A method and system for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models
By splitting the fluid dynamic system into multiple independent twin models according to functional units and coordinating them with a large language model, the problems of large scale, difficult calibration, high computational load and low diagnostic efficiency in the existing fluid dynamic system fault diagnosis are solved, and efficient and safe fault diagnosis is achieved.
Patent Information
- Application Number
- CN202611287677.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-10
AI Technical Summary
Existing fault diagnosis technologies for fluid dynamic systems suffer from problems such as large single-system-level model size, difficulty in parameter calibration, large simulation computation volume, long diagnosis time, lack of intelligent screening mechanism, low diagnosis efficiency, and the inability of large language models to actively command and make dynamic decisions.
The fluid dynamics system is divided into multiple independent twin models according to functional units. Each model is equipped with a standardized fault injection configuration file. Combined with a large language model, initial diagnosis, triage, injection strategy and iterative decision-making are performed. Dynamic adjustment and closed-loop iterative search are performed through similarity threshold to achieve safe and controllable fault injection parameters and improve diagnostic efficiency.
It significantly reduces the amount of simulation computation, improves diagnostic efficiency and intelligence, ensures the safety and reproducibility of injected parameters, supports rapid deployment across multiple models, and has good scalability and industrial practical value.
Smart Images

Figure CN123069892A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fluid dynamics system fault diagnosis technology, and in particular to a method and system for fluid dynamics system fault diagnosis based on the collaboration of multi-unit twin models and large language models. Background Technology
[0002] Fluid power systems, as the core power transmission and control units of modern industrial equipment, are widely used in key fields such as engineering machinery, aerospace, ships, and metallurgical equipment. With the continuous development of fluid power systems towards higher pressure, larger flow rates, higher integration, and greater intelligence, the system structure is becoming increasingly complex, and the coupling relationships between components are becoming tighter, significantly increasing the difficulty of fault diagnosis. Fluid power system faults typically manifest as various phenomena such as abnormal pressure, insufficient flow, and elevated temperature. Fault sources may involve many factors, including hydraulic pump wear, overflow valve set pressure deviation, valve core jamming, pipeline leakage, and seal aging. Furthermore, multiple fault modes often coexist, posing a severe challenge to rapid and accurate fault location.
[0003] Digital twin technology, by establishing a virtual simulation model of physical equipment, injecting hypothetical faults into the model, and comparing the characteristic differences between the simulation output and measured data, can pinpoint the source of the fault. In recent years, it has become an important technical means in the field of fault diagnosis for fluid dynamic systems. For example, Siemens' US-11567489-B2 patent discloses a digital twin-based equipment diagnosis method. This method hypothesizes faults for each element in the equipment twin model, runs the simulation, compares the simulation output with actual sensor data, and determines the fault location based on the similarity. However, this method still has the following shortcomings: First, a complete digital twin model of a fluid dynamics system involves the physical equations and coupling relationships of multiple components such as hydraulic pumps, motors, relief valves, proportional valves, hydraulic cylinders, and sensors. A single system-level model is large in scale, difficult to calibrate parameters, and has a large amount of simulation computation, which is not conducive to deployment on embedded terminals in the field. Second, the system-level model has a large number of fault injection points. This method adopts a mechanical traversal approach of injecting and comparing one by one, resulting in many simulation rounds, long diagnosis time, and a lack of intelligent screening mechanism. Third, the iterative decision-making adopts the method of taking the maximum similarity after matching a fixed threshold, without dynamic adjustment and closed-loop iteration mechanism, resulting in low diagnostic efficiency. Fourth, this method does not have natural language reasoning ability and cannot integrate multi-source heterogeneous information such as alarm semantics, trend descriptions, and historical cases for comprehensive judgment.
[0004] Meanwhile, with the rapid development of general artificial intelligence technology, large language models have begun to be applied in the field of industrial fault diagnosis. For example, the arXiv 2604.22777 paper proposes combining digital twins with FMEA knowledge, using a one-dimensional convolutional neural network to classify simulation data, and then generating interpretable diagnostic reports by a large language model; the CN120122611A patent discloses a fault diagnosis method for CNC systems based on RAG+LLM, which enhances the question-answering ability of the large language model through knowledge graph retrieval. However, in these methods, the large language model only intervenes after the simulation is completed or the classification is finished, responsible for generating reports or answering questions, playing a "post-interpretation" role. It cannot guide the diagnostic direction, select the model to be simulated and fault injection parameters before the simulation runs, nor can it dynamically adjust subsequent diagnostic strategies based on intermediate simulation results. In addition, the large language model itself has reliability issues in its deterministic numerical judgments such as deviation magnitude classification and trend significance determination. If fault injection parameter values are directly generated from it, there is also a risk of generating unreasonable or "illusionary" parameters, affecting the safety and credibility of the diagnosis.
[0005] Furthermore, patent CN117189713A discloses a digital twin and fault injection method for hydraulic systems, using a PCA-BiLSTM neural network to classify simulation outputs; patent CN114183437A discloses a fault injection method for electro-hydraulic actuators. However, these methods still employ a single system-level model without breaking it down into functional unit-level sub-models. Fault injection parameters are manually configured or driven by preset lists without intelligent filtering, and the neural network classifiers rely on a large amount of labeled data for training and have limited generalization ability.
[0006] In summary, existing fault diagnosis technologies for fluid dynamic systems are limited in their digital twin modeling capabilities due to the large scale, difficult calibration, and high computational cost resulting from single-system-level models. Regarding fault injection strategies, they suffer from limitations such as exhaustive enumeration, lack of priority ranking, and insufficient intelligent filtering. Furthermore, in the integration of large language models and digital twins, the large language models passively interpret information without playing an active command role, and suffer from prominent issues such as a lack of closed-loop iterative mechanisms, unreliable deterministic judgments, and uncontrollable parameter generation. Therefore, there is an urgent need in this field for a fault diagnosis method and system for fluid dynamic systems that can deploy digital twin models by functional units, fully leverage the active command and dynamic decision-making capabilities of large language models throughout the diagnostic process, and simultaneously ensure the safety and controllability of injected parameters. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for fault diagnosis of fluid dynamic systems based on the collaboration of multi-unit twin models and large language models, so as to solve the problems existing in the prior art.
[0008] To achieve the above objectives, the present invention provides the following solution: This invention provides a method for fault diagnosis of fluid dynamic systems based on the collaboration of multi-unit twin models and large language models, comprising the following steps: S1. The fluid dynamics system is broken down into multiple independently deployed twin models according to functional or structural units, and each twin model is equipped with a standardized fault injection configuration file; S2. When a fault event is detected, collect and save the fault data, organize the alarm information, trend analysis results, twin model catalog and rule engine pre-annotation results into prompt words, submit them to the large language model for initial diagnosis and triage, output the suspicion ranking of each functional unit, and only activate twin models with suspicion exceeding the preset threshold. S3. For the selected twin model, the large language model prioritizes each fault injection point based on alarm features and the expected symptom fields in the fault injection configuration file, and generates an ordered injection plan. S4. Drive the selected twin models to perform simulations sequentially according to the injection plan, extract the feature parameters of the simulation output, and calculate the similarity with the feature parameters of the actual data; S5. After completing the feature comparison for each injection scenario, the similarity result is fed back to the large language model. The large language model evaluates the current result according to the preset similarity threshold and decides the next action: when the similarity exceeds the first threshold, the match is confirmed and the injection stops; when the similarity is between the first and second thresholds, the result is recorded and the continuity parameter is fine-tuned to continue to the next injection scenario; when the similarity is below the second threshold, the remaining injection points of the current model are skipped and the next Siamese model is switched; when the injection scenarios of all selected models have not reached the matching threshold, the large language model re-evaluates and starts the models not selected in the initial diagnosis stage. S6. After iterative convergence, the large language model integrates alarm information, trend analysis, similarity ranking of each injection scenario, fault type corresponding to the injection scenario with the highest similarity, and retrieved historical cases to generate a diagnostic report.
[0009] Preferably, in step S1, the splitting by functional unit follows the principles of functional independence, weak coupling boundaries, and independent deployment. Each twin model is encapsulated in the form of a shared library and called through a unified interface specification. The unified interface specification includes interfaces for model initialization, warm start, single-step advancement, obtaining state vectors, and cleaning up resources. Each twin model is organized according to a unified directory structure. When the system starts, it locates the model package directory according to the model identifier and automatically loads the shared library and fault injection configuration file of each unit.
[0010] Preferably, the fault injection configuration file is in JSON format and includes a unique model identifier, model name, and fault injection array. Each injection point in the fault injection array includes a unique injection point identifier, fault type name, discrete parameter sequence, continuous parameter interval and step size, and expected symptom list. The large language model selects parameter values only from the discrete parameter sequence and the step size within the given interval.
[0011] Preferably, in step S2, the rule engine pre-annotation includes deviation magnitude classification, trend direction determination, and physical consistency check. The pre-annotation results are injected into prompt words in the form of data rather than input in the form of instructions. The prompt words are dynamically assembled by the prompt word assembler according to the current diagnosis stage and data characteristics. The diagnosis stage includes four stages: initial diagnosis triage, injection strategy, iterative decision-making, and comprehensive diagnosis. Each of the four stages constructs independent prompt words.
[0012] Preferably, the feature parameters include mean, variance, rate of change, and spectral features; the similarity value ranges from 0 to 1.
[0013] Preferably, when the diagnostic results indicate that the fault spans multiple units, multiple twin models are cascaded, and the simulation output of the previous twin model is used as the dynamic boundary input of the next twin model.
[0014] Preferably, the fault event is detected by the lower-level controller, which sends an alarm flag frame through the bus. After receiving the alarm, the data terminal automatically saves the operating data before and after the fault for a preset period of time. Alternatively, the fault event can be diagnosed by the host computer based on long-term time series data of the parameters to determine the fault status.
[0015] Preferably, in step S6, the diagnostic report includes the location of the faulty component and its confidence ranking, fault mode determination, handling suggestions, and details of the reasoning process.
[0016] This invention also provides a fault diagnosis system for fluid dynamic systems based on the collaboration of multi-unit twin models and large language models, including a data acquisition module, a twin model library, a rule engine module, a prompt word assembly module, a large language model reasoning module, a twin simulation engine, and a diagnostic process control module, wherein: The data acquisition module is used to collect real-time operating data of the fluid power system, detect fault events, and automatically save fault data. The twin model library is used to store multiple independent twin models and their fault injection configuration files, which are split according to functional or structural units. The rule engine module is used to make deterministic judgments on the collected data and output deviation classification, trend judgment and physical consistency check results; The prompt word assembly module is used to dynamically assemble prompt words for the large language model based on the current diagnostic stage and data characteristics; The large language model reasoning module is used to perform reasoning in four stages: initial diagnosis and triage, injection strategy, iterative decision-making, and comprehensive diagnosis. The twin simulation engine is used to drive twin model simulation according to the injection plan, extract feature parameters and calculate similarity; The diagnostic process control module is used to coordinate the execution of each module, manage iterations, and switch models.
[0017] Preferably, the large language model reasoning module outputs a ranking of the suspicion levels of each functional unit based on alarm information and trend analysis during the initial diagnosis and triage stage; outputs a priority ranking of fault injection points based on alarm characteristics and expected symptom fields during the injection strategy stage; outputs a decision to continue, stop, or switch models based on similarity feedback during the iterative decision stage; and integrates all information to output a diagnostic report during the comprehensive diagnosis stage.
[0018] The present invention achieves the following beneficial technical effects compared to the prior art: This invention provides a fault diagnosis method and system for fluid dynamic systems based on the collaboration of multi-unit twin models and a large language model. By splitting the complete fluid dynamic system into multiple independently deployed twin models according to functional units and equipping them with standardized fault injection configuration files, the scale of a single model and the difficulty of parameter calibration are significantly reduced, and the amount of simulation computation is greatly reduced, making it suitable for embedded terminal deployment. By upgrading the large language model from a traditional post-event interpretation role to an active commander throughout the entire process of initial diagnosis triage, injection strategy, iterative decision-making, and comprehensive diagnosis, the large language model can prioritize injection points based on alarm features and expected symptom fields before simulation to select the most valuable injection scenarios. After simulation, it can make dynamic decisions based on a three-level similarity threshold and achieve closed-loop adaptive iterative search, greatly reducing invalid simulation rounds and significantly improving diagnostic efficiency. The system improves both the intelligence level and the overall efficiency. Simultaneously, the rule engine's pre-annotation mechanism decouples deterministic judgments, such as deviation magnitude grading and trend significance determination, which large language models are less adept at, from the system. These judgments are pre-calculated by the rule engine and injected as data-based prompts. The large language model only needs to verify the pre-annotated results instead of reasoning from scratch, effectively improving inference quality. Furthermore, the large language model selects fault injection parameters from predefined discrete parameter sequences and within a given interval by step size, without generating the parameter values themselves. This fundamentally eliminates the risk of "illusion" parameters, ensuring the safety and reproducibility of the injected parameters. The mechanism of continuing to the next injection scenario when there is a partial match enables effective detection of composite faults, while the mechanism of adaptively expanding the investigation scope when there is no match achieves a balance between diagnostic coverage and efficiency. The standardized model package and automatic discovery mechanism support rapid deployment across multiple machine types, demonstrating good scalability and industrial practical value. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 A flowchart of the fluid dynamics system fault diagnosis method based on the collaboration of multi-unit twin model and large language model provided by the present invention; Figure 2 This is a schematic diagram of the functional unit division in this invention, taking an electro-hydraulic actuator as an example; Figure 3 This is a structural diagram of the fault injection configuration file in this invention; Figure 4 This is a flowchart of the iterative decision-making process in this invention; Figure 5 The diagram shows the system architecture for fault diagnosis of fluid dynamics systems based on the collaboration of multi-unit twin models and large language models provided by this invention. Detailed Implementation
[0021] 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 scope of protection of the present invention.
[0022] The purpose of this invention is to provide a method and system for fault diagnosis of fluid dynamic systems based on the collaboration of multi-unit twin models and large language models, so as to solve the problems existing in the prior art.
[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0024] Example 1: This invention provides a fault diagnosis method for fluid dynamic systems based on the collaboration of multi-unit twin models and large language models. The overall process is as follows: Figure 1 As shown, the process includes steps S1 (multi-unit twin model splitting and configuration), S2 (large language model initial diagnosis and triage), S3 (large language model injection strategy), S4 (twin simulation and feature comparison), S5 (large language model iterative decision-making), and S6 (large language model comprehensive diagnosis). The specific implementation methods for each step are described in detail below with reference to the accompanying drawings.
[0025] In step S1, the fluid dynamics system is decomposed into multiple independently deployed twin models according to functional units. The decomposition follows the principles of functional independence, weak coupling boundaries, and independent deployment. That is, each unit is physically relatively independent and corresponds to a clearly defined set of components. Unit boundaries are chosen at interfaces with weak parameter coupling, such as the pipeline from the pump outlet to the valve inlet, to allow each model to operate independently. Each unit is independently modeled, its functional model unit file is independently exported, and it is independently cross-compiled into a shared library. Taking an electro-hydraulic actuator as an example, such as... Figure 2 As shown, the complete system is divided into four functional units: a hydraulic power unit, a proportional valve drive unit, a hydraulic cylinder actuation unit, and a sensor unit. It should be noted that the above unit division is only an example; the actual division should be determined based on the equipment's hydraulic schematic diagram and fault mode analysis results. Each unit's twin model is encapsulated as a shared library (.so file) and is called by the upper-level main program through a unified C language interface specification. This interface specification includes the `init` interface for model initialization and loading parameter configuration, the `warm_start` interface for injecting historical state vectors to avoid cold start errors, the `step` interface for single-step progression with historical measured data as input and returning predicted output, the `get_state` interface for obtaining the current state vector for saving and restoring, and the `release` interface for cleaning up resources. Changing the model version only requires replacing the .so file and configuration file; it does not affect the upper-level application. Each twin model is organized into an independently loadable model package according to a unified directory structure. Specifically, a subdirectory named after the machine model is located in the specified root directory. This subdirectory contains the model manifest.json file, the shared library files corresponding to each functional unit, and the fault injection configuration file. The model manifest file contains the device name, model version number, and a list of units. Each unit in the unit list includes a unique unit identifier, unit name, shared library file name, fault injection configuration file name, input parameter list, output parameter list, and state vector dimension. Upon system startup, the system locates the corresponding model package directory based on the current device identifier, reads the model manifest file, and loads the shared libraries and fault injection configuration files for each unit sequentially according to the manifest declarations, thereby achieving rapid adaptation across multiple device models.
[0026] Each twin model is equipped with a standardized fault injection configuration file, which is in JSON format and has the following structure: Figure 3As shown, the system includes a unique model identifier, model name, shared library path, and a fault injection array. Each injection point in the fault injection array contains a unique injection point identifier, fault type name, parameter definition, and expected symptom list. Each parameter in the parameter definition corresponds to a predefined discrete value sequence. These discrete value sequences limit the range of parameter values that can be injected under this fault type. For example, the pump internal leakage coefficient can be selected as discrete values such as 0.05, 0.10, 0.20, and 0.30, and the relief valve set pressure offset can be selected as discrete values such as -5%, -10%, and -15%. The expected symptom list describes the expected symptom characteristics of the system after the fault injection, such as decreased oil supply pressure, decreased flow rate, decreased volumetric efficiency, or decreased system pressure and increased heat generation. The core design of this configuration file is that the large language model is only responsible for selecting appropriate parameter values from the discrete parameter value sequence and continuous parameters within the interval by step size in the subsequent injection strategy stage, without generating the parameter values themselves. This mechanism fundamentally eliminates the risk of the large language model generating unreasonable or phantom parameters, ensuring the safety and reproducibility of fault-injected parameters.
[0027] In step S2, when a fault occurs during the operation of the fluid power system, the lower-level controller detects the abnormal state and sends an alarm flag frame through the controller local area network bus or CANFD bus. After receiving the alarm, the data terminal automatically saves the operating data for a preset duration before and after the fault, such as 60 seconds before and after, including the sampled values of each sensor and their timestamps. After the operator initiates the fault diagnosis process through the human-machine interface, the rule engine first performs deterministic calculations on the collected real-time data to generate pre-labeled results. This pre-labeling includes three parts: deviation amplitude classification, trend direction determination, and physical consistency check. Deviation amplitude classification calculates the percentage deviation of the current value of each key parameter from the normal operating baseline and labels the specific value. Trend direction determination uses the Mann-Kendall test or the cumulative sum CUSUM method to determine the significance of the parameter change trend. Physical consistency check checks whether the physical change directions between related parameters are consistent, such as whether the direction of decrease in oil supply pressure is consistent with the direction of decrease in flow rate. These pre-labeled results are injected into the prompt words in the form of data, such as a format that indicates constant bias with a deviation of 52%, no significance in the Mann-Kendall test, and a cumulative sum of 0, rather than in the form of instructions such as "please determine whether it is a persistent bias". This reduces the understanding burden of large language models, allowing them to verify the pre-labeled results instead of reasoning from scratch.
[0028] The prompt word assembler dynamically assembles prompt words based on the current diagnostic stage and data characteristics. This includes selecting the corresponding rule set based on the stage and the type of anomaly hit, retrieving at most two few-sample cases by anomaly pattern to control the number of tokens and avoid information overload, and allocating the token budget according to the stage complexity. Independent prompt words are constructed for each of the four stages: initial diagnosis triage, injection strategy, iterative decision-making, and comprehensive diagnosis, without mixing them. The prompt words in the initial diagnosis triage stage include system role definitions, alarm information, trend analysis results, a twin model catalog, and pre-annotated results from the rule engine. The twin model catalog lists the name, covered components, fault injection point list, and expected symptoms of each functional unit model. The large language model assigns a suspicion score of zero to one and a judgment criterion to each functional unit based on the above information, and outputs a list of twin models to be activated. Only models with a suspicion score exceeding a preset threshold, such as 0.5, are activated, thereby filtering out low-probability functional units in the early stages of diagnosis and avoiding unnecessary subsequent simulation calculations.
[0029] In step S3, for the one or more twin models selected in step S2, the large language model prioritizes each fault injection point based on the degree of match between the actual alarm symptoms and the expected symptoms of each fault injection, the correspondence between trend analysis results and fault development patterns, and the actual distribution of each fault type in similar historical alarm cases retrieved by the retrieval enhancement generation (RAG), generating an ordered injection plan. Specifically, a gradual trend points to degenerative faults such as wear-related faults, while abrupt trends point to sudden faults such as jamming or breakage faults. When outputting the injection plan, the large language model explicitly specifies the identifiers of each injection point to be executed sequentially under each selected model and their order. In this step, the large language model selects parameter values only from the discrete parameter value sequence defined in the standardized fault injection configuration file, without automatically generating parameter values, ensuring that the parameters injected each time are within the predefined safety range and are reproducible.
[0030] In step S4, the diagnostic process control module drives the selected twin models to perform simulations sequentially according to the injection plan generated in step S3. Each simulation uses actual operating data before and after the fault as input to drive the selected twin model. Fault parameters corresponding to selected discrete values are applied to the model at specified injection points. After running the simulation, feature parameters of the simulation output are extracted. These feature parameters include the signal's mean, variance, rate of change, and spectral characteristics. Simultaneously, feature parameters of the same type are extracted from the actual collected fault data. The similarity between the simulation output feature parameters and the actual data feature parameters is calculated to obtain a similarity score for the injection scenario. This score ranges from zero to one.
[0031] In step S5, after completing the feature comparison for each injection scenario, the similarity result is fed back to the large language model. The large language model evaluates the current result according to the preset three-level similarity threshold and decides on the next action. Its decision-making process is as follows: Figure 4 As shown. When the similarity exceeds the first threshold, for example, 0.8, it is considered a confirmed match, all subsequent injections are stopped, and the system enters the comprehensive diagnosis stage. When the similarity is between the first and second thresholds, for example, 0.5 to 0.8, it is considered a partial match, the injection scenario and its similarity score are recorded, and then the next injection scenario in the injection plan is executed. This mechanism allows the system to approximate the real fault mode as closely as possible and can also detect complex fault scenarios where multiple faults coexist. When the similarity is below the second threshold, for example, 0.5, it is considered a mismatch, the remaining injection points of the current model are skipped, and the system switches to the next twin model in the suspicion ranking for diagnosis. When all injection scenarios of all selected models have not reached the matching threshold, that is, when the similarity of any injection scenario exceeds the first threshold, the large language model re-evaluates the current diagnostic results, determines whether the scope of investigation needs to be expanded, and starts the models that were not selected in the initial diagnosis stage due to their suspicion being below the threshold. This achieves adaptive expansion and contraction of the scope of investigation, balancing diagnostic efficiency and coverage. The above iterative decision-making process is achieved through multiple interactions between the large language model and the twin simulation engine. After each round of simulation, the large language model obtains feedback and makes the next decision until the stopping condition is met or all candidate models have been traversed.
[0032] In step S6, once the iteration converges (i.e., a confirmed matching injection scenario is found or all candidate scenarios have been traversed), the large language model enters the comprehensive diagnostic phase. The large language model integrates alarm information and trend analysis describing the complete fault phenomenon, the similarity ranking of each injection scenario and its corresponding fault type, the specific fault type corresponding to the injection scenario with the highest similarity, historical cases retrieved by RAG (Retrieval Enhancement), and a list of device functional units and typical fault modes. This generates a diagnostic report containing fault component location and its confidence ranking, fault mode determination, handling suggestions, and details of the reasoning process. The diagnostic report is simultaneously displayed on the operator's human-machine interface in both natural language and structured report formats. Operators can view the diagnostic summary and handling suggestions, while engineers can expand to view detailed records of each stage of the reasoning process.
[0033] Based on the above basic process, when diagnostic results indicate that the fault may span multiple functional units—for example, fluctuations in the hydraulic pump outlet pressure simultaneously affect the normal operation of the proportional valve—multiple twin models can be cascaded. Specifically, the simulation output of the previous twin model is used as the dynamic boundary input for the next twin model. For instance, in cross-unit fault diagnosis, the hydraulic power unit twin model is run first, and its simulation output includes outlet pressure fluctuation characteristics. This outlet pressure is then used as the dynamic boundary input for the proportional valve drive unit twin model. The proportional valve drive unit twin model is run to observe the response characteristics of the proportional valve under fluctuating supply pressure. Finally, the simulation results of the two models are combined for diagnosis. In the initial independent operation scheme, each model has its own rated boundary conditions. The cascaded expansion scheme is implemented as needed after the independent operation scheme has been verified.
[0034] The following is a specific example of a fault in the set pressure deviation of the relief valve in the hydraulic power unit of a certain type of electro-hydraulic actuator, to fully illustrate the method of the present invention. In the scenario setting of the embodiment, during equipment operation, the lower-level controller detects that the oil supply pressure is lower than the lower limit threshold and sends an alarm flag frame through the CANFD bus. After receiving the alarm, the data terminal automatically saves the operating data for 60 seconds before and after the fault. The operator initiates fault diagnosis through the human-machine interface touch screen.
[0035] In step S1, the model is pre-installed with four functional unit twin models, including the hydraulic power unit model identified as DT-HPU-01 and corresponding to the shared library hpu_twin.so and the fault injection configuration file hpu_inject.json; the proportional valve drive unit model identified as DT-VALVE-01 and corresponding to the shared library valve_twin.so and the fault injection configuration file valve_inject.json; the hydraulic cylinder actuation unit model identified as DT-CYL-01 and corresponding to the shared library cyl_twin.so and the fault injection configuration file cyl_inject.json; and the sensor unit model identified as DT-SENSOR-01 and corresponding to the shared library sensor_twin.so and the fault injection configuration file sensor_inject.json. The fault injection profile for the hydraulic power unit includes three injection points: the pump internal leakage wear injection point is identified as FI-PUMP-001, with a discrete value sequence of internal leakage coefficient parameters of 0.05, 0.10, 0.20, and 0.30, and the expected symptoms are decreased oil supply pressure, decreased flow rate, and reduced volumetric efficiency; the relief valve set pressure offset injection point is identified as FI-RV-001, with a discrete value sequence of pressure offset parameters of -5%, -10%, and -15%, and the expected symptoms are decreased system pressure and increased heat generation; the relief valve sticking injection point is identified as FI-RV-002, with a discrete value sequence of valve damping coefficient parameters of 0.3, 0.5, and 0.8, and the expected symptoms are pressure fluctuations and sluggish response.
[0036] In step S2, the rule engine pre-labels the collected data, obtaining results showing a supply pressure deviation of -52%, a Mann-Kendall test result that is not significant, and a cumulative sum of 0 indicating a constant deviation, and that the decrease in supply pressure and flow rate are in the same direction. The prompt word assembler organizes the alarm information (low supply pressure and abnormal flow rate), the trend analysis result (continuous decrease in supply pressure and a cumulative sum indicating a constant deviation), the twin model catalog (model names of the four functional units, a list of component fault injection points and expected symptoms), and the rule engine pre-labeling results into initial diagnosis triage prompt words and submits them to the large language model. After inference, the large language model outputs a suspicion degree of 0.85 for the hydraulic power unit, judging that the continuous pressure decrease trend points to a degradation fault, with pump wear or internal leakage of the relief valve being the most likely culprits; a suspicion degree of 0.25 for the proportional valve drive unit, judging that the abnormal flow rate may be a cascading effect of the pressure decrease, and that the proportional valve itself is less likely to be faulty; a suspicion degree of 0.10 for the hydraulic cylinder actuator unit; and a suspicion degree of 0.05 for the sensor unit. Because the suspicion level of the hydraulic power unit (0.85) exceeds the preset threshold of 0.5, the system only activates the twin model of the hydraulic power unit.
[0037] In step S3, the large language model, for the selected hydraulic power unit twin model, prioritizes each fault injection point based on the degree of match between the actual alarm symptoms (low oil supply pressure) and the expected symptoms of each fault injection, as well as the trend characteristics (constant deviation pointing to set pressure type offset faults). The injection plan is generated as follows: the overflow valve set pressure offset injection point is identified as FI-RV-001, the pump internal leakage injection point is identified as FI-PUMP-001, and the overflow valve stuck injection point is identified as FI-RV-002.
[0038] In steps S4 and S5, the first round of injection follows the injection plan, performing pressure offset injection from the overflow valve. The large language model selects -10% from the discrete value sequences of -5%, -10%, and -15% as the injection parameter. The simulation of the hydraulic power unit twin model is driven by actual operating data before and after the fault. Simulation output feature parameters, including mean, variance, rate of change, and spectral characteristics, are extracted and compared with the actual data feature parameters, with a similarity of 0.88. The large language model uses a three-level threshold to determine if the similarity of 0.88 is greater than the first threshold of 0.8, confirming a match. Injection is then stopped, and the process enters the comprehensive diagnostic phase.
[0039] In step S6, the large language model integrates all information to generate a diagnostic report. The diagnostic summary states that the overflow valve set pressure deviation causes the system pressure to remain low. Among the faulty components, the confidence level for the overflow valve set pressure deviation is 0.88, supported by evidence of a twin simulation similarity of 0.88, a matching pressure drop trend, and a high degree of agreement between the simulation characteristics and actual measurements after the overflow valve pressure deviation injection. The recommended course of action is to check the overflow valve set pressure and test the opening pressure, and recalibrate if necessary. The confidence level for internal leakage and wear in the hydraulic pump is 0.12, supported by evidence that the pressure drop trend can also be caused by pump wear, but internal leakage and wear in the pump... The leak injection similarity is only 0.35, therefore it is ruled out, and the recommended treatment is no treatment required, but regular monitoring of volumetric efficiency is suggested. The impact assessment is that the current deviation level is at the warning level and does not affect normal operation at present. It is recommended to handle it during the next scheduled inspection. The reasoning details are as follows: In the initial diagnosis and diversion stage, the hydraulic power unit was determined to be the most suspicious with a similarity of 0.85. In the injection strategy stage, pump internal leakage and relief valve pressure deviation were prioritized. The similarity of the first round of pump internal leakage injection was 0.35, which was mismatched, while the similarity of the second round of relief valve pressure deviation injection was 0.88, which was a match. Based on the comprehensive judgment, the relief valve set pressure deviation was determined to be the main fault. The diagnostic conclusion is displayed in both natural language and structured report formats on the operator's human-machine interface. The operator can view the diagnostic summary and treatment recommendations, and the engineer can expand to view the reasoning process of each stage.
[0040] Example 2: This invention also provides a fault diagnosis system for fluid dynamic systems based on the collaboration of multi-unit twin models and large language models, the system architecture of which is as follows: Figure 5As shown, the system includes a data acquisition module, a twin model library, a rule engine module, a prompt word assembly module, a large language model inference module, a twin simulation engine, and a diagnostic process control module. The data acquisition module collects real-time operating data of the fluid dynamics system, detects fault events, and automatically saves fault data. The twin model library stores multiple independent twin models broken down by functional units and their fault injection configuration files. The rule engine module performs deterministic judgments on the collected data, outputting deviation grading, trend determination, and physical consistency check results. The prompt word assembly module dynamically assembles prompt words for the large language model based on the current diagnostic stage and data characteristics. The large language model inference module outputs a ranking of the suspicion levels of each functional unit based on alarm information and trend analysis during the initial diagnosis triage stage; outputs a priority ranking of fault injection points based on alarm characteristics and expected symptom fields during the injection strategy stage; outputs a decision to continue, stop, or switch models based on similarity feedback during the iterative decision stage; and integrates all information to output a diagnostic report during the comprehensive diagnosis stage. The twin simulation engine drives twin model simulation according to the injection plan, extracts feature parameters, and calculates similarity. The diagnostic process control module coordinates the execution of each module according to steps S1 to S6 and manages iterations and model switching. Before simulation, the large language model executes initial diagnostic triage and injection strategies to actively guide the simulation direction. After each simulation, it performs iterative decision-making to dynamically adjust subsequent strategies. After iteration convergence, it performs comprehensive diagnosis to output the final conclusion, thus achieving active command and closed-loop adaptive search of the large language model throughout the entire diagnostic process. This system can be deployed on data terminals or edge computing devices in industrial settings, connecting to lower-level controllers and human-machine interface devices through standardized interfaces to achieve real-time fault diagnosis.
[0041] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0042] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.
[0043] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.
Claims
1. A method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models, characterized in that, Includes the following steps: S1. The fluid dynamics system is broken down into multiple independently deployed twin models according to functional or structural units, and each twin model is equipped with a standardized fault injection configuration file; S2. When a fault event is detected, collect and save the fault data, organize the alarm information, trend analysis results, twin model catalog and rule engine pre-annotation results into prompt words, submit them to the large language model for initial diagnosis and triage, output the suspicion ranking of each functional unit, and only activate twin models with suspicion exceeding the preset threshold. S3. For the selected twin model, the large language model prioritizes each fault injection point based on alarm features and the expected symptom fields in the fault injection configuration file, and generates an ordered injection plan. S4. Drive the selected twin models to perform simulations in sequence according to the injection plan, extract the feature parameters of the simulation output, and calculate the similarity with the feature parameters of the actual data; S5. After completing the feature comparison for each injection scenario, the similarity result is fed back to the large language model. The large language model evaluates the current result according to the preset similarity threshold and decides the next action: when the similarity exceeds the first threshold, the match is confirmed and the injection stops; when the similarity is between the first and second thresholds, the result is recorded and the continuity parameter is fine-tuned to continue to the next injection scenario; when the similarity is below the second threshold, the remaining injection points of the current model are skipped and the next Siamese model is switched; when the injection scenarios of all selected models have not reached the matching threshold, the large language model re-evaluates and starts the models not selected in the initial diagnosis stage. S6. After iterative convergence, the large language model integrates alarm information, trend analysis, similarity ranking of each injection scenario, fault type corresponding to the injection scenario with the highest similarity, and retrieved historical cases to generate a diagnostic report.
2. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, In step S1, the splitting by functional unit follows the principles of functional independence, weak coupling boundaries, and independent deployment. Each twin model is encapsulated in the form of a shared library and called through a unified interface specification. The unified interface specification includes interfaces for model initialization, warm start, single-step advancement, obtaining state vectors, and cleaning up resources. Each twin model is organized according to a unified directory structure. When the system starts, it locates the model package directory according to the model identifier and automatically loads the shared library and fault injection configuration file of each unit.
3. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, The fault injection configuration file is in JSON format and contains a unique model identifier, model name, and fault injection array. Each injection point in the fault injection array contains a unique injection point identifier, fault type name, discrete parameter sequence, continuous parameter interval and step size, and expected symptom list. The large language model selects parameter values only from the discrete parameter sequence and within a given interval by step size.
4. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, In step S2, the rule engine pre-annotation includes deviation magnitude classification, trend direction determination, and physical consistency check. The pre-annotation results are injected into prompt words in the form of data rather than input in the form of instructions. The prompt words are dynamically assembled by the prompt word assembler according to the current diagnosis stage and data characteristics. The diagnosis stage includes four stages: initial diagnosis triage, injection strategy, iterative decision-making, and comprehensive diagnosis. Each of the four stages constructs independent prompt words.
5. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, The characteristic parameters include mean, variance, rate of change, and spectral characteristics; The similarity value ranges from 0 to 1.
6. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, When the diagnostic results indicate that the fault spans multiple units, multiple twin models are cascaded, with the simulation output of the previous twin model serving as the dynamic boundary input for the next twin model.
7. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, The fault event is detected by the lower-level controller, which sends an alarm flag frame through the bus. After receiving the alarm, the data terminal automatically saves the operating data before and after the fault for a preset duration. Alternatively, the fault event can be diagnosed by the host computer based on long-term time series data of the parameters to determine the fault status.
8. The method for fault diagnosis of fluid dynamic systems based on the synergy of multi-unit twin models and large language models as described in claim 1, characterized in that, In step S6, the diagnostic report includes the location of the faulty component and its confidence ranking, fault mode determination, handling suggestions, and details of the reasoning process.
9. A fault diagnosis system for fluid dynamic systems based on the collaboration of a multi-unit twin model and a large language model, characterized in that, It includes a data acquisition module, a twin model library, a rule engine module, a prompt word assembly module, a large language model inference module, a twin simulation engine, and a diagnostic process control module, among which: The data acquisition module is used to collect real-time operating data of the fluid power system, detect fault events, and automatically save fault data. The twin model library is used to store multiple independent twin models and their fault injection configuration files, which are split according to functional or structural units. The rule engine module is used to make deterministic judgments on the collected data and output deviation classification, trend judgment and physical consistency check results; The prompt word assembly module is used to dynamically assemble prompt words for the large language model based on the current diagnostic stage and data characteristics; The large language model reasoning module is used to perform reasoning in four stages: initial diagnosis and triage, injection strategy, iterative decision-making, and comprehensive diagnosis. The twin simulation engine is used to drive twin model simulation according to the injection plan, extract feature parameters and calculate similarity; The diagnostic process control module is used to coordinate the execution of each module, manage iterations, and switch models.
10. The fluid dynamics system fault diagnosis system based on the synergy of multi-unit twin models and large language models as described in claim 9, characterized in that, The large language model reasoning module outputs a ranking of the suspicion levels of each functional unit based on alarm information and trend analysis during the initial diagnosis and triage stage; and outputs a priority ranking of fault injection points based on alarm characteristics and expected symptom fields during the injection strategy stage. During the iterative decision-making phase, the model continues, stops, or switches based on similarity feedback; during the comprehensive diagnosis phase, all information is integrated to output a diagnostic report.
Citation Information
Patent Citations
Fault modeling method for electro-hydraulic actuator of gas turbine
CN114183437A
Hydraulic system fault diagnosis method based on digital twin drive
CN117189713A
Numerical control system fault diagnosis method and system based on knowledge injection
CN120122611A
Method and device for efficient fault analysis through simulated faults in a digital twin
US11567489B2
