal temperature, flow fluctuation, current fluctuation, efficiency reduction, etc.) → Root cause layer (installation deviation, poor lubrication, silt wear, foreign object blockage, cavitation erosion, operating conditions deviating from design range, etc.) → Remedial measures layer (emergency shutdown, online operating condition adjustment, planned maintenance, component replacement, lubrication maintenance, etc.).
[0023] Among the above 12 proprietary semantic relationships, the basic association relationships include: composition relationship, fault occurrence, fault manifestation, cause, handling measures, and execution norm.
[0024] The aforementioned 28 engineering attribute dimensions include: design operating condition parameters, operating condition boundaries, fault evolution time threshold, fault confidence, maintenance history, handling effect evaluation, equipment geometric dimension parameters, component material performance parameters, installation and assembly accuracy threshold, rated hydraulic performance index, cavitation critical performance threshold, flow channel resistance characteristic parameters, efficiency attenuation allowable limit, vibration threshold judgment standard, spindle running swing limit, bearing lubrication performance parameters, motor electrical rated parameters, insulation performance safety threshold, equipment health status level, fault severity level, fault occurrence probability weight, secondary fault association risk, equipment remaining service life, maintenance cycle calibration parameters, spare parts compatibility model, maintenance operation safety procedures, industry standard compliance indicators, and data collection and sampling specifications.
[0025] In step S1 above, the embedded mechanism model includes a hydraulic mechanism model and a rotor dynamics and fault evolution model;
[0026] The hydraulic mechanism model includes: head-flow characteristic equation, shaft power calculation model, required net positive suction head (NPSH) model, pipeline resistance characteristic model, and efficiency characteristic model;
[0027] The rotor dynamics and fault evolution models include a bearing wear degradation trend model, a misalignment fault vibration characteristic evolution model, a cavitation development stage model, and a spindle fatigue life prediction model.
[0028] The method for deep integration of mechanism, knowledge and data in step S1 is as follows: the hydraulic mechanism model and rotor dynamics and fault evolution model are embedded into the knowledge graph of the vertical shaft axial flow pump; the named entity recognition and relation extraction method is used to automatically extract operation and maintenance entities and semantic relations from multi-source heterogeneous data, and the extraction results are deeply integrated with the mechanism model to construct a structured operation and maintenance knowledge graph, forming a three-element closed-loop fusion system of mechanism constraint knowledge, knowledge interpretation data and data optimization knowledge.
[0029] In step S1 above, the core mechanism model of the vertical shaft axial flow pump is embedded into the knowledge graph to set hard constraints for knowledge reasoning, thereby suppressing the illusion of professional knowledge of large models from the root and realizing the upgrade from static knowledge to dynamic evolutionary knowledge.
[0030] In step S1 above, after completing the construction of the dedicated knowledge graph ontology architecture for the vertical shaft axial flow pump, a deep integration of mechanism, knowledge, and data is carried out: First, core mechanism models such as head-flow characteristics, bearing wear degradation, and cavitation development are embedded into the knowledge graph to set hard constraints on engineering mechanisms for knowledge reasoning; further, named entity recognition and relation extraction artificial intelligence technologies are used to accurately extract equipment entities, fault entities, and corresponding semantic relationships from structured operation data and unstructured operation and maintenance documents; the extracted knowledge entities, semantic relationships, and mechanism models are deeply coupled and normalized to construct a structured and reasonable knowledge graph for the operation and maintenance of the vertical shaft axial flow pump, ultimately forming a three-element fusion system of mechanism constraint knowledge, knowledge interpretation data, and data optimization knowledge, realizing the upgrade from static operation and maintenance knowledge to dynamic evolutionary knowledge.
[0031] Step S2 above overcomes the limitations of general fine-tuning methods. Addressing the issues of scarce pump failure samples, difficulty in identifying early weak features, and severe professional knowledge illusion, it proposes a small-sample fine-tuning method that integrates contrastive learning with the Prompt thought chain. It optimizes the temporal awareness attention mechanism, incorporates a mechanism-constrained illusion suppression module, and constructs a highly reliable large-scale operation and maintenance model for vertical shaft axial flow pumps. The specific sub-steps are as follows:
[0032] In step S2 above, a small-sample fine-tuning dataset is constructed: Addressing the scarcity of real fault samples for vertical shaft axial flow pumps, simulation samples of the entire fault lifecycle are generated based on knowledge graphs and mechanism models. These are then integrated with real historical fault samples to construct a multi-turn dialogue-based small-sample fine-tuning dataset with core fields including operating condition description, feature data, fault evolution path, thought chain reasoning, handling solutions, and mechanism verification, replacing the traditional single question-and-answer pair dataset. The core types of this dataset include:
[0033] Status assessment Q&A pairs: Integrate real-time time series data with operating condition boundaries to achieve multi-dimensional assessment of equipment health status;
[0034] Fault Diagnosis Thinking Chain Data Pair: Decomposes fault diagnosis into a complete thinking chain of feature recognition, operating condition matching, mechanism verification, fault location, root cause analysis, and handling suggestions to improve the interpretability of model reasoning;
[0035] Fault evolution prediction data pairs: Based on the fault degradation model, construct the correspondence between fault characteristics and evolution trends at different stages, and train the model's long-term prediction capability;
[0036] Operation and maintenance handling and simulation verification data pair: Match the hierarchical handling plan, combine the digital twin simulation results, and construct the plan effectiveness verification data.
[0037] To address the aforementioned technical problem of scarce fault samples in vertical shaft axial flow pumps, this invention uses a knowledge graph as the knowledge base and a mechanism model as the constraint boundary to generate, in batches, simulated fault evolution samples covering the entire degradation stage. These simulated samples are then balanced and fused with real historical fault samples collected from the pump station site to construct a structured, multi-turn, dialogue-based small-sample fine-tuning dataset. This dataset uniformly adopts a six-dimensional field format: operating condition description, feature data, fault evolution path, thought chain reasoning, handling plan, and mechanism verification. Specifically, it is divided into four categories: state assessment question-and-answer pairs, fault diagnosis thought chain data pairs, fault evolution prediction data pairs, and operation and maintenance handling and simulation verification data pairs, providing a high-quality training foundation for fine-tuning large models in vertical domains.
[0038] In step S2 above, the parameter-efficient fine-tuning framework combined with contrastive learning is used to train the model as follows: The parameter-efficient fine-tuning framework with LoRA low-rank adaptation is adopted, the backbone parameters of the base model are frozen, and only the low-rank matrix is trained, which greatly reduces the data requirements and computing power costs for fine-tuning; at the same time, contrastive learning loss is added to the fine-tuning loss function, which brings the feature distance of similar fault samples closer and pushes the feature distance of dissimilar samples further apart in the feature space, solves the problem of poor generalization ability of the model under small sample size, and significantly improves the model's ability to distinguish and identify early weak fault features.
[0039] In step S2 above, the optimization method of the time-aware attention mechanism is as follows: a time-aware attention branch is added to the Transformer architecture to focus on the trend changes, mutation features, periodic features and degradation slope of long time-series data, and adaptively improve the attention weight of weak features of early faults.
[0040] To address the shortcomings of general large models in terms of insufficient sensitivity to long-term operational data and inability to capture slow degradation features, a time-aware attention branch is added to the Transformer architecture. This branch focuses on the trend changes, abrupt changes, periodic features, and degradation slope of long-term data with weights, adaptively increasing the attention weight of subtle changes in early fault features. This solves the core problem of low signal-to-noise ratio and difficulty in identifying early fault features in vertical shaft axial flow pumps.
[0041] In step S2 above, the illusion suppression module with dual constraints of mechanism and knowledge graph embeds a dual verification process based on knowledge graph and mechanism model at the model inference output layer: every inference result, diagnostic conclusion and treatment suggestion generated by the model must be matched with entities and relationships in the knowledge graph, and the engineering rationality is verified by the mechanism model; content that does not conform to the mechanism law or exceeds the boundary of the knowledge graph is automatically filtered, corrected and labeled with confidence, suppressing professional knowledge illusion from the root and ensuring that the model output fully conforms to the actual water conservancy project.
[0042] Step S3 above breaks through the limitations of existing technologies such as simple question-and-answer and isolated early warning, and constructs a closed-loop operation and maintenance logic for the entire process, realizing intelligent operation and maintenance throughout the entire lifecycle from early fault prediction to iterative optimization of handling effects.
[0043] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes: early fault intelligent early warning and evolution prediction, specifically: real-time access to spatiotemporally aligned operation data, capturing weak degradation characteristics of equipment and triggering graded early warnings, predicting the fault degradation trend of 72 hours and above based on the fault evolution model, and outputting fault probability, time window and risk level.
[0044] In other words, the system accesses multi-dimensional operational data in real time after spatiotemporal alignment. Through a finely tuned operation and maintenance model, it captures subtle changes in the operating status of equipment in real time, identifies early fault characteristics, and automatically triggers graded early warnings. At the same time, based on the fault evolution model, it predicts the deterioration trend of the fault in the next 72 hours or more, and gives the probability, time window and risk level of the fault occurrence, realizing the upgrade from "over-limit alarm" to "early prediction".
[0045] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes multi-dimensional coupled fault root cause diagnosis and inversion, specifically: for early warning faults, multi-parameter coupled correlation analysis is carried out by combining knowledge graphs, mechanism models and historical data to locate the root cause of the fault and distinguish between direct and indirect causes; the entire evolution path of the fault is inverted to clarify the key nodes of operation and maintenance management.
[0046] For early warning faults, the model combines knowledge graphs, mechanistic models, and historical data to conduct multi-parameter coupled correlation analysis, accurately locate the root cause of the fault, and distinguish between direct and indirect causes. At the same time, through the causal reasoning capabilities of the large model, the entire evolution process of the fault from its inception to its development is reversed, the fault development path is restored, the key nodes of operation and maintenance management are identified, and the recurrence of similar faults is prevented.
[0047] The fully closed-loop intelligent operation and maintenance system in step S3 above includes the generation of operation and maintenance disposal plans and digital twin simulation verification. Specifically, it involves generating graded and scenario-based disposal plans by combining fault diagnosis results, equipment operating conditions and pump station operation plans; and verifying the disposal plans through simulation of the vertical shaft axial flow pump digital twin model to predict the operating status and disposal effect.
[0048] The aforementioned model generates graded and scenario-based operation and maintenance solutions based on fault diagnosis results, current equipment operating conditions, and pump station operation plans (such as flood season drainage and water resource scheduling). These solutions include different levels such as emergency shutdown, online operating condition adjustment, planned maintenance, and temporary operation control. Simultaneously, it combines a digital twin model of the vertical shaft axial flow pump to simulate and verify the solutions, predict the equipment operating status, risk points, and expected effects after the solutions are implemented, and ensure the feasibility and safety of the solutions, avoiding secondary failures caused by blind implementation.
[0049] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes feedback on the handling effect and system self-iterative optimization. Specifically, it automatically collects operation data after handling and evaluates the handling effect; automatically updates new fault cases and handling data to the knowledge base and knowledge graph; and regularly performs incremental fine-tuning of the model to achieve system self-learning and self-optimization.
[0050] After the above-mentioned operation and maintenance procedures are completed, the system automatically collects the continuous operation data of the equipment after the procedures are completed, automatically evaluates the effectiveness of the procedures, and forms a complete operation and maintenance closed loop. At the same time, the newly added fault cases, procedures, effect data, and expert review conclusions are automatically updated into the knowledge base and knowledge graph. The model is incrementally fine-tuned regularly to achieve continuous self-learning and self-optimization of the system's operation and maintenance capabilities, solving the problem of knowledge solidification and inability to continuously evolve in traditional expert systems.
[0051] In step S3 above, the interactive full-scenario operation and maintenance system in natural language allows operation and maintenance personnel to initiate inspection queries, fault diagnosis, working condition simulation, ledger management and skills training instructions in natural language through the terminal. The system automatically generates natural language responses and visual reports that conform to water conservancy engineering specifications.
[0052] With natural language as the core human-computer interaction interface, the system supports natural language interaction for maintenance personnel in all scenarios, including daily inspections, fault handling, skills training, ledger management, and operating condition simulations. Maintenance personnel can directly initiate queries and commands using natural language via mobile phones and computers, such as "summarize the operating status and health trend of pump No. 3 this week" or "simulate the operating status of pump No. 2 when the guide bearing wear is 0.2mm under high water level conditions during the flood season, and provide pre-control measures." The system automatically calls upon models and knowledge bases to generate natural language responses and visual reports that conform to engineering specifications.
[0053] A fully closed-loop intelligent operation and maintenance system for large vertical shaft axial flow pumps based on a large language model, used to execute the above method, includes:
[0054] The data preprocessing module is used for spatiotemporal synchronization, noise reduction, and feature extraction of multi-source heterogeneous data;
[0055] The knowledge base construction module is used to build a dedicated knowledge graph that integrates mechanisms, knowledge, and data.
[0056] The model training module is used for fine-tuning large models with small samples, attention optimization, and hallucination suppression.
[0057] The closed-loop operation and maintenance module is used for fault early warning, root cause diagnosis, solution generation, simulation verification and self-iteration.
[0058] The human-computer interaction module is used for operation and maintenance interaction in all scenarios using natural language.
[0059] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model.
[0060] This invention solves the following technical problems: It addresses the issues of large differences in sampling frequency, asynchronous timing, and poor spatiotemporal coupling of multi-source heterogeneous data from large vertical shaft axial flow pumps, leading to distorted fault feature extraction and inability to identify early faults; it addresses the problems of weak early fault features, low signal-to-noise ratio, and scarce fault samples in vertical shaft axial flow pumps, resulting in ineffective capture by traditional methods and general-purpose large models, delayed early warnings, and high false negative rates; it addresses the problems of severe professional knowledge illusion in the vertical field of water conservancy pumps by general-purpose large-scale language models, where inference results do not conform to engineering mechanisms and actual operation and maintenance, leading to insufficient reliability of operation and maintenance decisions; and it addresses the problems of existing technologies having fragmented operation and maintenance processes, only achieving simple early warning and Q&A, failing to form a complete closed-loop operation and maintenance process of "early warning-diagnosis-handling-simulation-feedback-iteration," and lacking continuous self-learning capabilities.
[0061] Explanation of relevant technical terms:
[0062] Vertical shaft axial flow pump: a type of horizontal axial flow pump with low head and large flow rate. The pump body and motor are coaxially installed in the vertical shaft. It features a compact structure, convenient installation and maintenance, and high hydraulic efficiency. It is widely used in water conservancy projects such as flood control and drainage, and water resource allocation.
[0063] Large Language Model: A deep learning model trained from massive amounts of text data, possessing powerful natural language understanding, generation, and complex logical reasoning capabilities, and is the core intelligent carrier of this invention.
[0064] Knowledge graph: a technique that uses graph structures to describe semantic relationships between entities, consisting of "entity-relationship-entity" triples. In this invention, it provides a structured professional domain knowledge foundation for large models.
[0065] LoRA Low-Rank Adaptation: A parameter-efficient fine-tuning method that freezes the backbone parameters of a pre-trained large model and trains only the low-rank matrix of the Transformer layer, significantly reducing the computational and data requirements for fine-tuning.
[0066] Contrastive learning: a self-supervised learning method that improves the model's feature discrimination ability by allowing the model to learn the commonalities of samples of the same class and the differences of samples of different classes. In this invention, it is used to solve the problem of poor generalization ability of the model under small sample conditions.
[0067] SCADA system: Data acquisition and monitoring control system, widely used in the automated data acquisition and equipment monitoring of industrial infrastructure.
[0068] Any techniques not mentioned in this invention are based on existing technologies.
[0069] Compared with the prior art, the present invention has the following beneficial effects:
[0070] Improve the accuracy and timeliness of fault diagnosis: By integrating multi-dimensional data and domain knowledge, the model can discover data correlations and early fault symptoms that are difficult for the human eye to detect, realizing the transformation from "post-event alarm" to "pre-event prediction" and significantly reducing unplanned downtime.
[0071] The illusion of professional knowledge is fundamentally suppressed: through the knowledge graph and mechanism constraint verification module that integrates "mechanism-knowledge-data", the accuracy rate of water conservancy professional knowledge Q&A is ≥98%, and the operation and maintenance decisions output by the model are fully consistent with the engineering mechanism and the actual situation on site, and the reliability of decision-making is greatly improved.
[0072] Lowering the barrier to operation and maintenance and improving efficiency: Natural language interaction allows operation and maintenance personnel to obtain accurate operation and maintenance guidance without having in-depth data analysis or expert system operation knowledge, significantly improving work efficiency and quality.
[0073] Achieving the accumulation and evolution of operational knowledge: The knowledge base and models can be continuously updated and optimized with the accumulation of new data and new cases, so that the "wisdom" of the entire system continues to grow, avoiding the problems of knowledge solidification and difficulty in updating in traditional expert systems. Attached Figure Description
[0074] Figure 1 This is the overall flowchart of a method for constructing multi-dimensional intelligent operation and maintenance of water pumps based on a large language model according to the present invention.
[0075] Figure 2 This is a schematic diagram illustrating the construction of the multi-dimensional knowledge base for water pumps in this invention.
[0076] Figure 3 This is a system architecture block diagram of the present invention. Detailed Implementation
[0077] To better understand the present invention, the following embodiments further illustrate the content of the present invention, but the content of the present invention is not limited to the following embodiments.
[0078] like Figures 1-3 As shown, a method for constructing multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model includes the following steps:
[0079] S1 Constructs a multi-dimensional knowledge base for large vertical shaft axial flow pumps that deeply integrates mechanism, knowledge, and data: It performs spatiotemporal synchronization and preprocessing of multi-source heterogeneous data, designs a knowledge graph ontology architecture for large vertical shaft axial flow pumps, embeds hydraulic and fault evolution mechanism models, and completes the three-dimensional deep integration of mechanism, knowledge, and time-series data.
[0080] S2 Shaft Axial Flow Pump Operation and Maintenance Large Language Model Small Sample Enhancement Fine-tuning and Illusion Suppression: Construct a multi-turn dialogue-based small sample fine-tuning dataset, adopt a parameter-efficient fine-tuning framework combined with contrastive learning to train the model, optimize the temporal awareness attention mechanism, and add an illusion suppression module with dual constraints of mechanism and knowledge graph to obtain a vertical domain-specific large operation and maintenance model;
[0081] S3 constructs a fully closed-loop intelligent operation and maintenance system: based on a dedicated operation and maintenance big model and a multi-dimensional knowledge base, it realizes closed-loop operation and maintenance of large vertical shaft axial flow pumps throughout the entire life cycle, including early warning, diagnosis, handling, simulation, feedback and iteration, and supports interactive operation and maintenance management in all scenarios using natural language.
[0082] Among them, the large vertical shaft axial flow pump is a vertical shaft axial flow pump unit with a single unit rated flow rate ≥25m³ / s, a single unit power ≥1000kW, and an impeller diameter ≥2000mm.
[0083] In step S1 above, this step breaks through the limitations of simple knowledge extraction in existing technologies, designs a knowledge graph ontology architecture dedicated to vertical shaft axial flow pumps, integrates hydraulic mechanism models and fault evolution models, realizes spatiotemporal synchronization and deep fusion of multi-source data, and provides a professional knowledge base without illusions for model reasoning.
[0084] In step S1 above, the specific method for spatiotemporal synchronization and preprocessing of multi-source heterogeneous data is as follows: Addressing the issues of temporal asynchrony and feature coupling distortion caused by large differences in sampling frequencies among multiple sensors in a vertical shaft axial flow pump, a spatiotemporal alignment algorithm combining cubic spline interpolation resampling with cross-correlation delay estimation is employed to unify the data from vibration sensors, temperature sensors, pressure sensors, electrical sensors, hydraulic sensors, equipment status sensors, and environmental auxiliary sensors to a 1ms time base, achieving a temporal synchronization error ≤0.5ms. An adaptive wavelet denoising algorithm is used to filter operating noise and retain weak early fault features, completing data cleaning, normalization, and feature extraction to generate a standardized time-series dataset.
[0085] In step S1 above, the multi-source heterogeneous data includes structured operation data and unstructured operation and maintenance data; the structured operation data includes vibration acceleration, vibration velocity, vibration displacement, bearing temperature, winding temperature, cooling water temperature, inlet and outlet pressure, flow rate, head, voltage, current, power, speed and upstream and downstream water levels; the unstructured operation and maintenance data includes equipment design drawings, operation manuals, maintenance records, fault cases, expert experience and water conservancy industry standards and specifications.
[0086] The knowledge graph ontology architecture for vertical shaft axial flow pumps in step S1 above breaks through the limitations of general knowledge graphs. It adopts a 6-level entity architecture, 12 proprietary semantic relationships and 28 engineering attribute dimensions, specifically for the operation and maintenance characteristics of vertical shaft axial flow pumps.
[0087] The 6-layer entity architecture consists of: system layer, component layer, failure mode layer, symptom and characteristic layer, root cause layer, and remedial measures layer.
[0088] The 12 proprietary semantic relationships include 6 basic association relationships and 6 core relationships specific to vertical shaft axial flow pumps, which fully depict the dynamic evolution of faults. The 6 basic association relationships include: composition relationship, fault occurrence, fault manifestation, cause, treatment measures, and execution specifications. The 6 core relationships specific to vertical shaft axial flow pumps include: evolution sequence, fault coupling relationship, mechanism constraint, treatment priority, operating condition adaptability, and degradation trend.
[0089] The 28 engineering attribute dimensions include 28 exclusive attributes for each entity, such as design operating condition parameters, operating condition boundaries, fault evolution time thresholds, fault confidence, operation and maintenance history records, and handling effect evaluation, to achieve accurate matching between knowledge and engineering practice.
[0090] The above 6-level entity architecture includes the following dedicated entity levels: System layer (vertical shaft axial flow pump unit) → Component layer (impeller, main shaft, thrust bearing, guide bearing, synchronous motor, inlet and outlet water channels, coupling, cooling and lubrication system, mechanical seal) → Failure mode layer (blade cavitation, rotor misalignment, rotor imbalance, bearing wear, winding overheating, seal leakage, flow channel blockage, etc.) → Symptom characteristic layer (excessive vibration, abnormal temperature, flow fluctuation, current fluctuation, efficiency reduction, etc.) → Root cause layer (installation deviation, poor lubrication, silt wear, foreign object blockage, cavitation erosion, operating conditions deviating from design range, etc.) → Remedial measures layer (emergency shutdown, online operating condition adjustment, planned maintenance, component replacement, lubrication maintenance, etc.).
[0091] Among the above 12 proprietary semantic relationships, the basic association relationships include: composition relationship, fault occurrence, fault manifestation, cause, handling measures, and execution norm.
[0092] The aforementioned 28 engineering attribute dimensions include: design operating condition parameters, operating condition boundaries, fault evolution time threshold, fault confidence, maintenance history, handling effect evaluation, equipment geometric dimension parameters, component material performance parameters, installation and assembly accuracy threshold, rated hydraulic performance index, cavitation critical performance threshold, flow channel resistance characteristic parameters, efficiency attenuation allowable limit, vibration threshold judgment standard, spindle running swing limit, bearing lubrication performance parameters, motor electrical rated parameters, insulation performance safety threshold, equipment health status level, fault severity level, fault occurrence probability weight, secondary fault association risk, equipment remaining service life, maintenance cycle calibration parameters, spare parts compatibility model, maintenance operation safety procedures, industry standard compliance indicators, and data collection and sampling specifications.
[0093] In step S1 above, the embedded mechanism model includes a hydraulic mechanism model and a rotor dynamics and fault evolution model;
[0094] The hydraulic mechanism model includes: head-flow characteristic equation, shaft power calculation model, required net positive suction head (NPSH) model, pipeline resistance characteristic model, and efficiency characteristic model;
[0095] The rotor dynamics and fault evolution models include a bearing wear degradation trend model, a misalignment fault vibration characteristic evolution model, a cavitation development stage model, and a spindle fatigue life prediction model.
[0096] The method for deep integration of mechanism, knowledge and data in step S1 is as follows: the hydraulic mechanism model and rotor dynamics and fault evolution model are embedded into the knowledge graph of the vertical shaft axial flow pump; the named entity recognition and relation extraction method is used to automatically extract operation and maintenance entities and semantic relations from multi-source heterogeneous data, and the extraction results are deeply integrated with the mechanism model to construct a structured operation and maintenance knowledge graph, forming a three-element closed-loop fusion system of mechanism constraint knowledge, knowledge interpretation data and data optimization knowledge.
[0097] In step S1 above, the core mechanism model of the vertical shaft axial flow pump is embedded into the knowledge graph to set hard constraints for knowledge reasoning, thereby suppressing the illusion of professional knowledge of large models from the root and realizing the upgrade from static knowledge to dynamic evolutionary knowledge.
[0098] In step S1 above, after completing the construction of the dedicated knowledge graph ontology architecture for the vertical shaft axial flow pump, a deep integration of mechanism, knowledge, and data is carried out: First, core mechanism models such as head-flow characteristics, bearing wear degradation, and cavitation development are embedded into the knowledge graph to set hard constraints on engineering mechanisms for knowledge reasoning; further, named entity recognition and relation extraction artificial intelligence technologies are used to accurately extract equipment entities, fault entities, and corresponding semantic relationships from structured operation data and unstructured operation and maintenance documents; the extracted knowledge entities, semantic relationships, and mechanism models are deeply coupled and normalized to construct a structured and reasonable knowledge graph for the operation and maintenance of the vertical shaft axial flow pump, ultimately forming a three-element fusion system of mechanism constraint knowledge, knowledge interpretation data, and data optimization knowledge, realizing the upgrade from static operation and maintenance knowledge to dynamic evolutionary knowledge.
[0099] Step S2 above overcomes the limitations of general fine-tuning methods. Addressing the issues of scarce pump failure samples, difficulty in identifying early weak features, and severe professional knowledge illusion, it proposes a small-sample fine-tuning method that integrates contrastive learning with the Prompt thought chain. It optimizes the temporal awareness attention mechanism, incorporates a mechanism-constrained illusion suppression module, and constructs a highly reliable large-scale operation and maintenance model for vertical shaft axial flow pumps. The specific sub-steps are as follows:
[0100] In step S2 above, a small-sample fine-tuning dataset is constructed: Addressing the scarcity of real fault samples for vertical shaft axial flow pumps, simulation samples of the entire fault lifecycle are generated based on knowledge graphs and mechanism models. These are then integrated with real historical fault samples to construct a multi-turn dialogue-based small-sample fine-tuning dataset with core fields including operating condition description, feature data, fault evolution path, thought chain reasoning, handling solutions, and mechanism verification, replacing the traditional single question-and-answer pair dataset. The core types of this dataset include:
[0101] Status assessment Q&A pairs: Integrate real-time time series data with operating condition boundaries to achieve multi-dimensional assessment of equipment health status;
[0102] Fault Diagnosis Thinking Chain Data Pair: Decomposes fault diagnosis into a complete thinking chain of feature recognition, operating condition matching, mechanism verification, fault location, root cause analysis, and handling suggestions to improve the interpretability of model reasoning;
[0103] Fault evolution prediction data pairs: Based on the fault degradation model, construct the correspondence between fault characteristics and evolution trends at different stages, and train the model's long-term prediction capability;
[0104] Operation and maintenance handling and simulation verification data pair: Match the hierarchical handling plan, combine the digital twin simulation results, and construct the plan effectiveness verification data.
[0105] To address the aforementioned technical problem of scarce fault samples in vertical shaft axial flow pumps, this invention uses a knowledge graph as the knowledge base and a mechanism model as the constraint boundary to generate, in batches, simulated fault evolution samples covering the entire degradation stage. These simulated samples are then balanced and fused with real historical fault samples collected from the pump station site to construct a structured, multi-turn, dialogue-based small-sample fine-tuning dataset. This dataset uniformly adopts a six-dimensional field format: operating condition description, feature data, fault evolution path, thought chain reasoning, handling plan, and mechanism verification. Specifically, it is divided into four categories: state assessment question-and-answer pairs, fault diagnosis thought chain data pairs, fault evolution prediction data pairs, and operation and maintenance handling and simulation verification data pairs, providing a high-quality training foundation for fine-tuning large models in vertical domains.
[0106] In step S2 above, the parameter-efficient fine-tuning framework combined with contrastive learning is used to train the model as follows: The parameter-efficient fine-tuning framework with LoRA low-rank adaptation is adopted, the backbone parameters of the base model are frozen, and only the low-rank matrix is trained, which greatly reduces the data requirements and computing power costs for fine-tuning; at the same time, contrastive learning loss is added to the fine-tuning loss function, which brings the feature distance of similar fault samples closer and pushes the feature distance of dissimilar samples further apart in the feature space, solves the problem of poor generalization ability of the model under small sample size, and significantly improves the model's ability to distinguish and identify early weak fault features.
[0107] In step S2 above, the optimization method of the time-aware attention mechanism is as follows: a time-aware attention branch is added to the Transformer architecture to focus on the trend changes, mutation features, periodic features and degradation slope of long time-series data, and adaptively improve the attention weight of weak features of early faults.
[0108] To address the shortcomings of general large models in terms of insufficient sensitivity to long-term operational data and inability to capture slow degradation features, a time-aware attention branch is added to the Transformer architecture. This branch focuses on the trend changes, abrupt changes, periodic features, and degradation slope of long-term data with weights, adaptively increasing the attention weight of subtle changes in early fault features. This solves the core problem of low signal-to-noise ratio and difficulty in identifying early fault features in vertical shaft axial flow pumps.
[0109] In step S2 above, the illusion suppression module with dual constraints of mechanism and knowledge graph embeds a dual verification process based on knowledge graph and mechanism model at the model inference output layer: every inference result, diagnostic conclusion and treatment suggestion generated by the model must be matched with entities and relationships in the knowledge graph, and the engineering rationality is verified by the mechanism model; content that does not conform to the mechanism law or exceeds the boundary of the knowledge graph is automatically filtered, corrected and labeled with confidence, suppressing professional knowledge illusion from the root and ensuring that the model output fully conforms to the actual water conservancy project.
[0110] Step S3 above breaks through the limitations of existing technologies such as simple question-and-answer and isolated early warning, and constructs a closed-loop operation and maintenance logic for the entire process, realizing intelligent operation and maintenance throughout the entire lifecycle from early fault prediction to iterative optimization of handling effects.
[0111] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes: early fault intelligent early warning and evolution prediction, specifically: real-time access to spatiotemporally aligned operation data, capturing weak degradation characteristics of equipment and triggering graded early warnings, predicting the fault degradation trend of 72 hours and above based on the fault evolution model, and outputting fault probability, time window and risk level.
[0112] In other words, the system accesses multi-dimensional operational data in real time after spatiotemporal alignment. Through a finely tuned operation and maintenance model, it captures subtle changes in the operating status of equipment in real time, identifies early fault characteristics, and automatically triggers graded early warnings. At the same time, based on the fault evolution model, it predicts the deterioration trend of the fault in the next 72 hours or more, and gives the probability, time window and risk level of the fault occurrence, realizing the upgrade from "over-limit alarm" to "early prediction".
[0113] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes multi-dimensional coupled fault root cause diagnosis and inversion, specifically: for early warning faults, multi-parameter coupled correlation analysis is carried out by combining knowledge graphs, mechanism models and historical data to locate the root cause of the fault and distinguish between direct and indirect causes; the entire evolution path of the fault is inverted to clarify the key nodes of operation and maintenance management.
[0114] For early warning faults, the model combines knowledge graphs, mechanistic models, and historical data to conduct multi-parameter coupled correlation analysis, accurately locate the root cause of the fault, and distinguish between direct and indirect causes. At the same time, through the causal reasoning capabilities of the large model, the entire evolution process of the fault from its inception to its development is reversed, the fault development path is restored, the key nodes of operation and maintenance management are identified, and the recurrence of similar faults is prevented.
[0115] The fully closed-loop intelligent operation and maintenance system in step S3 above includes the generation of operation and maintenance disposal plans and digital twin simulation verification. Specifically, it involves generating graded and scenario-based disposal plans by combining fault diagnosis results, equipment operating conditions and pump station operation plans; and verifying the disposal plans through simulation of the vertical shaft axial flow pump digital twin model to predict the operating status and disposal effect.
[0116] The aforementioned model generates graded and scenario-based operation and maintenance solutions based on fault diagnosis results, current equipment operating conditions, and pump station operation plans (such as flood season drainage and water resource scheduling). These solutions include different levels such as emergency shutdown, online operating condition adjustment, planned maintenance, and temporary operation control. Simultaneously, it combines a digital twin model of the vertical shaft axial flow pump to simulate and verify the solutions, predict the equipment operating status, risk points, and expected effects after the solutions are implemented, and ensure the feasibility and safety of the solutions, avoiding secondary failures caused by blind implementation.
[0117] In step S3 above, the fully closed-loop intelligent operation and maintenance system includes feedback on the handling effect and system self-iterative optimization. Specifically, it automatically collects operation data after handling and evaluates the handling effect; automatically updates new fault cases and handling data to the knowledge base and knowledge graph; and regularly performs incremental fine-tuning of the model to achieve system self-learning and self-optimization.
[0118] After the above-mentioned operation and maintenance procedures are completed, the system automatically collects the continuous operation data of the equipment after the procedures are completed, automatically evaluates the effectiveness of the procedures, and forms a complete operation and maintenance closed loop. At the same time, the newly added fault cases, procedures, effect data, and expert review conclusions are automatically updated into the knowledge base and knowledge graph. The model is incrementally fine-tuned regularly to achieve continuous self-learning and self-optimization of the system's operation and maintenance capabilities, solving the problem of knowledge solidification and inability to continuously evolve in traditional expert systems.
[0119] In step S3 above, the interactive full-scenario operation and maintenance system in natural language allows operation and maintenance personnel to initiate inspection queries, fault diagnosis, working condition simulation, ledger management and skills training instructions in natural language through the terminal. The system automatically generates natural language responses and visual reports that conform to water conservancy engineering specifications.
[0120] With natural language as the core human-computer interaction interface, the system supports natural language interaction for maintenance personnel in all scenarios, including daily inspections, fault handling, skills training, ledger management, and operating condition simulations. Maintenance personnel can directly initiate queries and commands using natural language via mobile phones and computers, such as "summarize the operating status and health trend of pump No. 3 this week" or "simulate the operating status of pump No. 2 when the guide bearing wear is 0.2mm under high water level conditions during the flood season, and provide pre-control measures." The system automatically calls upon models and knowledge bases to generate natural language responses and visual reports that conform to engineering specifications.
[0121] A fully closed-loop intelligent operation and maintenance system for large vertical shaft axial flow pumps based on a large language model, used to execute the above method, includes:
[0122] The data preprocessing module is used for spatiotemporal synchronization, noise reduction, and feature extraction of multi-source heterogeneous data;
[0123] The knowledge base construction module is used to build a dedicated knowledge graph that integrates mechanisms, knowledge, and data.
[0124] The model training module is used for fine-tuning large models with small samples, attention optimization, and hallucination suppression.
[0125] The closed-loop operation and maintenance module is used for fault early warning, root cause diagnosis, solution generation, simulation verification and self-iteration.
[0126] The human-computer interaction module is used for operation and maintenance interaction in all scenarios using natural language.
[0127] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model.
[0128] Application Example: Digital Twin System of the Jiyun River Estuary Pumping Station in Tianjin
[0129] The Tianjin Jiyun River Estuary Pumping Station is a core hub for flood control and drainage in the basin. It is equipped with six large vertical shaft axial flow pumps, each with a design flow rate of 50 m³ / s, primarily responsible for flood drainage during the flood season and water resource allocation during the non-flood season. The above-described scheme in this embodiment constructs a fully intelligent operation and maintenance system for this pumping station.
[0130] Multi-dimensional knowledge base construction
[0131] For the vertical shaft axial flow pump of this pumping station, historical operating data from the past 5 years were collected, including 12 vibration measurement points (20kHz sampling), 8 temperature measurement points (1Hz sampling), 4 pressure measurement points (10Hz sampling), and full-dimensional operating data such as flow rate, head, voltage and current, and upstream and downstream water levels. At the same time, unstructured data such as equipment design manuals, 86 maintenance reports, 12 operating specifications, 68 failure cases, and industry standards were also collected.
[0132] A spatiotemporal alignment algorithm based on cubic spline interpolation resampling and cross-correlation delay estimation was adopted to unify all data to a 1ms time base, with a timing synchronization error of ≤0.4ms. Data preprocessing was completed through adaptive wavelet denoising. A dedicated knowledge graph ontology architecture adapted to the vertical shaft axial flow pump of the pumping station was designed, embedding 8 core mechanism models such as head-flow characteristics, bearing wear degradation, and cavitation evolution. The final knowledge graph contains 1260 entities and 3820 relation triples, realizing the deep integration of mechanism, knowledge, and data.
[0133] Fine-tuning and optimization of the large-scale operation and maintenance model
[0134] Based on real fault samples from the pump station, multiple fault full-cycle evolution samples were generated by mechanistic model simulation, and a fine-tuning dataset with a complete thought chain was constructed. The LoRA+contrastive learning small-sample fine-tuning strategy was adopted to efficiently fine-tune the parameters of the Deepseek-32B base model. At the same time, a time-aware attention branch was added, and a mechanistic constraint illusion suppression module was embedded. Finally, the training and deployment of the dedicated operation and maintenance model for the vertical shaft axial flow pump were completed.
[0135] Deployment and Application Effects of the Fully Closed-Loop Operation and Maintenance System
[0136] The system officially went online in May 2026, achieving 24 / 7 intelligent operation and maintenance of four vertical shaft axial flow pumps. The core application effects are as follows:
[0137] Early Fault Warning Case: Before the flood season, the system monitored that the temperature of the guide bearing of the No. 2 vertical shaft axial flow pump was slowly rising at a rate of 0.3℃ per hour. At the same time, the signal-to-noise ratio of the high-frequency vibration component (10-20kHz) increased by 3dB. The model identified this as an early wear fault in the guide bearing and triggered a yellow warning 78 hours in advance (the current mainstream technology has a maximum warning advance of only 48 hours, and most traditional SCADA systems can only achieve over-limit alarms, without early prediction capabilities). It predicted that the bearing temperature would exceed the limit after 72 hours and generated a diagnostic report: "Warning: Early wear of the guide bearing of pump No. 2, confidence level 96%. Root cause: Frequent start-stop conditions before the flood season led to poor lubrication film formation and wear deterioration. Recommendation: Stop the machine to check the bearing wear, replace the lubricating oil, and adjust the lubrication control logic under start-stop conditions." Maintenance personnel shut down the machine for inspection and found that the guide bearing was worn by 0.15mm (this early wear is only 1 / 5 of the accuracy of conventional detection with existing technology, and the existing vibration monitoring system cannot capture such a weak wear feature). After taking the recommended measures, the equipment returned to normal operation, avoiding an unplanned shutdown accident during the critical period of flood drainage.
[0138] Quantitative operational indicators: After 12 months of operation, the system has achieved 3 early fault warnings, with warning lead times exceeding 72-78 hours (far exceeding the maximum warning lead time of 48 hours for existing technologies, completely solving the core problem of existing technologies that "can only alarm for exceeding limits but cannot predict in advance"). The fault diagnosis accuracy rate is 96.5%, and the false alarm rate is reduced by 68.3% compared to the original SCADA system. The accuracy rate of professional knowledge Q&A and operation and maintenance decision generation reaches 98.7%, with no serious professional knowledge illusion problems. A total of 126 hours of unplanned downtime has been reduced, the daily inspection efficiency of pump station operation and maintenance personnel has increased by 55%, and the training cycle for new employees has been shortened by 62%, achieving significant economic and safety benefits.
[0139] This invention presents a knowledge graph ontology architecture specifically for the operation and maintenance of large vertical shaft axial flow pumps. It achieves deep integration of hydraulic mechanism models, fault evolution models, and knowledge graphs, constructing a knowledge base driven by a "mechanism-knowledge-data" ternary model. This fundamentally solves the problem of severe professional knowledge illusion and poor adaptability of general-purpose large models in the vertical field of hydraulic pumps. Based on a multi-source heterogeneous data spatiotemporal synchronization method using cubic spline interpolation resampling and cross-correlation delay estimation, it solves the problem of fault feature coupling distortion caused by large differences in sampling frequencies and temporal asynchrony among multiple sensors in large vertical shaft axial flow pumps. Simultaneously, adaptive wavelet denoising fully preserves weak early fault features, providing a high-quality data foundation for fault identification. The small-sample fine-tuning method integrating comparative learning and the Prompt thought chain, combined with temporal-aware attention mechanism optimization, solves the core problems of scarce fault samples and difficulty in identifying weak early fault features in hydraulic pumps, significantly improving the model's fault warning lead time, diagnostic accuracy, and generalization ability. The "early warning-diagnosis-handling-simulation-feedback-iteration" fully closed-loop intelligent operation and maintenance method based on a large language model achieves a closed-loop process encompassing fault evolution prediction, root cause inversion, digital twin simulation verification of handling solutions, and system self-optimization. This solves the problems of fragmented operation and maintenance processes and the inability to continuously iterate and evolve in existing technologies. The large-model professional knowledge illusion suppression method, which embeds mechanistic constraints, incorporates dual verification of knowledge graphs and mechanistic models in the model reasoning stage. It automatically filters and corrects outputs that do not conform to engineering principles, achieving an accuracy rate of ≥98% for water conservancy professional knowledge question answering, significantly improving the reliability of operation and maintenance decisions.
Claims
1. A method for constructing multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model, characterized in that: Includes the following steps: S1 Constructs a multi-dimensional knowledge base for large vertical shaft axial flow pumps that deeply integrates mechanism, knowledge, and data: It performs spatiotemporal synchronization and preprocessing of multi-source heterogeneous data, designs a knowledge graph ontology architecture for large vertical shaft axial flow pumps, embeds hydraulic and fault evolution mechanism models, and completes the three-dimensional deep integration of mechanism, knowledge, and time-series data. S2 Shaft Axial Flow Pump Operation and Maintenance Large Language Model Small Sample Enhancement Fine-tuning and Illusion Suppression: Construct a multi-turn dialogue-based small sample fine-tuning dataset, adopt a parameter-efficient fine-tuning framework combined with contrastive learning to train the model, optimize the temporal awareness attention mechanism, and add an illusion suppression module with dual constraints of mechanism and knowledge graph to obtain a vertical domain-specific large operation and maintenance model; S3 constructs a fully closed-loop intelligent operation and maintenance system: based on a dedicated operation and maintenance big model and a multi-dimensional knowledge base, it realizes closed-loop operation and maintenance of large vertical shaft axial flow pumps throughout the entire life cycle, including early warning, diagnosis, handling, simulation, feedback and iteration, and supports interactive operation and maintenance management in all scenarios using natural language. Among them, the large vertical shaft axial flow pump is a vertical shaft axial flow pump unit with a single unit rated flow rate ≥25m³ / s, a single unit power ≥1000kW, and an impeller diameter ≥2000mm.
2. The method according to claim 1, characterized in that, In step S1, the specific method for spatiotemporal synchronization and preprocessing of multi-source heterogeneous data is as follows: a spatiotemporal alignment algorithm combining cubic spline interpolation resampling with cross-correlation delay estimation is used to unify the data from vibration sensors, temperature sensors, pressure sensors, electrical sensors, hydraulic sensors, equipment status sensors, and environmental auxiliary sensors to a 1ms time base, with a time synchronization error ≤0.5ms; an adaptive wavelet denoising algorithm is used to filter operating noise and retain weak features of early faults, completing data cleaning, normalization, and feature extraction to generate a standardized time series dataset; In step S1, the multi-source heterogeneous data includes structured operation data and unstructured operation and maintenance data; the structured operation data includes vibration acceleration, vibration velocity, vibration displacement, bearing temperature, winding temperature, cooling water temperature, inlet and outlet pressure, flow rate, head, voltage, current, power, speed and upstream and downstream water levels; the unstructured operation and maintenance data includes equipment design drawings, operation manuals, maintenance records, fault cases, expert experience and water conservancy industry standards and specifications.
3. The method according to claim 1 or 2, characterized in that, The knowledge graph ontology architecture for the vertical shaft axial flow pump in step S1 adopts a 6-level entity architecture, 12 proprietary semantic relationships, and 28 engineering attribute dimensions; The 6-layer entity architecture consists of: system layer, component layer, failure mode layer, symptom and characteristic layer, root cause layer, and remedial measures layer. The 12 proprietary semantic relationships include 6 basic association relationships and 6 core relationships specific to vertical shaft axial flow pumps, which fully depict the dynamic evolution of faults. The 6 basic association relationships include: composition relationship, fault occurrence, fault manifestation, cause, treatment measures, and execution specifications. The 6 core relationships specific to vertical shaft axial flow pumps include: evolution sequence, fault coupling relationship, mechanism constraint, treatment priority, operating condition adaptability, and degradation trend. The 28 engineering attribute dimensions include design operating condition parameters, operating condition boundaries, fault evolution time thresholds, fault confidence, operation and maintenance history records, and evaluation of handling effects.
4. The method according to claim 1 or 2, characterized in that, In step S1, the embedded mechanism model includes a hydraulic mechanism model and a rotor dynamics and fault evolution model; The hydraulic mechanism model includes: head-flow characteristic equation, shaft power calculation model, required net positive suction head (NPSH) model, pipeline resistance characteristic model, and efficiency characteristic model; The rotor dynamics and fault evolution models include a bearing wear degradation trend model, a misalignment fault vibration characteristic evolution model, a cavitation development stage model, and a spindle fatigue life prediction model. The method for deep integration of mechanism, knowledge and data in step S1 is as follows: the hydraulic mechanism model and rotor dynamics and fault evolution model are embedded into the knowledge graph of the vertical shaft axial flow pump; the named entity recognition and relation extraction method is used to automatically extract operation and maintenance entities and semantic relations from multi-source heterogeneous data, and the extraction results are deeply integrated with the mechanism model to construct a structured operation and maintenance knowledge graph, forming a three-element closed-loop fusion system of mechanism constraint knowledge, knowledge interpretation data and data optimization knowledge.
5. The method according to claim 1 or 2, characterized in that, In step S2, a small-sample fine-tuning dataset is constructed: Addressing the scarcity of real fault samples for vertical shaft axial flow pumps, simulation samples of the entire fault lifecycle are generated based on knowledge graphs and mechanism models. These are then integrated with real historical fault samples to construct a multi-turn dialogue-based small-sample fine-tuning dataset with core fields including operating condition description, feature data, fault evolution path, thought chain reasoning, handling solutions, and mechanism verification. This dataset replaces the traditional single question-and-answer pair dataset. The core types of this dataset include: Status assessment Q&A pairs: Integrate real-time time series data with operating condition boundaries to achieve multi-dimensional assessment of equipment health status; Fault Diagnosis Thinking Chain Data Pair: Decomposes fault diagnosis into a complete thinking chain of feature recognition, operating condition matching, mechanism verification, fault location, root cause analysis, and handling suggestions to improve the interpretability of model reasoning; Fault evolution prediction data pairs: Based on the fault degradation model, construct the correspondence between fault characteristics and evolution trends at different stages, and train the model's long-term prediction capability; Operation and maintenance handling and simulation verification data pair: Match the hierarchical handling plan, combine the digital twin simulation results, and construct the plan effectiveness verification data.
6. The method according to claim 1 or 2, characterized in that, In step S2, the parameter-efficient fine-tuning framework combined with contrastive learning is used to train the model as follows: the parameter-efficient fine-tuning framework with LoRA low-rank adaptation is adopted, the backbone parameters of the base model are frozen, and only the low-rank matrix is trained; at the same time, contrastive learning loss is added to the fine-tuning loss function to bring the feature distance of similar fault samples closer and push the feature distance of dissimilar samples further apart in the feature space, so as to solve the problem of poor generalization ability of the model under small sample size and improve the model's discrimination and recognition ability for early weak fault features. In step S2, the optimization method of the time-aware attention mechanism is as follows: a time-aware attention branch is added to the Transformer architecture to focus on the trend changes, mutation features, periodic features and degradation slope of long time-series data, and adaptively improve the attention weight of weak features of early faults.
7. The method according to claim 1 or 2, characterized in that, In step S2, the illusion suppression module with dual constraints of mechanism and knowledge graph embeds a dual verification process based on knowledge graph and mechanism model in the model inference output layer: every inference result, diagnosis conclusion and treatment suggestion generated by the model must be matched with entities and relationships in the knowledge graph, and the engineering rationality is verified by the mechanism model; content that does not conform to the mechanism law or exceeds the boundary of the knowledge graph is automatically filtered, corrected and labeled with confidence, suppressing professional knowledge illusion from the root and ensuring that the model output fully conforms to the actual water conservancy project.
8. The method according to claim 1 or 2, characterized in that, In step S3, the fully closed-loop intelligent operation and maintenance system includes: early fault intelligent early warning and evolution prediction, specifically: real-time access to spatiotemporally aligned operation data, capturing weak degradation characteristics of equipment and triggering graded early warnings, predicting the fault degradation trend of 72 hours and above based on the fault evolution model, and outputting fault probability, time window and risk level. And / or, in step S3, the fully closed-loop intelligent operation and maintenance system includes multi-dimensional coupled fault root cause diagnosis and inversion, specifically: for early warning faults, multi-parameter coupled correlation analysis is carried out by combining knowledge graphs, mechanism models and historical data to locate the root cause of the fault and distinguish between direct and indirect causes; the entire evolution path of the fault is inverted to clarify the key nodes of operation and maintenance management; And / or, the fully closed-loop intelligent operation and maintenance system in step S3 includes the generation of operation and maintenance handling plans and digital twin simulation verification. Specifically, it involves: generating graded and scenario-based handling plans by combining fault diagnosis results, equipment operating conditions, and pump station operation plans; and verifying the plans through simulation verification of the vertical shaft axial flow pump digital twin model to predict the operating status and handling effects. And / or, in step S3, the fully closed-loop intelligent operation and maintenance system includes feedback on the handling effect and system self-iterative optimization, specifically: automatically collecting operational data after handling and evaluating the handling effect; automatically updating new fault cases and handling data to the knowledge base and knowledge graph; periodically performing incremental fine-tuning of the model to achieve system self-learning and self-optimization. And / or, in step S3, the natural language interactive full-scenario operation and maintenance supports operation and maintenance personnel to initiate inspection queries, fault diagnosis, working condition simulation, ledger management and skills training instructions in natural language through the terminal. The system automatically generates natural language responses and visualization reports that conform to water conservancy engineering specifications.
9. A fully closed-loop intelligent operation and maintenance system for large vertical shaft axial flow pumps based on a large language model, characterized in that, For performing the method according to any one of claims 1-8, comprising: The data preprocessing module is used for spatiotemporal synchronization, noise reduction, and feature extraction of multi-source heterogeneous data; The knowledge base construction module is used to build a dedicated knowledge graph that integrates mechanisms, knowledge, and data. The model training module is used for fine-tuning large models with small samples, attention optimization, and hallucination suppression. The closed-loop operation and maintenance module is used for fault early warning, root cause diagnosis, solution generation, simulation verification and self-iteration. The human-computer interaction module is used for operation and maintenance interaction in all scenarios using natural language.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for multi-dimensional intelligent operation and maintenance of large vertical shaft axial flow pumps based on a large language model as described in any one of claims 1-8.
