Hydropower station oil and gas pressure system fault prediction method fusing perception data and knowledge graph

Abstract

The invention belongs to the field of big data and artificial intelligence, and discloses a hydropower station oil and gas pressure system fault prediction method fusing perception data and a knowledge graph, which is characterized by comprising the following steps: collecting and preprocessing multi-modal data; performing feature extraction and cross-modal correlation analysis; constructing and dynamically updating a knowledge graph; fusing the physical mechanism model and the data driving model; constructing a fault mode library and reasoning knowledge; training and optimizing a multi-modal prediction model; and generating a fault early warning and maintenance strategy. The method has the main beneficial technical effects that the accuracy of fault prediction is improved, the fault root cause analysis capability is improved, the comprehensive perception and cognition capability is enhanced, the decision support capability is enhanced, the interpretability and credibility of the model are improved, and a whole-process intelligent operation and maintenance closed loop is realized.

Description

Technical Field

[0001] The present invention belongs to the field of big data and artificial intelligence, and specifically involves the combination of mechanical performance degradation analysis of the oil and gas pressure system of a hydropower station, multi-source heterogeneous data fusion, knowledge graph construction, and multimodal perception models. It aims to achieve early fault warning and predictive maintenance of equipment through multi-sensor data perception and dynamic knowledge reasoning, and discloses a fault prediction method for the oil and gas pressure system of a hydropower station that integrates perception data and knowledge graphs. Background Art

[0002] The pressure oil and gas tanks and pressure air tanks in a hydropower station's oil and gas pressure system utilize a gas-liquid linkage to achieve coordinated regulation of pressure and air, improving system stability and flexibility. As a core auxiliary system for hydropower units, the operating status of a hydropower station's oil and gas pressure system directly impacts power generation efficiency and safety. Traditional operations and maintenance rely on manual inspections and scheduled maintenance. While most hydropower stations have established data collection platforms to collect vibration, oil, and temperature data from their oil and gas equipment, these data are not linked, making unified analysis difficult.

[0003] Most existing work uses LSTM and Transformer models to predict relevant sensor data. These methods detect abnormalities in the data by detecting frequent fault anomalies, but ignore cross-modal information such as oil state and environmental parameters. Furthermore, single time-series prediction algorithms struggle to capture the degradation trends of complex oil and gas pressure systems, including the degradation mechanisms of fatigue cracks and seal failure. Furthermore, during hydropower generation operations, the operating environment of the oil and gas pressure system fluctuates with environmental changes and maintenance. The impact of oil and gas composition on the pressure system frequently changes, necessitating continuous updating of the relevant operational knowledge graph to prevent prediction model failure.

[0004] CN117150032A discloses an intelligent maintenance system and method for hydropower station generator sets, relating to the field of hydropower station equipment management. The system includes: a data interface module for data interoperability with other systems within the hydropower station generator set; a database construction module for collecting data from the generator set and building a database and database model based on this data; a regular maintenance module for automatically conducting equipment status assessments and defect statistical analysis within a preset time period, generating corresponding reports, and automatically pushing maintenance work orders and executing processes based on the equipment's regular work list; and a precision maintenance module for intelligently tracking and comprehensively determining the equipment's operating status, issuing alarms when equipment anomalies occur, and automatically generating component diagnostic sheets. This system enables precise location of abnormal equipment, automatic triggering of abnormality handling processes, and precise equipment maintenance, reducing unnecessary equipment losses and improving the accuracy and efficiency of detecting equipment defects and abnormalities. However, there is still room for improvement in fault prediction.

[0005] In light of this, the present invention integrates heterogeneous data such as vibration, oil, pressure, and temperature to construct a unified feature space. Based on degradation characteristics, the knowledge graph dynamically updates causal relationships and failure modes. Furthermore, the system combines physical mechanism models with deep learning models to eliminate noise interference, improving the accuracy and robustness of oil and gas system fault prediction. Summary of the Invention

[0006] To solve the above problems, the purpose of the present invention is to disclose a method for predicting faults in the oil and gas pressure system of a hydropower station by integrating perception data and knowledge graphs, which is achieved by adopting the following technical solutions.

[0007] A method for predicting faults in an oil and gas pressure system of a hydropower station by integrating sensory data and a knowledge graph, characterized by comprising the following steps:

[0008] Step 1: Multimodal data acquisition and preprocessing;

[0009] Step 2: Feature extraction and cross-modal correlation analysis;

[0010] Step 3: Knowledge graph construction and dynamic update;

[0011] Step 4: Fusion of physical mechanism model and data-driven model;

[0012] Step 5: Fault mode library construction and knowledge reasoning;

[0013] Step 6: Multimodal prediction model training and optimization;

[0014] Step 7: Generate fault warning and maintenance strategy.

[0015] The present invention improves the generalization and dynamics of fault prediction of the oil and gas pressure system of a hydropower station.

[0016] The present invention has the following main beneficial technical effects:

[0017] 1. Improve the accuracy of fault prediction: Through cross-modal correlation analysis, hidden connections between different sensor data and knowledge graph information are revealed to capture early and subtle fault characteristics.

[0018] 2. Improve the ability to analyze the root cause of faults: Through the topological structure and rich associated information of the knowledge graph, when a potential fault is predicted, it can be quickly traced to the possible faulty components, failure modes and their propagation paths, facilitating the precise location of the core of the problem.

[0019] 3. Enhanced comprehensive perception and cognitive capabilities: By integrating multimodal perception data (such as vibration, temperature, pressure and other sensor data) and knowledge graphs (including structured knowledge such as equipment structure, historical failures, maintenance records, expert rules, etc.), a more comprehensive and in-depth understanding of system status is achieved, breaking through the limitations of a single data source or model.

[0020] 4. Enhanced decision-making support capabilities: The fault warning and maintenance strategy generation part not only predicts the probability of failure, but also integrates the relevant information provided by the knowledge graph (such as spare parts status, maintenance costs, scope of impact, and maintenance best practices) to generate smarter and more actionable maintenance suggestions and response strategies, assisting operation and maintenance personnel in making optimal decisions and reducing unplanned downtime.

[0021] 5. Improve the interpretability and credibility of the model: As a background knowledge base, the knowledge graph provides semantic explanation and support for the prediction results of the data-driven model, enhancing the credibility of the model results and user acceptance.

[0022] 6. Realize a closed loop of intelligent operation and maintenance throughout the entire process: This method forms a complete closed loop of intelligent analysis and prediction from data collection, processing, analysis, model building to final early warning and decision support, significantly improving the intelligence level and efficiency of operation and maintenance of key equipment (oil and gas pressure systems) of hydropower stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Schematic diagram of the process of the present application. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand and implement this patent, the implementation methods of this application are now described in detail in conjunction with the drawings in the specification.

[0025] Please see Figure 1 A method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graphs is characterized by comprising the following steps:

[0026] Step 1: Multimodal data acquisition and preprocessing;

[0027] Step 2: Feature extraction and cross-modal correlation analysis;

[0028] Step 3: Knowledge graph construction and dynamic update;

[0029] Step 4: Fusion of physical mechanism model and data-driven model;

[0030] Step 5: Fault mode library construction and knowledge reasoning;

[0031] Step 6: Multimodal prediction model training and optimization;

[0032] Step 7: Generate fault warning and maintenance strategy.

[0033] The present invention improves the generalization and dynamics of fault prediction of the oil and gas pressure system of a hydropower station.

[0034] The embodiments of the present invention are described in detail as follows:

[0035] Step 1: Multimodal data acquisition and preprocessing: This step uses vibration, liquid level, and pressure sensors installed on the oil and gas pressure tank to sense its status data in real time, eliminates fluctuations and noise in the sensed data, and uses a filtering algorithm to smooth the sensed data. The specific sub-steps are as follows:

[0036] Step 1-1: Install vibration sensors on the bearing seat, gearbox housing, and seal connections of the pressure oil tank to monitor the equipment vibration frequency in real time. Use a magnetic oil particle sensor to detect the metal particle concentration, and a spectrum analyzer to detect the oil oxidation product index and liquid level. Pressure sensors are installed at the oil pump outlet, the top of the oil tank, and the seal interface to collect pressure. Use an infrared thermal imager to monitor the surface temperature distribution of the oil tank. Synchronize the data from these multiple sensors to a centralized monitoring center via Industrial Ethernet (such as Profinet) or the OPC UA protocol. Each sensor samples at 1 kHz, and its value is recorded as follows:

[0037] ,

[0038] in It is a time domain sensing signal (which can be a vibration signal, pressure signal, or oil index). is the amplitude of the i-th sensor, is the sampling frequency of the i-th sensor, is the phase value of the i-th sensor used for calibration.

[0039] Step 1-2: Perform N-layer discrete wavelet decomposition on the perception signal to obtain the corresponding detail coefficients and approximation coefficients, as shown in the formula:

[0040] ,

[0041] in is the coefficient of the Nth layer wavelet decomposition, that is, the relevant low-frequency component; is the detail coefficient of the kth layer, corresponding to high-frequency components of different scales.

[0042] Process the decomposition coefficients using soft thresholding:

[0043] ,

[0044] ,

[0045] c is the original coefficient (such as and ), is the threshold, is the noise standard deviation, M is the signal length;

[0046] The signal strength after reconstructing the perception signal and eliminating noise is:

[0047] ,

[0048] Steps 1-3: Perform Z-score normalization (mean = 0, standard deviation = 1) on the oil and gas pressure system sensing data to ensure that the data from different sensors are in the same dimension, as shown in the following formula: And use the 3σ principle to eliminate outliers in the pressure sensor data (such as points that exceed the mean ± 3 times the standard deviation):

[0049]

[0050] in and are the mean and variance respectively.

[0051] Step 2: Feature Extraction and Cross-modal Correlation Analysis: This step extracts the time-domain, frequency-domain, and operational-related features of the oil and gas pressure system's mechanical wear and vibration, as reflected in the sensing signals. This allows for cross-modal alignment of multi-source sensors and better serves early fault detection. This step specifically includes the following sub-steps:

[0052] Step 2-1: Extract and calculate the time domain features of each sensing signal of the oil and gas pressure system, including root mean square (RMS), peak-to-peak value (PPV), and kurtosis features, to detect bearing wear or gear cracks;

[0053] The RMS value is calculated as follows to measure the energy intensity and vibration intensity of the signal:

[0054] ,

[0055] is the i-th sampling point of signal x, and N is the total number of sampling points used to calculate the signal;

[0056] The peak-to-peak value calculation formula is as follows, which is used to measure the dynamic range and impact characteristics of the signal, especially the abnormal characteristics:

[0057] ,

[0058] The kurtosis characteristic calculation formula is as follows to reflect the signal peak characteristics:

[0059] ,

[0060] As shown above, and are the mean and variance of the signal respectively.

[0061] Step 2-2: Use Fast Fourier Transform (FFT) to extract the main frequency components and combine it with envelope analysis to detect early faults such as crack resonance, as shown in the formula:

[0062] ,

[0063] in is the frequency signal, f is the frequency variable, n is the nth frequency cost, and N is the total frequency component.

[0064] Step 2-3: Extract oil-related features, including the temperature difference between the tank surface and the ambient temperature, the acid value and moisture content characteristics of the oil spectral analysis, and the metal particle concentration characteristics.

[0065] Step 2-4: Use the Pearson correlation coefficient to analyze the relationship between the vibration signal and the oil particle concentration. This is used to extract the coupling relationship between the metal particle concentration and the vibration spectrum shift caused by oil pump wear for subsequent knowledge graph construction. The specific relationship is shown in the formula:

[0066] ,

[0067] r is the correlation coefficient, xi and yi are the values ​​of the two sensors or modes respectively, and are the means of n samples respectively.

[0068] Step 3: Knowledge graph construction and dynamic update. This step is to build a fault-related knowledge graph based on the working mechanism of the oil and gas pressure system, provide more clear entity relationships and the correlation mapping relationship between performance and faults, and dynamically update it in a timely manner. It specifically includes the following sub-steps:

[0069] Step 3-1: Use the named entity method to extract equipment entities (such as "oil and gas pressure tank", "oil pump", "seal", and "bearing") and fault types (such as "seal failure", "bearing wear", and "oil pump cavitation"):

[0070] ,

[0071] in is the entity type (e.g., "device"), is a property (such as "pressure"), is the attribute value (such as "1.5 MPa").

[0072] Step 3-2: Extract the monitoring relationship and causal relationship between the equipment entity fault and the perception data features according to the working mechanism specifications, monitoring relationship (such as {"oil sensor", "monitoring", "oil particle concentration"}) and causal relationship (such as {"seal wear", "causal", "oil leakage"} and {"oil leakage", "causal", "pressure fluctuation"}).

[0073] Step 3-3: Use a large language model or natural language processing tool to extract equipment failures, related indicators, and causal relationships from the work mechanism specifications and technical analysis maintenance manuals, build relevant knowledge graphs, and use graph neural networks to represent and model entity relationships:

[0074] ,

[0075] in is the embedding vector of the l-th layer node, is the adjacency matrix, is the learnable weight matrix, is the activation function.

[0076] Step 3-4: Update the knowledge graph based on the monitoring attribute indicators and fault phenomena during real-time operation (e.g., when the oil particle concentration is > 13, the seal is worn). Specifically, it is expressed as follows:

[0077] ,

[0078] in To add new entities or attributes (such as "oil pump", "particle concentration"), To add new relationships (such as connection, cause and effect).

[0079] Step 4: Fusion of the physical model and the data model. This step uses the knowledge graph generated in step 3 to further construct a physical mechanism model, integrates the data-driven prediction model, and simulates the oil flow and pressure model of the oil and gas pressure system to eliminate the impact of noise. At the same time, it better predicts the oil pump outlet pressure fluctuation and accurately predicts oil and gas system failures. The specific sub-steps include the following:

[0080] Step 4-1: Use the fluid dynamics model to establish a correlation model between oil fluid density, velocity, and pressure, and predict the pressure fluctuation at the oil and gas pressure tank port:

[0081] ,

[0082] in is the fluid density, is the fluid velocity, For pressure, is the dynamic viscosity, For external force.

[0083] Step 4-2: Use the Fourier heat conduction equation (∂T / ∂t = α(∂²T / ∂x²)) to calculate the surface temperature distribution of the oil tank. Combined with the thermal boundary conditions (such as the convection heat transfer coefficient h = 10 W / m²·K), the thermal aging rate of the seal is predicted.

[0084] Step 4-3: Divide the sensor signal data into 1024-point blocks and use the Transformer model to extract temporal dependencies for predicting the remaining service life of bearing wear:

[0085] ,

[0086] , are input data, query, key, and value matrices respectively.

[0087] Step 4-4: Perform weighted fusion of the data obtained from the physical model and the predicted value of the data model to reduce the impact of data noise on the prediction results:

[0088] ,

[0089] Such as oil and gas port pressure physical model is the value obtained by kinetic model simulation calculation in step 4-1, is the predicted value obtained in step 4-3; α is the model weight, which can be taken as 0.6.

[0090] Step 5: Fault mode library construction and knowledge reasoning. This step extracts fault modes based on the previous steps and historical data, and builds relevant network models for fault classification and causal reasoning. It specifically includes the following sub-steps:

[0091] Step 5.1: Build a failure mode library based on historical failure cases and classify the failure modes into mechanical failures and oil failures. Mechanical failures include bearing wear, gear breakage, and seal aging, while oil failures include emulsification, oxidation, and excessive moisture. The corresponding parameter features for each failure mode are as follows:

[0092] ,

[0093] in For the kth fault feature (such as "abnormal oil pressure"), based on this pattern library, historical fault data is classified through machine learning methods such as random forest, and new fault patterns (such as "oil pump cavitation") are automatically added.

[0094] Step 5.2: Use the Monte Carlo simulation method to perform probabilistic prediction reasoning on oil and gas faults based on the fault pattern library and attribute characteristics:

[0095] ,

[0096] in It is the posterior probability after the feature appears. For example, multiple Monte Carlo simulations are performed on the temperature field of the oil tank to predict the probability (e.g., 95%) that the thermal stress of the seal exceeds the critical value (>150 MPa), triggering a maintenance warning.

[0097] Step 5.3: Use Bayesian network for causal reasoning and determine the probability of related faults based on attribute indicators, as shown in the formula:

[0098] ,

[0099] If the probability of oil particle concentration > 12 is 0.8 and the probability of vibration spectrum shift > 20% is 0.7, the probability of bearing wear is calculated (P = 0.92).

[0100] Step 6: Multimodal prediction model training and optimization. This step is to train and verify the end-to-end model involved in the above steps. The loss function is designed based on the training results of the above model to balance the minimum mean square error and cross entropy, and the training learning rate is adjusted according to the data features. The results are then verified. It specifically includes the following sub-steps.

[0101] Step 6-1: Use a weighted index as the loss function to balance the minimum mean square error (MSE) and cross entropy (CrossEntropy). MSE is used to optimize the prediction accuracy of the physical model, while CrossEntropy is used to meet the probability prediction accuracy of fault type identification:

[0102] ,

[0103] in is the loss function weight.

[0104] Step 6-2: Use Bayesian optimization to adjust the learning rate of the Transformer prediction model, and add L2 regularization (λ=0.01) and early stopping mechanism (patience=10) to the model training to prevent overfitting. At the same time, use the Adam optimizer for gradient descent optimization training, as shown below:

[0105] ,

[0106] in and are model parameters and learning rate, respectively.

[0107] Step 6-3: Use 5-fold cross-validation to ensure the generalization ability of the model under different working conditions (such as high load and low load), and evaluate the model's accuracy, F1 score, and recall rate.

[0108] Step 7: Fault warning and maintenance strategy generation. This step uses the Z-score to calculate the threshold based on historical data and generates corresponding maintenance actions based on different fault probabilities to determine whether to shut down the machine for immediate inspection or regular maintenance. The specific steps include the following:

[0109] Step 7-1: Use Z-score to calculate and set the fault warning threshold from historical data:

[0110] ,

[0111] If exponential moving average is used, an early warning is triggered when the short-term trend threshold of oil particle concentration reaches EMA + 3σ (e.g. 150 µm).

[0112] Step 7-2: Generate maintenance recommendations based on the probability of failure, such as immediate shutdown for inspection if the probability is high (e.g., if the probability of oil oxidation degree > 8 exceeds 0.8, immediately shut down for inspection and replace the oil); and regular maintenance for low probability (e.g., if RUL > 200 hours, regular maintenance every 150 hours):

[0113] .

[0114] The present invention has the following main beneficial technical effects:

[0115] 1. Improve the accuracy of fault prediction: Through cross-modal correlation analysis, hidden connections between different sensor data and knowledge graph information are revealed to capture early and subtle fault characteristics.

[0116] 2. Improve the ability to analyze the root cause of faults: Through the topological structure and rich associated information of the knowledge graph, when a potential fault is predicted, it can be quickly traced to the possible faulty components, failure modes and their propagation paths, facilitating the precise location of the core of the problem.

[0117] 3. Enhanced comprehensive perception and cognitive capabilities: By integrating multimodal perception data (such as vibration, temperature, pressure and other sensor data) and knowledge graphs (including structured knowledge such as equipment structure, historical failures, maintenance records, expert rules, etc.), a more comprehensive and in-depth understanding of system status is achieved, breaking through the limitations of a single data source or model.

[0118] 4. Enhanced decision-making support capabilities: The fault warning and maintenance strategy generation part not only predicts the probability of failure, but also integrates the relevant information provided by the knowledge graph (such as spare parts status, maintenance costs, scope of impact, and maintenance best practices) to generate smarter and more actionable maintenance suggestions and response strategies, assisting operation and maintenance personnel in making optimal decisions and reducing unplanned downtime.

[0119] 5. Improve the interpretability and credibility of the model: As a background knowledge base, the knowledge graph provides semantic explanation and support for the prediction results of the data-driven model, enhancing the credibility of the model results and user acceptance.

[0120] 6. Realize a closed loop of intelligent operation and maintenance throughout the entire process: This method forms a complete closed loop of intelligent analysis and prediction from data collection, processing, analysis, model building to final early warning and decision support, significantly improving the intelligence level and efficiency of operation and maintenance of key equipment (oil and gas pressure systems) of hydropower stations.

[0121] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions described in the claims, including equivalent alternatives to the technical features of the technical solutions described in the claims. Equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graphs, characterized by: The following steps are involved: Step 1: Multimodal data acquisition and preprocessing; Step 2: Feature extraction and cross-modal correlation analysis; Step 3: Knowledge graph construction and dynamic update; Step 4: Fusion of physical mechanism model and data-driven model; Step 5: Fault mode library construction and knowledge reasoning; Step 6: Multimodal prediction model training and optimization; Step 7: Generate fault warning and maintenance strategy.

2. The method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 1, the vibration, liquid level, and pressure sensors installed on the oil and gas pressure tank are used to sense its status data in real time, eliminate fluctuations and noise in the sensed data, and use a filtering algorithm to smooth the sensed data. The specific sub-steps are as follows: Step 1-1: Install vibration sensors on the bearing seat, gearbox housing, and seal connections of the pressure oil tank to monitor the equipment vibration frequency in real time. Use a magnetic oil particle sensor to detect the metal particle concentration, and a spectrum analyzer to detect the oil oxidation product index and liquid level. Pressure sensors are installed at the oil pump outlet, the top of the oil tank, and the seal interface to collect pressure. Use an infrared thermal imager to monitor the surface temperature distribution of the oil tank. Synchronize the above sensor data to the centralized monitoring center via Industrial Ethernet or OPC UA protocol. Each sensor samples at 1 kHz, and its value is recorded as follows: ,in is the time domain perception signal, is the amplitude of the i-th sensor, is the sampling frequency of the i-th sensor, is the phase value of the i-th sensor used for calibration; Step 1-2: Perform N-layer discrete wavelet decomposition on the perception signal to obtain the corresponding detail coefficients and approximation coefficients, as shown in the formula: ,in is the coefficient of the Nth layer wavelet decomposition, that is, the relevant low-frequency component; is the detail coefficient of the kth layer, corresponding to high-frequency components of different scales; Process the decomposition coefficients using soft thresholding: , , c is the original coefficient (such as and ), is the threshold, is the noise standard deviation, M is the signal length; The signal strength after reconstructing the perception signal and eliminating noise is: ; Step 1-3: Perform Z-score normalization on the oil and gas pressure system sensing data. The normalization method is: mean 0, standard deviation 1; to ensure that the data from different sensors are in the same dimension, as shown in the following formula: And use the 3σ principle to eliminate outliers in the pressure sensor data: ,in and are the mean and variance respectively.

3. The method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 2, the time domain, frequency domain, and operation-related features of the mechanical wear and vibration of the oil and gas pressure system reflected in the sensing signal are extracted to achieve cross-modal alignment of multi-source sensors and better serve early fault detection. This specifically includes the following sub-steps: Step 2-1: Extract and calculate the time domain features of each sensing signal of the oil and gas pressure system, including the root mean square value, peak-to-peak value, and kurtosis features, to detect bearing wear or gear cracks; The RMS value is calculated as follows to measure the energy intensity and vibration intensity of the signal: , is the i-th sampling point of signal x, and N is the total number of sampling points used to calculate the signal; The peak-to-peak value calculation formula is as follows, which is used to measure the dynamic range and impact characteristics of the signal, especially the abnormal characteristics: ; The kurtosis characteristic calculation formula is as follows to reflect the signal peak characteristics: , as shown above, and are the mean and variance of the signal respectively; Step 2-2: Use fast Fourier transform to extract the main frequency components and combine it with envelope analysis to detect early faults such as crack resonance, as shown in the formula: ,in is the frequency signal, f is the frequency variable, n is the nth frequency cost, and N is the total frequency component; Step 2-3: Extract oil-related features, including the temperature difference between the tank surface and the ambient temperature, the acid value and moisture content characteristics of the oil spectral analysis, and the metal particle concentration characteristics; Step 2-4: Use the Pearson correlation coefficient to analyze the relationship between the vibration signal and the oil particle concentration. This is used to extract the coupling relationship between the metal particle concentration and the vibration spectrum shift caused by oil pump wear for subsequent knowledge graph construction. The specific relationship is shown in the formula: , r is the correlation coefficient, xi and yi are the values ​​of the two sensors or modes respectively, and are the means of n samples respectively.

4. The method for predicting oil and gas pressure system faults in a hydropower station by integrating sensory data and knowledge graphs according to claim 1 is characterized by: In step 3, a fault-related knowledge graph is constructed based on the working mechanism of the oil and gas pressure system to provide clearer entity relationships and the correlation mapping relationship between performance and faults, and is dynamically updated in a timely manner. The specific sub-steps include: Step 3-1: Use named entity method to extract equipment entities and fault types: ,in is the entity type (such as "device"), for properties (such as "pressure"), is the attribute value (such as "1.5MPa"); Step 3-2: Extract the monitoring relationship and causal relationship between the equipment entity fault and the perception data features according to the working mechanism specification; Step 3-3: Use a large language model or natural language processing tool to extract equipment failures, related indicators, and causal relationships from the work mechanism specifications and technical analysis maintenance manuals, build relevant knowledge graphs, and use graph neural networks to represent and model entity relationships: ,in is the embedding vector of the l-th layer node, is the adjacency matrix, is the learnable weight matrix, is the activation function; Step 3-4: Update the knowledge graph based on the monitoring attribute indicators and fault phenomena during real-time operation, specifically expressed as follows: ,in To add new entities or attributes (such as "Oil Pump", "Particle Concentration"), To add new relationships (such as connection, cause and effect).

5. The method for predicting oil and gas pressure system faults in a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 4, the knowledge graph generated in step 3 is used to further build a physical mechanism model, integrate the data-driven prediction model, and simulate the oil flow and pressure model of the oil and gas pressure system to eliminate the influence of noise. At the same time, it can better predict the oil pump outlet pressure fluctuation and accurately predict the oil and gas system failure. The specific sub-steps include the following: Step 4-1: Use the fluid dynamics model to establish a correlation model between oil fluid density, velocity, and pressure, and predict the pressure fluctuation at the oil and gas pressure tank port: ,in is the fluid density, is the fluid velocity, For pressure, is the dynamic viscosity, is an external force; Step 4-2: Use the Fourier heat conduction equation: ∂T / ∂t = α(∂²T / ∂x²) to calculate the surface temperature distribution of the oil tank and predict the thermal aging rate of the seal in combination with the thermal boundary conditions; Step 4-3: Divide the sensor signal data into 1024-point blocks and use the Transformer model to extract temporal dependencies for predicting the remaining service life of bearing wear: , , are input data, query, key, and value matrices respectively; Step 4-4: Perform weighted fusion of the data obtained from the physical model and the predicted value of the data model to reduce the impact of data noise on the prediction results: , such as the oil and gas port pressure physical model is the value obtained by kinetic model simulation calculation in step 4-1, is the predicted value obtained in step 4-3; α is the model weight, which is taken as 0.

6.

6. The method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 5, based on the previous steps and historical data, the fault patterns are extracted and a network model is established for fault classification and causal reasoning. This includes the following sub-steps: Step 5.1: Build a failure mode library based on historical failure cases and classify the failure modes into mechanical failures and oil failures. Mechanical failures include bearing wear, gear breakage, and seal aging, while oil failures include emulsification, oxidation, and excessive moisture. The corresponding parameter features for each failure mode are as follows: ,in The kth fault feature is used as the basis for classifying historical fault data through machine learning methods such as random forest and automatically adding new fault patterns. Step 5.2: Use the Monte Carlo simulation method to perform probabilistic prediction reasoning on oil and gas faults based on the fault pattern library and attribute characteristics: ,in The posterior probability after the feature appears, such as performing multiple Monte Carlo simulations on the temperature field of the oil tank to predict the probability that the thermal stress of the seal exceeds the critical value, triggering a maintenance warning; Step 5.3: Use Bayesian network for causal reasoning and determine the probability of related faults based on attribute indicators, as shown in the formula: .

7. The method for predicting faults in the oil and gas pressure system of a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 6, the end-to-end model involved in the above steps is trained and verified. The loss function is designed based on the training results of the above model to balance the minimum mean square error and cross entropy, and the training learning rate is optimized and adjusted according to the data features. The results are then verified. The specific sub-steps include the following: Step 6-1: Use weighted indicators as the loss function to balance the minimum mean square error and cross entropy, optimize the prediction accuracy of the physical model through MSE, and meet the probability prediction accuracy of fault type identification through CrossEntropy: ,in is the loss function weight; Step 6-2: Use Bayesian optimization to adjust the learning rate of the Transformer prediction model, and add L2 regularization and early stopping mechanism to the model training to prevent overfitting. At the same time, use the Adam optimizer for gradient descent optimization training, as shown below: ,in and are model parameters and learning rate respectively; Step 6-3: Use 5-fold cross-validation to ensure the generalization ability of the model under different working conditions and evaluate the model's accuracy, F1 score, and recall rate.

8. The method for predicting oil and gas pressure system faults in a hydropower station by integrating sensory data and knowledge graph according to claim 1 is characterized by: In step 7, a Z-score is used to calculate a threshold based on historical data, and corresponding maintenance actions are generated based on different failure probabilities to determine whether to shut down the machine for immediate inspection or scheduled maintenance. The specific steps include the following: Step 7-1: Use Z-score to calculate and set the fault warning threshold from historical data: ,When using exponential moving average calculation, when the short-term trend threshold of oil particle concentration is EMA + 3σ, an early warning is initiated; Step 7-2: Generate maintenance recommendations based on the probability of failure. If the probability is high, the machine should be shut down for inspection immediately; if the probability is low, regular maintenance should be performed. .

Images