Hydropower station governor main pressure distribution valve fault diagnosis early warning system and method

Abstract

The application provides a hydroelectric station governor main pressure distribution valve fault diagnosis early warning system and method, relates to the technical field of hydroelectric station governor main pressure distribution valve fault diagnosis. The hydroelectric station governor main pressure distribution valve fault diagnosis early warning system comprises a governor main pressure distribution valve and a fault diagnosis early warning system, a plurality of sensors are arranged on the governor main pressure distribution valve, which are used for monitoring various operation parameters of the main pressure distribution valve in real time, pressure, temperature and displacement, the early warning system is connected with the sensors, can analyze the collected data in real time, and the fault diagnosis early warning system comprises a multi-modal sensor array, an edge computing unit and a digital twin platform. Through the collaborative design of multi-modal fusion anti-interference, edge computing real-time response and meta-learning small sample adaptation, the core pain points of traditional methods in data dependence, real-time performance and false alarm rate are solved, and a high-reliability and high-precision intelligent operation and maintenance solution for the hydroelectric station speed regulation system is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis of the main pressure distribution valve of the governor of a hydropower station, in particular to a fault diagnosis and early warning system and method for the main pressure distribution valve of the governor of a hydropower station. BACKGROUND

[0002] A hydropower station is a renewable energy facility that utilizes the kinetic and potential energy of water flow to generate electricity. It converts the energy of water into mechanical energy through a water turbine, and then converts the mechanical energy into electrical energy through a generator. The governor of a hydropower station is one of the core control devices of the generator unit of the hydropower station, mainly used to adjust the speed and output power of the water turbine to ensure stable operation of the generator unit. Its main function is to automatically adjust the guide vane opening of the water turbine according to the load change of the power grid or the command signal, thereby controlling the speed of the water turbine and the output power of the generator to maintain the stability of the power grid frequency. In order to achieve accurate adjustment of the guide vane opening, a main pressure distribution valve is used to connect the governor control system and the water turbine actuator. According to the control signal of the governor, the flow of hydraulic oil entering the guide vane is accurately adjusted, and by changing the position of the valve core, the oil path is controlled and the oil flow direction is controlled, thereby adjusting the guide vane opening. The main pressure distribution valve is the core control component of the governor system of a hydropower station. Its failure may cause the guide vane to lose control, and further cause serious accidents such as unit overspeed and runaway, which poses a major threat to the safety of hydropower station equipment and the power grid. Therefore, it is of great importance and necessity to set up a fault diagnosis and early warning system for the main pressure distribution valve of the governor of a hydropower station.

[0003] There are still some problems in the use of the existing fault diagnosis and early warning system for the main pressure distribution valve of the governor of a hydropower station. The traditional fault diagnosis system for the main pressure distribution valve of the governor of a hydropower station mainly adopts a cloud centralized computing mode, i.e. the data of the field sensors are transmitted to a central server through a network for processing and analysis. Although it has real-time simulation function, the traditional cloud architecture cannot meet the millisecond-level response requirement of the governor. When the main pressure distribution valve is abnormal, the main pressure distribution valve frequently moves up and down. The delay of cloud computing will cause the diagnosis result to lag behind, missing the best intervention opportunity. Therefore, the technical personnel in this field provide a fault diagnosis and early warning system for the main pressure distribution valve of the governor of a hydropower station to solve the problems in the above background technology. SUMMARY

[0004] (I) Technical problems solved

[0005] In view of the deficiencies of the prior art, the water power station governor main pressure distribution valve fault diagnosis early warning system and method provided by the present application solves the problem that the traditional water power station governor main pressure distribution valve fault diagnosis system mainly adopts a cloud centralized computing mode, that is, the field sensor data is transmitted to a central server through a network for processing and analysis, although the real-time simulation function is provided, the traditional cloud architecture is difficult to meet the millisecond level response requirement of the governor, when the main pressure distribution valve is abnormal, the main pressure distribution valve frequently moves up and down, and the cloud computing delay will cause the diagnosis result to lag, and the best intervention opportunity is missed.

[0006] (II) Technical solutions

[0007] To achieve the above object, the present application is realized by the following technical solutions: a water power station governor main pressure distribution valve fault diagnosis early warning system, comprising a governor main pressure distribution valve and a fault diagnosis early warning system, a plurality of sensors are arranged on the governor main pressure distribution valve for real-time monitoring of pressure, temperature and displacement parameters of the main pressure distribution valve, the early warning system is connected with the sensors, and the collected data can be analyzed in real time;

[0008] The fault diagnosis early warning system comprises a multi-modal sensor array for real-time collection of operation state data of the governor main pressure distribution valve;

[0009] An edge computing unit is in communication connection with the multi-modal sensor array, a lightweight meta-learning model is deployed to analyze the operation state data in real time, the edge computing unit is connected with a water power station SCADA system through an RS-485 / Modbus TCP interface, real-time fault warning information is pushed to a monitoring large screen, and historical data is stored to an SQL database;

[0010] A digital twin platform is in communication connection with the edge computing unit, and real-time mapping of the physical device state and generation of a dynamic early warning strategy are realized.

[0011] Preferably, the multi-modal sensor array comprises a fiber grating vibration sensor arranged on the surface of the main pressure distribution valve body for collecting vibration frequency spectrum data of the valve core, the sampling frequency is 10 kHz, and whether the energy proportion of 2-5 times the rotation frequency exceeds 15% is monitored;

[0012] An ultrasonic oil particle size on-line monitor is arranged in the interior of the governor oil tank for detecting the oil pollution degree, and an early warning is triggered when the particle size level rises from 6 to 8;

[0013] A piezoelectric pressure sensor array is arranged at the inlet of the main pressure distribution valve control cavity, the sampling frequency is 10 kHz, and the pressure pulsation signal is collected for identifying pressure fluctuation abnormalities;

[0014] Non-contact laser displacement sensor is arranged on the stroke path of the main pressure regulating valve core, with an accuracy of ±0.01 mm, and is used for real-time monitoring of the core displacement signal, replacing the traditional potentiometer to solve the temperature drift problem.

[0015] Preferably, the meta-learning model comprises a pre-training module, a meta-learner is constructed using the historical fault data of the Three Gorges and Gezhouba hydropower stations based on a public hydropower dataset, and a transfer learning framework is adopted in the training process, and the model architecture is a hybrid neural network.

[0016] When the field adaptation module is deployed at the target hydropower station, only 10-20 fault samples are needed to fine-tune the pre-trained model, dynamically optimize the feature extraction layer parameters, and adapt to the operating characteristics of the vibration, oil pressure and displacement parameters under the conditions of unit no-load, grid connection and heavy load.

[0017] Preferably, the digital twin platform comprises a three-dimensional modeling unit, a high-precision three-dimensional model of the main pressure regulating valve is constructed based on a general finite element analysis software, and the hydraulic system parameters, oil pressure, flow rate, mechanical valve core displacement and friction coefficient are integrated.

[0018] The real-time mapping unit communicates with the governor control system through a communication protocol in the field of industrial automation, obtains sensor data in real time and updates the virtual model state, and the mapping frequency is 100 Hz.

[0019] The simulation verification unit performs digital twin simulation on suspected fault scenarios, calculates parameters such as valve core natural frequency and pressure response curve, and verifies the fault diagnosis results.

[0020] The data update delay of the digital twin platform is less than 10 ms, the virtual model visualization interface supports three-dimensional cross-sectional display, and the valve core displacement, pressure field distribution and oil flow line dynamics can be viewed in real time.

[0021] Preferably, the multi-modal sensor array further comprises a mounting structure, the mounting structure comprises a fiber grating vibration sensor fixing arm made of 304 stainless steel and fixed to the surface of the main pressure regulating valve body by bolts, the sensor probe and the valve body contact surface are coated with heat-conducting silicone grease to ensure that the vibration signal transmission efficiency is greater than 90%, the ultrasonic oil particle size instrument mounting seat is arranged at the bottom of the governor oil tank, a blowdown valve is arranged at the bottom, the probe surface is automatically cleaned regularly to avoid the accumulation of oil sludge affecting the measurement accuracy, and the piezoelectric pressure sensor array support is isolated from mechanical vibration by shock-absorbing rubber pads, and the support is connected with the main pressure regulating valve control cavity inlet flange.

[0022] The method for fault diagnosis and early warning of the main pressure regulating valve of the governor of the hydropower station comprises the following steps:

[0023] S1. Multi-modal data acquisition and edge real-time processing;

[0024] S1.1. Isomeric sensor deployment and parameter configuration, fiber grating vibration sensor, installed on the surface of the main pressure regulating valve valve body at 3 measuring points, radial and axial positions, real-time transmission of wavelength change data through grating demodulator, oil particle size monitoring installed at the bottom of the governor oil tank and the oil return pipeline, monitoring the change of oil pollution level, piezoelectric pressure sensor array installed at the inlet of the main pressure regulating valve control cavity, forming a pressure field spatial distribution monitoring, non-contact laser displacement sensor installed directly above the valve core stroke path, real-time displacement calculation through triangulation method;

[0025] S1.2. Edge computing unit data preprocessing, hardware platform uses NVIDIA Jetson AGX Orin edge computing module, equipped with ARM architecture CPU+GPU, inference speed 2000FPS, 5-point sliding window mean filtering is used for vibration signal to eliminate high-frequency noise, its formula is:

[0026] ;

[0027] Wherein, is the original vibration signal amplitude at the th time point, is the filtered amplitude at the th time point;

[0028] Based on the 3σ criterion, the instantaneous jump data of oil particle size is eliminated;

[0029] S1.3. Normalize the vibration spectrum (0-10kHz), oil pressure (0-40MPa) and displacement (0-50mm) parameters to the [-1, 1] interval respectively;

[0030] S2. Multi-modal feature fusion and meta-learning diagnosis;

[0031] S2.1. Time-frequency analysis branch, input vibration signal STFT time-frequency spectrum, network structure, convolution layer 1: 3x3 convolution kernel, 64 channels, ReLU activation, pooling layer 2x2, convolution layer 2: 3x3 convolution kernel, 128 channels, LeakyReLU activation, fully connected layer output 128-dimensional frequency domain feature vector, time series prediction branch input oil pressure and displacement time series, network structure BiLSTM layer: 128 hidden units, bidirectional propagation, Attention layer calculates time step weight, output 64-dimensional time series feature vector;

[0032] S2.2. Multi-modal feature fusion, input frequency domain features + time series features concatenated into a 192-dimensional vector, encoder 6-layer Transformer block, each layer containing self-attention, feedforward network, attention weight calculation output 256-dimensional multi-modal feature vector after weighted fusion;

[0033] S3. Meta-learning model inference;

[0034] S3.1. Model architecture, pre-training phase based on fault samples of Three Gorges hydropower station historical data, using hybrid neural network to train meta-learner, save model parameters, target power station only needs 20 fault samples to fine-tune top fully connected layer, freeze bottom feature extraction layer;

[0035] S3.2. Fault probability calculation, output layer Softmax function generates 4-class fault probability, main matching dynamic, proportional valve fault, servomotor internal leakage and oil deterioration, trigger condition is any fault probability >= 95% and lasts for 3 seconds, trigger early warning;

[0036] S4. Dynamic threshold early warning and digital twin simulation verification;

[0037] S4.1. Prediction model training, input historical vibration spectrum, oil pressure and displacement data, sliding window length 200 points, network structure 2 layers of LSTM and fully connected layer, predict future 5 seconds parameter normal range, then calculate threshold, normal fluctuation range threshold = predicted value ± 3σ, where σ is the predicted residual standard deviation, automatically tighten the threshold dynamic adjustment according to the oil temperature change rate;

[0038] S4.2. Digital twin platform simulation verification, build main matching valve 3D model, integrate hydraulic system parameters, oil pressure, flow and valve core friction coefficient, get sensor data from edge computing unit through OPC UA protocol, drive model state update, when suspected fault triggers, digital twin platform runs the following simulation, apply step control command, monitor valve core displacement overshoot > 15% and oscillation times > 3 times, simulate particle size Class 8, verify system pressure loss > 5%, finally generate decision, if multi-modal fusion result is consistent with simulation verification, trigger sound and light alarm and push strategy, switch standby governor or adjust PID parameters, if not consistent, mark as "to be observed" and trigger manual review process.

[0039] Preferably, the multi-modal feature fusion further comprises the following sub-steps:

[0040] A1. Wavelet denoising processing is performed on the vibration spectrum data to suppress environmental noise interference;

[0041] A2. D-S evidence theory is used to fuse the output results of CNN branch and BiLSTM branch at decision layer, and fault probability is calculated.

[0042] Preferably, the dynamic threshold early warning comprises: real-time monitoring of oil temperature change rate, when the oil temperature rise rate > 0.5℃ / min, the alarm thresholds of oil pressure and vibration parameters are automatically tightened, according to the governor operating conditions, empty load, grid connection and heavy load are divided into cluster centers, and Mahalanobis distance is used to calculate the deviation degree of feature vector.

[0043] (Three) beneficial effects

[0044] The application provides a hydroelectric station governor main pressure distribution valve fault diagnosis and early warning system and method. The following beneficial effects are provided:

[0045] In the application, the multi-modal data cross-validation reduces the false positive rate from 8% of the traditional threshold method to <2%. When the oil temperature rises at a rate of >0.5℃ / min, the alarm threshold is dynamically adjusted to avoid false negatives caused by a fixed threshold.

[0046] In the application, the local edge end completes data cleaning, feature extraction and preliminary diagnosis, avoids cloud transmission delay, and still guarantees fault detection continuity when the network of remote hydroelectric stations is unstable.

[0047] In the application, a high-accuracy model can still be quickly trained in a data-scarce scenario, and the model can be migrated to different hydroelectric stations and only needs a small amount of data to adapt.

[0048] In the application, through the collaborative design of multi-modal fusion anti-interference, edge computing real-time response and meta-learning small sample adaptation, the core pain points of traditional methods in data dependence, real-time performance and false positive rate are solved, and a high-reliability and high-precision intelligent operation and maintenance solution for the hydroelectric station speed regulation system is provided. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The system schematic diagram of the application is shown. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0051] Embodiment:

[0052] As shown in Figure 1 The application embodiment provides a hydroelectric station governor main pressure distribution valve fault diagnosis and early warning system, which comprises a governor main pressure distribution valve and a fault diagnosis and early warning system. A plurality of sensors are arranged on the governor main pressure distribution valve for real-time monitoring of pressure, temperature and displacement parameters of the main pressure distribution valve. The early warning system is connected with the sensors and can perform real-time analysis on the collected data.

[0053] The fault diagnosis and early warning system comprises a multi-modal sensor array for collecting real-time operation state data of the governor main pressure distribution valve, a communication link of the multi-modal sensor array adopts a dual-channel redundant design, including an RS-485 wired channel and a LoRa wireless channel, and any channel failure is automatically switched;

[0054] An edge computing unit is in communication connection with the multi-modal sensor array, a light-weight meta-learning model is deployed to analyze the operation state data in real time, the edge computing unit is connected with a hydropower station SCADA system through an RS-485 / Modbus TCP interface, real-time fault early warning information is pushed to a monitoring large screen, and historical data is stored to a SQL database, the edge computing unit adopts a dual-machine hot standby redundant design, when the main unit fails, it is automatically switched to the standby unit to ensure continuous operation of the system, and the switching time is less than 100 ms;

[0055] A digital twin platform is in communication connection with the edge computing unit, and a dynamic early warning strategy is generated by mapping the physical device state in real time.

[0056] Preferably, the multi-modal sensor array comprises a fiber grating vibration sensor arranged on the surface of the main pressure distribution valve body, for collecting valve core vibration frequency spectrum data, the sampling frequency is 10 kHz, and whether the energy proportion of 2-5 times the rotation frequency exceeds 15% is monitored;

[0057] An ultrasonic oil particle size online monitor is arranged inside the governor oil tank, for detecting oil pollution degree, and an early warning is triggered when the particle size level rises from level 6 to level 8;

[0058] A piezoelectric pressure sensor array is arranged at the inlet of the main pressure distribution valve control cavity, with a sampling frequency of 10 kHz, for collecting pressure pulsation signals and identifying pressure fluctuation abnormalities;

[0059] A non-contact laser displacement sensor is arranged on the valve core stroke path of the main pressure distribution valve, with an accuracy of ±0.01 mm, for monitoring valve core displacement signals in real time, and replacing the traditional potentiometer to solve the temperature drift problem.

[0060] The meta-learning model comprises a pre-training module, based on public hydropower data sets, a meta-learner is constructed using historical fault data of the Three Gorges and Gezhouba hydropower stations, a transfer learning framework is adopted in the training process, and the model architecture is a hybrid neural network;

[0061] When the on-site adaptation module is deployed at the target hydropower station, only 10-20 fault samples are needed to fine-tune the pre-trained model, dynamically optimize the feature extraction layer parameters, and adapt to the operation characteristics of vibration, oil pressure and displacement parameters under the conditions of unit no-load, grid connection and heavy load.

[0062] The digital twin platform comprises a three-dimensional modeling unit, a high-precision three-dimensional model of the main pressure regulating valve is constructed based on a general finite element analysis software, and hydraulic system parameters, oil pressure, flow and mechanical valve core displacement and friction coefficient are integrated;

[0063] The real-time mapping unit communicates with the speed regulator control system through a communication protocol in the field of industrial automation, obtains sensor data in real time and updates the virtual model state, and the mapping frequency is 100 Hz;

[0064] The simulation verification unit performs digital twin simulation on suspected fault scenarios, calculates valve core natural frequency, pressure response curve and other parameters, and verifies fault diagnosis results;

[0065] The data update delay of the digital twin platform is less than 10 ms, the virtual model visual interface supports three-dimensional section display, the valve core displacement, pressure field distribution and oil flow line dynamics can be viewed in real time, the three-dimensional model of the digital twin platform integrates the calculation module of ANSYS or COMSOL software, supports online adjustment of hydraulic parameters and re-simulation, and the simulation results are rendered to the browser end through WebGL technology.

[0066] The multi-modal sensor array further comprises a mounting structure, the mounting structure comprises a fiber grating vibration sensor fixing arm made of 304 stainless steel and fixed to the surface of the main pressure regulating valve body by bolts, the sensor probe is coated with heat-conducting silicone grease on the contact surface with the valve body to ensure that the vibration signal transmission efficiency is greater than 90%, the ultrasonic oil particle size instrument mounting seat is arranged at the bottom of the speed regulator oil tank, a blowdown valve is arranged at the bottom, the probe surface is automatically cleaned regularly to avoid the accumulation of oil sludge affecting the measurement accuracy, and the piezoelectric pressure sensor array support is isolated from mechanical vibration by shock-absorbing rubber pads, and the support is connected with the main pressure regulating valve control cavity inlet flange.

[0067] The method for fault diagnosis and early warning of the main pressure regulating valve of the speed regulator of the hydropower station comprises the following steps:

[0068] S1. Multi-modal data acquisition and edge real-time processing;

[0069] S1.1. Heterogeneous sensor deployment and parameter configuration, the fiber grating vibration sensor is installed on the surface of the main pressure regulating valve body at three measuring points in the radial and axial positions, wavelength change data is transmitted in real time through the grating demodulator, the oil particle size monitor is installed at the bottom of the speed regulator oil tank and the oil return pipeline to monitor the oil pollution level change, and the piezoelectric pressure sensor array is installed at the inlet of the main pressure regulating valve control cavity to form a pressure field spatial distribution monitor, and the non-contact laser displacement sensor is installed directly above the valve core stroke path to calculate the displacement in real time through the triangulation method;

[0070] S1.2. Edge computing unit data preprocessing, hardware platform uses NVIDIA Jetson AGX Orin edge computing module, carries ARM architecture CPU+GPU, inference speed 2000FPS, 5-point sliding window mean filtering is used for vibration signal to eliminate high-frequency noise, its formula is:

[0071] ;

[0072] wherein, is the original vibration signal amplitude of the th time point, is the filtered amplitude of the th time point;

[0073] Based on 3σ criterion, the instantaneous jump data of oil particle size is eliminated;

[0074] S1.3. The vibration spectrum (0-10kHz), oil pressure (0-40MPa) and displacement (0-50mm) parameters are normalized to [-1, 1] interval respectively;

[0075] S2. Multi-modal feature fusion and meta-learning diagnosis;

[0076] S2.1. Time-frequency analysis branch, input vibration signal STFT time-frequency spectrum, network structure, convolution layer 1: 3x3 convolution kernel, 64 channels, ReLU activation, pooling layer 2x2, convolution layer 2: 3x3 convolution kernel, 128 channels, LeakyReLU activation, full connection layer output 128-dimensional frequency domain feature vector, time series prediction branch input oil pressure and displacement time series, network structure BiLSTM layer: 128 hidden units, bidirectional propagation, Attention layer calculates time step weight, output 64-dimensional time series feature vector;

[0077] S2.2. Multi-modal feature fusion, input frequency domain features + time series features are spliced into 192-dimensional vectors, encoder 6-layer Transformer block, each layer contains self-attention, feedforward network, attention weight calculation outputs 256-dimensional multi-modal feature vectors after weighted fusion;

[0078] S3. Meta-learning model inference;

[0079] S3.1. Model architecture, pre-training phase based on fault samples of Three Gorges hydropower station historical data, using hybrid neural network to train meta-learner, save model parameters, target power station only needs 20 fault samples to fine-tune the top full connection layer, freeze the bottom feature extraction layer;

[0080] S3.2. Failure probability calculation, the output layer Softmax function generates 4 categories of failure probabilities, main matching twitch, proportional valve failure, servomotor internal leakage and oil deterioration, the trigger condition is that any failure probability ≥ 95% and lasts for 3 seconds, triggering the early warning;

[0081] S4. Dynamic threshold early warning and digital twin simulation verification;

[0082] S4.1. Prediction model training, input historical vibration spectrum, oil pressure and displacement data, sliding window length 200 points, network structure 2 layers of LSTM and full connection layer, predict future 5 seconds parameter normal range, then calculate threshold value, normal fluctuation range threshold = predicted value ± 3σ, where σ is the predicted residual standard deviation, automatically tighten the threshold value dynamically according to the oil temperature change rate;

[0083] S4.2. Digital twin platform simulation verification, build a 3D model of the main matching pressure valve, integrate hydraulic system parameters, oil pressure, flow and valve core friction coefficient, get sensor data from edge computing unit through OPC UA protocol, drive model state update, when suspected failure is triggered, digital twin platform runs the following simulation, applies step control command, monitors valve core displacement overshoot > 15% and oscillation times > 3 times, simulates particle size Class 8, verifies system pressure loss > 5%, finally generates a decision, if the multimodal fusion result is consistent with the simulation verification, trigger audible and light alarm and push strategy, switch standby governor or adjust PID parameters, if not consistent, mark as "to be observed" and trigger manual review process

[0084] The multi-modal feature fusion further includes the following sub-steps:

[0085] A1. Wavelet denoising processing is performed on the vibration spectrum data to suppress environmental noise interference.

[0086] A2. D-S evidence theory is used to fuse the output results of the CNN branch and the BiLSTM branch at the decision layer to calculate the failure probability.

[0087] The dynamic threshold early warning includes real-time monitoring of the oil temperature change rate, when the oil temperature rise rate > 0.5℃ / min, the alarm threshold of oil pressure and vibration parameters is automatically tightened, according to the governor operating conditions, empty load, grid connection and heavy load are divided into cluster centers, and Mahalanobis distance is used to calculate the deviation degree of feature vector.

[0088] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing and early warning of faults of a main pressure distribution valve of a hydroelectric power station governor, characterized in that: Comprising the following steps: S1. Multi-modal data acquisition and edge real-time processing; S1.

1. Heterogeneous sensor deployment and parameter configuration, multi-modal sensor array includes fiber grating vibration sensor, set on the surface of the main pressure regulating valve, fiber grating vibration sensor, installed on the surface of the main pressure regulating valve at 3 measuring points, radial and axial positions, real-time transmission of wavelength change data through grating demodulator, oil particle size monitoring is installed at the bottom of the governor oil tank and the oil return pipeline to monitor the oil pollution level change, piezoelectric pressure sensor array is installed at the inlet of the main pressure regulating valve control cavity to form a pressure field space distribution monitoring, non-contact laser displacement sensor is installed directly above the valve core stroke path to calculate the displacement amount in real time through triangulation method; S1.

2. Data preprocessing of edge computing unit, hardware platform uses NVIDIA Jetson AGX Orin edge computing module, equipped with ARM architecture CPU+GPU, inference speed 2000FPS, 5-point sliding window mean filtering is used for vibration signal to eliminate high-frequency noise, its formula is: ； wherein, is the raw vibration signal amplitude at the th time point, is the filtered amplitude at the th time point; Based on 3σ criterion, the instantaneous jump data of oil particle size is removed; S1.

3. Normalizing the vibration spectrum 0-10kHz, oil pressure 0-40MPa and displacement 0-50mm parameters to [-1, 1] interval respectively; S2. Multi-modal feature fusion and meta-learning diagnosis; S2.

1. Time-frequency analysis branch, input vibration signal STFT time-frequency spectrum, network structure, convolution layer 1: 3x3 convolution kernel, 64 channels, ReLU activation, pooling layer 2x2, convolution layer 2: 3x3 convolution kernel, 128 channels, LeakyReLU activation, full connection layer output 128-dimensional frequency domain feature vector, time series prediction branch input oil pressure and displacement time series, network structure BiLSTM layer: 128 hidden units, bidirectional propagation, Attention layer calculates time step weight, output 64-dimensional time series feature vector; S2.

2. Multi-modal feature fusion, input frequency domain features + time series features are spliced into 192-dimensional vectors, encoder 6-layer Transformer block, each layer contains self-attention, feedforward network, attention weight calculation outputs 256-dimensional multi-modal feature vector after weighted fusion; S3. Meta-learning model inference; S3.

1. Model architecture, in the pre-training stage, based on the fault samples of the Three Gorges Hydropower Station historical data, use hybrid neural network to train meta-learner, save model parameters, target power station only needs 20 fault samples to fine-tune the top full connection layer, freeze the bottom feature extraction layer; S3.

2. Fault probability calculation, output layer Softmax function generates 4-class fault probability, main pressure regulating valve fluctuation, proportional valve fault, servomotor internal leakage and oil deterioration, trigger condition is that any fault probability ≥95% and lasts for 3 seconds, trigger alarm; S4. Dynamic threshold early warning and digital twin simulation verification; S4.

1. Predictive model training, input historical vibration spectrum, oil pressure and displacement data, sliding window length 200 points, network structure 2 layers of LSTM and fully connected layer, predict future 5 seconds parameter normal range, then calculate threshold, normal fluctuation range threshold = predicted value ± 3σ, where σ is the standard deviation of the prediction residual, automatically tighten the threshold value according to the oil temperature change rate for dynamic adjustment; S4.

2. Digital twin platform simulation verification, build a three-dimensional model of the main pressure regulating valve, integrate hydraulic system parameters, oil pressure, flow and valve core friction coefficient, obtain sensor data from the edge computing unit through OPC UA protocol, drive model state update, when suspected failure triggers, the digital twin platform runs the following simulation, applies a step control command, monitors the valve core displacement overshoot > 15% and the oscillation frequency > 3 times, simulates a particle size of Class 8, verifies that the system pressure loss > 5%, and finally generates a decision, if the multi-modal fusion result is consistent with the simulation verification, trigger the sound and light alarm and push the strategy, switch to the standby governor or adjust the PID parameters, if not, mark as "to be observed" and trigger the manual review process.

2. The method for diagnosing and early warning of the failure of the main governing valve of the hydroelectric station governor according to claim 1, characterized in that: The dynamic threshold early warning monitors the oil temperature change rate in real time, when the oil temperature rise rate > 0.5℃ / min, the alarm threshold of oil pressure and vibration parameters is automatically tightened, according to the governor operating conditions, empty load, grid connection and heavy load are divided into cluster centers, and the Mahalanobis distance is used to calculate the deviation degree of the feature vector.

3. The method for diagnosing and early warning of the failure of the main governing valve of the hydroelectric station governor according to claim 1, characterized in that: The multi-modal feature fusion further includes the following sub-steps: A1. Wavelet denoising processing is performed on the vibration spectrum data to suppress environmental noise interference; A2. D-S evidence theory is used to fuse the output results of the CNN branch and the BiLSTM branch at the decision layer to calculate the fault probability.

4. The main pressure distributing valve fault diagnosis and early warning system of the hydroelectric power station governor, adopting the main pressure distributing valve fault diagnosis and early warning method of any one of claims 1 to 3, characterized in that: A plurality of sensors are arranged on the main pressure regulating valve of the governor for real-time monitoring of the pressure, temperature and displacement parameters of the main pressure regulating valve, the early warning system is connected with these sensors, and the collected data can be analyzed in real time; The fault diagnosis and early warning system comprises a multi-modal sensor array for real-time acquisition of the operating state data of the governor main pressure regulating valve; An edge computing unit is in communication connection with the multi-modal sensor array, a lightweight meta-learning model is deployed to analyze the operating state data in real time, the edge computing unit is connected with the hydropower station SCADA system through RS-485 / Modbus TCP interface, real-time fault warning information is pushed to the monitoring screen, and historical data is stored in the SQL database; A digital twin platform is in communication connection with the edge computing unit, which maps the physical device state in real time and generates a dynamic early warning strategy.

5. The hydroelectric power station governor main distributing valve fault diagnosis and early warning system according to claim 4, characterized in that: The digital twin platform comprises a three-dimensional modeling unit, which constructs a high-precision three-dimensional model of the main pressure regulating valve based on a general finite element analysis software, integrates hydraulic system parameters, oil pressure, flow and mechanical valve core displacement and friction coefficient; A real-time mapping unit communicates with the governor control system through a communication protocol in the field of industrial automation, obtains sensor data in real time and updates the virtual model state, and the mapping frequency is 100Hz; The simulation verification unit performs digital twin simulation on suspected fault scenarios, calculates the valve core natural frequency and pressure response curve parameters, and verifies the fault diagnosis results; The data update delay of the digital twin platform is less than 10 ms, the virtual model visualization interface supports three-dimensional section display, and the valve core displacement, pressure field distribution, and oil flow line dynamics can be viewed in real time.

6. The hydroelectric power station governor main distributing valve fault diagnosis and early warning system according to claim 4, characterized in that: The multi-modal sensor array also includes a mounting structure, which includes a fiber grating vibration sensor fixing arm made of 304 stainless steel and fixed to the surface of the main governing valve body by bolts. The sensor probe and the valve body contact surface are coated with heat-conducting silicone grease to ensure that the vibration signal transmission efficiency is greater than 90%. The ultrasonic oil particle size instrument mounting seat is arranged at the bottom of the governor oil tank, and a blowdown valve is arranged at the bottom to automatically clean the probe surface periodically and avoid the accumulation of oil sludge affecting the measurement accuracy. The piezoelectric pressure sensor array support is isolated from mechanical vibration by shock-absorbing rubber pads, and the support is connected with the main governing valve control cavity inlet flange.

7. The hydroelectric power station governor main distributing valve fault diagnosis and early warning system according to claim 4, characterized in that: The meta-learning model includes a pre-training module, which uses the historical fault data of the Three Gorges and Gezhouba hydropower stations to build a meta-learner based on public water and electricity data sets. The training process uses a transfer learning framework, and the model architecture is a hybrid neural network. When the field adaptation module is deployed at the target hydropower station, only 10-20 fault samples are needed to fine-tune the pre-trained model, dynamically optimize the feature extraction layer parameters, and adapt to the operating characteristics of vibration, oil pressure and displacement parameters under the conditions of unit no-load, grid connection and heavy load.

8. The hydroelectric power station governor main distributing valve fault diagnosis and early warning system according to claim 4, characterized in that: The fiber grating vibration sensor is arranged on the surface of the main governing valve body to collect valve core vibration spectrum data, with a sampling frequency of 10 kHz, and monitors whether the energy proportion of 2-5 times the rotation frequency exceeds 15%; The ultrasonic oil particle size online monitor is arranged inside the governor oil tank to detect oil contamination. When the particle size level rises from level 6 to level 8, a pre-warning is triggered. The piezoelectric pressure sensor array is arranged at the inlet of the main governing valve control cavity, with a sampling frequency of 10 kHz, for collecting pressure pulsation signals and identifying pressure fluctuation abnormalities. The non-contact laser displacement sensor is arranged on the valve core stroke path of the main governing valve, with an accuracy of ±0.01 mm, for real-time monitoring of valve core displacement signals, replacing traditional potentiometers to solve the problem of temperature drift.

Images