Water turbine guide vane water leakage detection method and system

Abstract

The invention discloses a water turbine guide vane water leakage detection method and system, belongs to the technical field of hydropower station equipment monitoring, and aims at solving the technical problems that in an existing guide vane water leakage detection method, precision is limited, water leakage points cannot be accurately judged, and shutdown inspection is relied on. A pressure sensor and an RTD temperature sensor are integrated, and real-time and accurate monitoring of pressure and temperature is achieved; the FFT spectrum analysis technology is combined with pressure calibration, so that the accuracy of water leakage detection is improved; meanwhile, a multi-stage judgment strategy and a feature fusion model based on LSTM, SVM and logistic regression are adopted, intelligent prediction and graded early warning of the water leakage trend are achieved, the precision and reliability of water leakage detection are effectively improved, the situations of false alarm and missing alarm are greatly reduced, the water leakage problem can be found and early warned in time, therefore, the equipment downtime is shortened, the maintenance cost is reduced, and the working efficiency is improved. And the operation efficiency and safety of the water turbine are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of water turbine operation monitoring in a hydropower station, and in particular to a water turbine guide vane leakage detection method and system. Background Art

[0002] In the field of hydropower station equipment monitoring technology, leak detection of turbine guide vanes has always been a key link in ensuring the safe and efficient operation of turbine units. Currently, leak detection of turbine guide vanes mainly relies on water level monitoring, flow meter testing, and manual visual inspection.

[0003] Water level monitoring methods typically indirectly infer leakage by monitoring changes in water levels upstream and downstream of the turbine. The basic principle is that abnormal water level changes may indicate a leak. Flow meter detection methods directly measure the flow rate through the turbine and compare it with the theoretical value to determine whether there is a leak. Manual visual inspection methods often require rotating the main shaft to observe whether there are signs of leakage at the guide vane seal while the turbine is shut down.

[0004] However, these existing technologies all have significant problems and shortcomings. Water level monitoring methods have limited accuracy, and water level changes can be affected by a variety of factors, making it difficult to accurately determine the location and extent of leaks. Flow meter detection methods also face accuracy issues and are unable to detect early signs of leaks in real time. Manual visual inspection methods are not only complex and require downtime, increasing operation and maintenance costs, but also have difficulty detecting subtle leak trends and cannot provide timely warnings of leak problems.

[0005] For example, CN110207905B discloses a guide vane leak detection system, which primarily includes a distance sensor, a central controller, and a liquid level detection device. Its operating principle is to use a liquid level gauge to detect changes in the liquid level at the vent to determine whether the guide vane is leaking. While this method can detect leaks to a certain extent, it still has the following drawbacks: Limited detection accuracy: Liquid level detection methods rely on changes in liquid level height to determine water leakage. However, in actual operation, the liquid level height may be affected by various factors, such as water flow fluctuations and temperature changes, which may lead to errors in the detection results. Inability to provide real-time warnings: Although the system can detect water leaks, it may not issue early warnings in the early stages of a leak, resulting in the leak not being addressed promptly and increasing the risk of equipment damage. Complex installation and maintenance: The installation and maintenance of liquid level detection devices require certain professional skills and experience, which increases operation and maintenance costs.

[0006] For example, CN114483417B discloses a method for quickly identifying water leakage defects in turbine guide vanes based on voiceprint recognition. This method sets multiple voiceprint collection points to collect voiceprint information near the turbine guide vanes and uses an RNN neural network model for training to identify water leakage defects in the guide vanes. Although this method improves the speed and accuracy of water leakage detection, it still has the following shortcomings: Environmental noise interference: Voiceprint recognition methods are easily affected by environmental noise. Especially in a noisy environment such as a hydropower station, the voiceprint signal may be seriously polluted, affecting the accuracy of the recognition results.

[0007] Complex model training: The training of RNN neural network models requires a large amount of sample data and computing resources, and the training process is relatively complex, requiring high professional capabilities of technical personnel.

[0008] Limited application scenarios: Although this method can quickly identify water leakage defects, its identification effect may be affected in certain specific scenarios (such as high-load operation of the unit, poor water quality, etc.).

[0009] In summary, existing turbine guide vane leakage detection technology still has many flaws and shortcomings, such as insufficient accuracy, inability to accurately identify the leak point, and reliance on shutdown inspections. These limitations make it unable to meet the requirements of modern hydropower stations for safe and efficient operation. Therefore, the development of a real-time, efficient, and accurate turbine guide vane leakage detection method and system has important practical significance and application value. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a method and system for detecting water leakage in guide vanes of turbines, so as to solve the specific technical problems existing in the field of water leakage detection in guide vanes of turbine units in hydropower stations. Specifically, the current water leakage detection in guide vanes mainly relies on methods such as water level monitoring, flow meter detection and manual visual inspection. However, these methods have limitations such as limited accuracy, inability to accurately determine the leakage point, and reliance on shutdown inspection, and cannot meet the requirements of modern hydropower stations for safe and efficient operation.

[0011] In order to solve the above technical problems, the present invention adopts the following technical solutions: Specifically, the present invention provides a method and system for detecting water leakage in a turbine guide vane, including the following key steps and modules: Sensor Installation and Data Collection: Piezoelectric sensors are installed near the upper and lower guide vane seals. An RTD temperature sensor is integrated with the pressure sensor to ensure real-time monitoring of the pressure sensor's operating temperature. The sensor module monitors pressure and temperature changes inside and outside the guide vane seal cavity in real time. A sampling frequency of between 1000Hz and 10,000Hz is recommended.

[0012] Pressure Calibration and Spectrum Analysis: The data processing module calibrates the pressure data collected by the sensor to correct for thermal drift errors caused by temperature changes. Spectrum analysis is used to extract the amplitude of the main frequency and the energy percentage of the high-frequency components of the pressure signal to identify abnormal periodic fluctuations.

[0013] Multi-level decision-making strategy: The anomaly detection module uses a multi-level decision-making strategy, including primary filtering, secondary verification, and final decision-making. Primary filtering determines whether there are preliminary signs of water leakage by monitoring changes in pressure differential. Secondary verification analyzes changes in high-frequency component energy. The final decision-making process combines multiple factors (such as pressure differential, frequency component, load, and speed) and uses a classification model to comprehensively score and determine whether to issue an alarm.

[0014] Model training and prediction: The model training module continuously optimizes leak detection accuracy and response speed by learning from historical data and training models. It uses a feature fusion model using LSTM, SVM, and logistic regression to automatically learn and adapt to data patterns under different working conditions, enabling intelligent leak prediction.

[0015] Early Warning and Feedback: The Early Warning and Feedback module sets different levels of alarm signals based on the model's predictions, ensuring operators can respond quickly. It also provides a user interface that displays real-time pressure data, spectrum analysis results, and early warning status, and supports historical data query capabilities.

[0016] System Integration and Maintenance: This module is responsible for effectively integrating all submodules to ensure the efficient operation of the entire turbine guide vane leakage detection system. Regular calibration and maintenance plans, including sensor calibration, system software updates, and hardware inspections, ensure the long-term stability and accuracy of the system.

[0017] The present invention provides a method and system for detecting water leakage in a turbine guide vane, which has the following beneficial effects: 1. The present invention solves the limitations of existing technologies (such as water level monitoring, flow meter detection, and manual visual inspection) in the field of guide vane leakage detection in hydropower stations, such as limited accuracy, inability to accurately determine the leakage point, and reliance on shutdown inspections, thereby achieving real-time, efficient, and accurate leakage detection.

[0018] 2. By real-time monitoring of pressure and temperature changes inside and outside the guide vane sealing cavity and using spectrum analysis and pressure calibration technology, the present invention achieves high-precision water leakage detection.

[0019] 3. The system of the present invention adopts a multi-level judgment strategy, combining the comprehensive scoring of pressure difference, high-frequency energy proportion and classification model, which effectively reduces false alarms and missed alarms and improves the reliability and adaptability of detection.

[0020] 4. Based on the feature fusion model of LSTM, SVM and logistic regression, the present invention can automatically learn and adapt to data patterns under different working conditions, realizing intelligent water leakage prediction.

[0021] 5. The system of the present invention provides a three-level early warning mechanism, which outputs different levels of alarm signals (yellow primary warning, orange intermediate warning, and red alarm signal) according to the severity of the water leakage risk, helping operators to respond quickly and effectively avoid further deterioration of the water leakage problem.

[0022] 6. The design of the system integration module of the present invention ensures efficient collaboration between the sub-modules, regular calibration and maintenance plans, and fault diagnosis functions, ensuring the long-term stability and accuracy of the system.

[0023] 7. The present invention can detect water leakage problems in a timely manner, reduce equipment downtime and maintenance costs, improve the operating efficiency and safety of the turbine, and has significant economic and social benefits.

[0024] 8. The present invention installs piezoelectric sensors at the seal adjacent positions of the upper and lower end faces of the guide vanes, and integrates RTD temperature sensors with the pressure sensors, thereby achieving real-time monitoring of pressure and temperature and improving the comprehensiveness and accuracy of monitoring.

[0025] 9. The present invention processes the pressure and temperature data collected by the sensor, uses the FFT (Fast Fourier Transform) method to perform spectrum analysis on the pressure signal, extracts the main frequency amplitude and the energy proportion of the high-frequency component, and quickly identifies abnormal periodic fluctuations.

[0026] 10. The present invention uses thermal drift error correction to correct pressure data, thereby improving the accuracy of pressure monitoring.

[0027] 11. The present invention effectively solves the key technical problems in the field of guide vane leakage detection of hydropower station turbines through innovative features such as multi-sensor fusion monitoring, spectrum analysis and pressure calibration, multi-level judgment strategy, intelligent leakage prediction model, three-level early warning mechanism, and system integration and maintenance, providing strong support for the safe and efficient operation of hydropower stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the technical process of the detection method of the present invention; Figure 2 This is a technical flow diagram of the detection method of Example 2 of the present invention; Figure 3 This is a flow chart of the LSTM-SVM-LR model based on feature fusion in Example 2 of the present invention; Figure 4 This is a module structure diagram of the detection system of Examples 3 and 4 of the present invention; Figure 5 This is a schematic diagram of the interface of intelligent auxiliary software for turbine guide vane detection based on the LSTM-SVM-LR algorithm with feature fusion according to Example 4 of the present invention; Figure 6 This is a comparison chart of the guide vane leakage prediction accuracy of different models for a power plant from January to December in Example 4 of the present invention. DETAILED DESCRIPTION

[0029] The technical solutions of the present invention are further described below with reference to the accompanying drawings and embodiments: Example 1 like Figure 1 As shown, this embodiment provides a method for detecting water leakage in a turbine guide vane, which specifically includes the following steps: Step 1: Sensor Installation and Data Collection 1.1. Sensor selection and installation location Piezoelectric sensor: Use high-sensitivity, fast-response piezoelectric sensors and install them adjacent to the upper and lower end seals of the guide vanes to ensure that pressure fluctuations caused by guide vane seal failure can be captured; Integration of pressure sensor and RTD (Resistance Temperature Detector) temperature sensor: The pressure sensor and RTD temperature sensor are integrated and installed in the same location to monitor the pressure changes inside and outside the sealed cavity and the operating temperature of the pressure sensor in real time.

[0030] 1.2 Installation steps Clean the guide vane installation area to ensure there is no oil, dirt, or impurities that may affect the sensor installation accuracy; Use special installation tools to firmly fix the piezoelectric sensor, pressure sensor and RTD temperature sensor on the guide vane, ensuring close contact between the sensor and the guide vane surface to reduce measurement errors; Connect the cables between the sensor and the data acquisition system to ensure that the connection is firm and not loose, and perform insulation tests to prevent short circuits or leakage.

[0031] 1.3 Data Collection Settings Set the sampling frequency of the data acquisition system to 5000 Hz (adjustable between 1000 Hz and 10000 Hz based on actual needs) to ensure that subtle changes in the pressure signal can be captured; Configure the data acquisition system so that it can receive and store the pressure and temperature data collected by the sensor in real time.

[0032] Step 2: Pressure Calibration 2.1 Initial calibration After the sensor is installed, perform initial calibration and record the initial temperature Pressure value under ; At set temperature Under this condition, the pressure value of the pressure sensor is calculated according to formula (1) : (1) Where, It is the rate of change of pressure due to temperature change, determined by experimental calibration; for example, recording the readings of the pressure sensor at different ambient temperatures and using linear regression to determine the temperature compensation coefficient The specific value range of the pressure is determined, and fine-tuned according to the actual working conditions of the hydropower station to ensure the accuracy of the pressure calibration; usually in the actual operating environment of the hydropower station, combined with historical data, through regression analysis or machine learning algorithms (such as linear regression) to determine The specific value of, for example, within a specific temperature range (such as 20°C to 40°C), is obtained by calibration with a large amount of historical data. The typical value is 0.002MPa / °C (this value is only an example and needs to be determined through experiments).

[0033] 2.2 Thermal Drift Error Correction According to the formula, the corrected pressure is obtained Expressed as: (2).

[0034] Step 3: Spectrum Analysis 3.1 FFT Transform Perform FFT (fast Fourier transform) on the collected discrete pressure signal to obtain the frequency domain representation of the pressure signal; Use professional signal processing software or programming languages ​​(such as MATLAB, Python, etc.) to implement FFT transformation.

[0035] 3.2 Extraction of main frequency amplitude and high-frequency component energy ratio Calculate the amplitude spectrum of the FFT result to obtain the amplitude of different frequency components; Determine the main frequency amplitude, that is, the frequency component corresponding to the maximum value in the amplitude spectrum; Calculate the energy ratio of high-frequency components, that is, the ratio of the energy of high-frequency components (usually components with frequencies higher than a set threshold) to the total energy.

[0036] Step 4: Calculate pressure difference and rate of change 4.1 Pressure difference calculation According to formula (3), calculate the pressure difference between the outside and inside of the sealed cavity : (3) Where, 、 Respectively represent the pressure inside and outside the sealing cavity; 4.2 Calculation of pressure change rate According to the instantaneous change rate of the sealed chamber pressure, calculate the change rate of pressure over time: (4) (5) Where, 、 Respectively represent the pressure inside and outside the sealing chamber The instantaneous rate of change at a moment.

[0037] In practical applications, numerical differentiation methods (such as the central difference method) can be used to approximate the pressure change rate.

[0038] Step 5: Multi-level decision strategy 5.1 Primary filtration By comparing the pressure difference The dynamic threshold is set to determine whether there are preliminary signs of water leakage. The specific empirical formula is as follows: (6) Where, is the turbine load; is the turbine speed; 、 、 Represent different empirical coefficients, obtained through historical data or regression fitting Dynamic threshold setting method: Determine the empirical coefficient based on historical data or regression fitting 、 、 , so that the threshold can be dynamically adjusted as the turbine load and speed change.

[0039] 5.2 Secondary Verification Analyze the changes in high-frequency component energy. If the high-frequency component energy ratio exceeds the set benchmark value, further verify the possibility of water leakage. The benchmark value formula for the high-frequency component energy ratio is as follows: (7) Where, Indicates the instantaneous rate of change of the seal chamber pressure; It is the benchmark value of high-frequency energy proportion; Historical benchmark values ​​set for experience; 、 、 、 、 、 、 They represent different empirical coefficients, obtained by fitting historical data; is the natural base of the exponential function; The benchmark value of the high-frequency component energy ratio can be determined based on historical data or experiments.

[0040] 5.3 Final Decision Data sequence construction: The pressure inside and outside the sealed cavity, the fluctuation frequency, the high-frequency component energy ratio and the instantaneous change rate of the pressure inside and outside the sealed cavity together constitute the turbine guide vane pressure data sequence; the turbine load and speed As auxiliary input data sequence, the two data sequences together constitute the model input first-level parameter sequence : (8) Where, 、 Respectively represent the fluctuation frequency of the pressure inside and outside the sealing chamber; 、 They represent the proportion of high-frequency component energy of the internal and external pressures of the sealed cavity respectively; 、 denote the load and speed of the turbine at time t respectively; Sliding window processing: Use sliding window technology to construct time series samples, each sample contains the past time step data (such as = 10), and obtain the model input secondary parameter sequence : (9) Where, 、 、 The classification representation model is 、 and Input; Indicates time that has passed; Model input and prediction: The secondary parameter sequence is input into the trained LSTM (Long Short-Term Memory) model to extract temporal information; the SVM (Support Vector Machine) model is combined to extract static information; and finally, logistic regression is used to integrate temporal features and classification probabilities to generate the final probability of water leakage prediction. .

[0041] Alarm decision: If the final probability If the value exceeds the set threshold (such as 0.8), the water leakage detection result is output and a red alarm signal is triggered; if it is between 0.5 and 0.8, an orange intermediate alarm is triggered; if it is lower than 0.5, no alarm is triggered.

[0042] This first embodiment describes in detail the various steps of a method for detecting water leakage in turbine guide vanes, including sensor installation and data acquisition, pressure calibration, spectrum analysis, pressure difference and rate of change calculation, and a multi-stage determination strategy. By refining the specific operation methods and calculation formulas for each step, the method's operability and accuracy are improved.

[0043] Example 2 In another preferred embodiment, based on the above embodiment 1, Figure 2 As shown, this embodiment provides a method for detecting water leakage of a turbine guide vane, the method comprising the following steps: Piezoelectric sensors are installed at the seal adjacent to the upper and lower end faces of the guide vane, and RTD temperature sensors are integrated with the pressure sensor to ensure that the working temperature of the pressure sensor can be monitored in real time.

[0044] S1: Pressure calibration, correcting the thermal drift error of pressure data; the output signal of the pressure sensor will change at different temperatures. and temperature There are certain dependencies, which can be expressed as: (1) Where, At the initial temperature The pressure value under is the rate of change of pressure due to temperature change, determined by experimental calibration; The corrected pressure can be expressed as: (2) The specific value of is determined by experimental calibration; According to the corrected pressure, the pressure inside and outside the sealing cavity are obtained respectively 、 .

[0045] S2: Extract the main frequency amplitude of the pressure signal using spectrum analysis method and the energy proportion of high-frequency components ,Abnormal periodic fluctuations may be caused by water flow disturbance caused by leakage; The sampling frequency of the pressure sensor outside the sealed chamber is (number of samples per second), the number of sampling points is N , at a certain moment The collected pressure data constitutes a discrete signal ,in , perform FFT transformation on the discrete pressure signal and convert it into a frequency domain signal: (10) Where, is the frequency component The complex value at , representing the frequency the amplitude and phase of is the imaginary unit in complex numbers, are discrete points; The distance between each frequency component for: (11) Calculate the magnitude spectrum of the FFT result (Take the absolute value) to get the amplitude of different frequency components, and find the frequency component with the largest amplitude in the amplitude spectrum , which is the main frequency amplitude of the pressure fluctuation in the upper part of the sealing chamber at this moment: (12) The total energy of the pressure signal outside the sealed chamber is the sum of the powers of all frequency components of the signal. , in the frequency domain, the energy is obtained by square of the amplitude of the frequency component: (13) The frequency range of the high frequency part is from arrive , the energy of the high-frequency component of the pressure signal outside the sealed chamber Proportion for: (14) Similarly, the main frequency amplitude of the pressure fluctuation in the lower part of the sealed chamber can be obtained and ; In particular, the sampling frequency It should be high enough to ensure no distortion, and the preferred value is 1000~10000HZ; Record the main frequency range of pressure fluctuations under normal operating conditions (such as historical data or laboratory calibration) and establish a benchmark database as a reference for judging abnormalities.

[0046] S3: Calculate the pressure difference between the inside and outside of the sealed chamber : (3) During normal operation, the pressure difference between the inside and outside of the sealing chamber should be within a specific range. When the pressure difference decreases significantly or fluctuates abnormally, it indicates that the seal has failed and there is water leakage.

[0047] S4: Calculate the pressure change rate over time based on the instantaneous change rate of the sealed chamber pressure : (4) (5) Where, 、 Respectively represent the pressure inside and outside the sealing chamber The instantaneous rate of change at a moment.

[0048] S5: A multi-level judgment strategy is adopted to determine whether the sealing cavity is leaking. The LSTM-SVM-LR model based on feature fusion is adopted. The specific process is as follows Figure 3 As shown: S51, primary filtration (pressure difference below dynamic threshold) The change in pressure difference is used to determine whether there are preliminary signs of water leakage. If the pressure difference is lower than the set dynamic threshold, it is preliminarily judged as abnormal and a warning signal is triggered; Data input: According to the pressure difference between the inside and outside of the sealed cavity provided by the sensor ; Dynamic Threshold The following empirical formula can be used to calculate: (6) Where, is the turbine load; is the turbine speed; 、 、 Represent different empirical coefficients, obtained through historical data or regression fitting like , then the warning is triggered, a yellow primary warning signal is output, and the secondary verification stage is entered.

[0049] S52, Secondary Verification (High-frequency component energy ratio exceeds the reference value) Changes in high-frequency component energy are associated with anomalies (such as leaks, vibrations, etc.) Secondary verification further verifies the presence of water leaks by analyzing the signal in the frequency domain.

[0050] Based on multiple factors such as turbine operating status, load, speed, pressure change rate, etc., the benchmark value of the high-frequency component energy ratio can be obtained The empirical formula is (7) Where, Indicates the instantaneous rate of change of the seal chamber pressure; Historical benchmark values ​​set for experience; 、 、 、 、 、 、 They represent different empirical coefficients, obtained by fitting historical data; Is the natural base of the exponential function; empirical value As a reference benchmark, the calculation of high-frequency energy is corrected to make it consistent with the historical operating characteristics of the equipment; the additional sine wave term takes into account the impact of periodic vibration on the proportion of high-frequency energy; Fit the reference values ​​of the high-frequency energy ratio of the pressure signals inside and outside the sealed cavity respectively. If or If there are signs of water leakage, an orange intermediate warning signal will be output, and the final decision-making stage will be entered; 、 、 、 They are the high-frequency energy proportion and baseline value of the pressure signals inside and outside the sealed cavity, respectively.

[0051] S52. Final decision (comprehensive score of classification model) By combining multiple factors (such as pressure difference, frequency component, load, speed, etc.), the classification model is used to make the final decision. The probability score output by the model determines whether to alarm; the pressure inside and outside the sealing chamber 、 , Fluctuation frequency , high-frequency component energy ratio And the instantaneous rate of change of pressure inside and outside the sealed chamber and Together they constitute the turbine guide vane pressure data sequence, turbine load T and speed ω As auxiliary input data sequences, the two data sequences together constitute the model input first-level parameter sequence: (8) The training set is divided into training set, validation set and test set in a ratio of 70%-15%-15%. The training set (70%) is used for parameter learning of the LSTM model, optimizing weights and fitting data patterns. The validation set (15%) is used for hyperparameter adjustment, such as learning rate and number of hidden layer units, to prevent overfitting. The test set (15%) is used to finally evaluate the generalization ability of the model to ensure that it performs well on unseen data. The data is constructed using the sliding window technology to construct time series samples, each sample contains the past Time step data is collected to ensure that the model learns short-term and long-term temporal dependencies and obtains the model input secondary parameter sequence : (9) Will Input into the LSTM model to extract timing information.

[0052] The gate mechanism of LSTM cells is enhanced: The LSTM cell consists of a forget gate, an input gate, an output gate, and a cell state update. This embodiment adjusts the gate mechanism based on characteristics such as the pressure fluctuation rate to make it more suitable for sealing leakage prediction.

[0053] (1) Forget Gate The forget gate controls the LSTM cell state in time Should it be forgotten? When the sealed chamber pressure changes dramatically or the high-frequency energy increases abnormally, the model can reduce its reliance on past states and pay more attention to current inputs. Modify the calculation method of the forget gate: (15) Where, Represents the output of the forget gate (between 0 and 1); Represents the weight matrix of the forget gate; Represents the bias of the forget gate; Indicates the hidden state at the previous moment; Represents the input of the current time step; Represents the Sigmoid activation function; The proportion of high-frequency energy introduced and pressure fluctuation rate , so that the forget gate can dynamically adjust the degree of forgetting according to the current state of the system.

[0054] (2) Input gate The input gate controls the current input Effects on cell status when loading When the change is large, the system state is different, so the input gate can depend on To dynamically adjust: (16) Introducing load As additional input, adjust the input gate's acceptance of new information; when When the change is small, it means that the turbine is in a stable operating state. Moderate, update information normally, when If the change is large, it may indicate abnormal conditions. Adjustments may be needed to prevent excessive model updates; At the same time, the LSTM cell needs to calculate new candidate states: (17) Where, Represents the weight of the input gate; represents the candidate cell state; 、 Represents the weight matrix before and after input gate adjustment; 、 represents the bias of the input gate; Represents the hyperbolic tangent activation function, ensuring that the output is between [-1, 1]; High-frequency peak As input, it helps the model focus on high-frequency abnormal signals of the system; Final state of LSTM cell for: (18) Where, express The cell state at a moment in time; Represents the convolution operation.

[0055] (3) Output gate Output of the current LSTM cell: (19) Where, 、 Represents the weight matrix and bias of the output gate; is the weight of the output gate; is the final LSTM hidden state output (i.e., LSTM prediction result); Speed Affects the weight of the output gate, reduced at low speed , to avoid false alarms, increase the , enhance the detection of abnormal fluctuations; LSTM generates time series prediction probabilities for: (20) 1) LSTM training: Use binary cross entropy to optimize the loss function: (twenty one) Using two-category labels, This means that there is abnormal leakage in the guide vane seal. It means the guide vane seal is normal. is the predicted probability of the LSTM model; Using the Adam (Adaptive Moment Estimation) optimizer: (twenty two) Where, It is the set of model parameters, including the weight matrix and bias term of the LSTM network; is the learning rate, 、 They are the first-order momentum estimate and the second-order momentum estimate of the gradient respectively; is a small constant that prevents division by zero; in each iteration, The prediction error (loss function) of the model will be gradually reduced according to the gradient update, thereby improving the accuracy of turbine guide vane leakage detection.

[0056] 2) LSTM model connection layer design: After LSTM processes the time series, the hidden state of the output It will be used as the final time series feature for water leakage prediction; the connection layer is responsible for fusing the LSTM output and the SVM prediction result; the final hidden state predicted by LSTM And the SVM model output results Combine to form a new feature vector : (twenty three) Where, Contains LSTM time series features, representing time series information; SVM output , providing prior information for static classification prediction; the proportion of high-frequency energy and pressure change rate As supplementary information, it improves the generalization ability of the model.

[0057] 3) SVM model feature weighting: Input the data of the first-level parameter sequence into the SVM model to extract static information based on the current data; Weighted, the probability of LSTM output As a weight factor, adjust the influence of the input feature; adjust the weight of the input feature according to the probability, so as to obtain the weighted feature vector : (twenty four) Each input feature Corresponding to a label ; This means that there is abnormal leakage in the guide vane seal. This indicates that the guide vane seal is normal; the decision function of SVM is: (25) Where, is the weight vector of the hyperplane, is the bias term; The decision value output by SVM is converted into a probability value through the Sigmoid activation function: (26) Where, Indicates the probability of water leakage in the sample under the SVM model; The cross entropy loss function is used to minimize the training error of SVM, optimize the hyperplane parameters of SVM, and use the SMO (Sequential Minimal Optimization) optimization algorithm to solve.

[0058] This example uses logistic regression to integrate the time series features of LSTM and SVM classification probability , build an LSTM-SVM-LR model based on feature fusion to obtain the final output leakage prediction probability .

[0059] First calculate the weighted sum of the input features : (27) Where, is the transposed weight vector of logistic regression, ; is the bias term; Output after linear transformation (without activation function); Specifically expanded as follows: (28) Use Sigmoid activation function to convert linear output Mapped to [0, 1], the final leakage probability is obtained: (29) The training goal of logistic regression is to minimize the binary cross entropy loss: (30) Using two-category labels, This means that there is abnormal leakage in the guide vane seal. It means that the guide vane seal is normal; is the predicted probability of the logistic regression model; The final classification decision result is (31) Where, is the classification threshold (default 0.5, can be adjusted to optimize recall or precision); if , a red alarm signal is output, prompting staff to immediately check whether there is a risk of water leakage.

[0060] Example 3 In another preferred embodiment, based on the above embodiments 1 and 2, as Figure 4 As shown, this embodiment provides a turbine guide vane leakage detection system, which is a detection system for a turbine guide vane leakage detection method described in embodiments 1 and 2, and includes the following modules: 1. Sensor module 1.1 Sensor selection and configuration Piezoelectric sensor: Use industrial-grade piezoelectric sensors with high sensitivity and anti-interference capabilities. The range covers the possible pressure fluctuation range of the turbine guide vane seal cavity, and the response time is less than 1ms, ensuring rapid capture of small pressure changes. Pressure sensor: High-precision, high-stability pressure sensor with temperature compensation function, measurement accuracy better than ±0.1%FS, long-term stability better than ±0.05%FS / year; RTD temperature sensor: PT100 or PT1000 platinum resistance temperature sensor is selected, with a measurement range of -50°C to +200°C and an accuracy of ±0.1°C, ensuring accurate monitoring of temperature changes; Integrated installation design: Integrates the pressure sensor and RTD temperature sensor in the same sealed housing to reduce wiring complexity and ensure the synchronization of pressure and temperature data through calibration.

[0061] 1.2 Installation and Calibration Perform sensor zero point calibration and sensitivity test before installation to ensure the accuracy of the initial state; Regularly perform zero-point and full-scale calibration on the sensor to eliminate zero-point drift and span error.

[0062] 1.3 Data Collection and Transmission Configure a high sampling rate data acquisition card (such as NI-9234) to support multi-channel synchronous acquisition, and the sampling frequency can be adjusted to 10kHz to meet the needs of high-frequency signal capture; The sensor signal is transmitted to the data acquisition module through a shielded cable to reduce electromagnetic interference.

[0063] 2. Data processing module 2.1 Data Preprocessing After collecting sensor data in real time, digital filtering (such as Butterworth filter) is performed to remove high-frequency noise and retain valid signals; Perform sliding average filtering on the pressure data (window size 5 to 10 sampling points) to suppress random interference; Outliers (such as sudden pressure changes exceeding 3 times the standard deviation) were detected and corrected using linear interpolation.

[0064] 2.2 Spectrum Analysis Implementation Use Python's NumPy and SciPy libraries to implement FFT transformation, set the Hanning window function to reduce spectrum leakage, and set the spectrum resolution to Δf=fs / N (fs is the sampling rate and N is the number of FFT points); Extract the main frequency amplitude (maximum value of the amplitude spectrum) and the energy ratio of the high-frequency component (the ratio of the energy with frequency > fc to the total energy, where fc is the cutoff frequency, determined based on experiments).

[0065] 2.3 Calculation of pressure difference and rate of change Real-time calculation of pressure difference , a sliding average filter (window length 10 seconds) is used to reduce noise interference; Calculating the rate of pressure change , a Savitzky-Golay filter (third-order polynomial, window length 5 seconds) is used to smooth the derivative calculation to avoid numerical oscillations.

[0066] 3. Anomaly Detection Module 3.1 Implementation of multi-level decision strategy Primary filtering: dynamic threshold Through machine learning models (such as random forests) combined with dynamic adjustment of turbine operating conditions (load, speed), historical data training optimizes threshold curves; Secondary verification: high-frequency energy ratio In the calculation, a dynamic update mechanism of historical benchmark values ​​is added, and the benchmark values ​​are recalibrated every quarter. .

[0067] 3.2 Final decision logic LSTM model input: The first-level parameter sequence was normalized (Z-score), and a three-layer LSTM network (64 hidden layer neurons) was used with a dropout rate of 0.2 to prevent overfitting.

[0068] Fusion of SVM and logistic regression: The SVM uses the RBF (Radial Basis Function Kernel) kernel function with C=1.0 and γ=0.1. The logistic regression weight is the weighted average of the LSTM output probability and the SVM classification probability (weight 0.6:0.4).

[0069] 4. Model training module 4.1 Dataset Division Collect historical data and divide it into training set, validation set, and test set according to the ratio of 7:1.5:1.5.

[0070] 4.2 Sliding Window Construction The window size is set to 20 time steps, the step size is 1, and 1000 samples are generated (assuming a total duration of 1000 seconds and a sampling rate of 1 Hz to simplify the scenario).

[0071] 4.3 Model Fusion LSTM extracts temporal features, SVM classifies static features, and logistic regression integrates them to generate the final probability.

[0072] 5. Early warning and feedback module Alarm level setting Yellow alert: ∈[0.3,0.5) Orange warning: ∈[0.5,0.8) Red Alarm: ≥0.8 6. System integration and maintenance Regularly calibrate sensors and update model parameters to ensure efficient system operation.

[0073] This embodiment describes in detail the composition and functions of a turbine guide vane leakage detection system, including a sensor module, a data processing module, an anomaly detection module, a model training module, a warning and feedback module, and a system integration and maintenance module. By refining the specific implementation methods and precautions of each module, the operability and accuracy of the system are improved, providing a complete solution for turbine guide vane leakage detection.

[0074] Example 4 In another preferred embodiment, based on the above embodiment 3, as Figure 4 As shown, this embodiment provides a turbine guide vane leakage detection system, which is specifically as follows: A turbine guide vane leakage detection system includes the following modules: a sensor module, a data processing module, an anomaly detection module, a model training module, an early warning and feedback module, and a system integration and maintenance module. The sensor module is responsible for real-time monitoring of pressure and temperature changes inside and outside the guide vane seal cavity. This module primarily includes a piezoelectric sensor, a pressure sensor, and an RTD temperature sensor. First, piezoelectric sensors are installed at the guide vane root and at the upper and lower portions of the seal cavity. These sensors rapidly respond to pressure changes and convert them into electrical signals, enabling real-time monitoring of pressure inside and outside the seal cavity. Second, the pressure sensor is integrated with the piezoelectric sensor to record pressure data inside and outside the seal cavity, which is then regularly collected by a data acquisition system. To ensure the accuracy of pressure data, an RTD temperature sensor is integrated into the system to monitor operating temperature. Because the output signal of the pressure sensor can be affected by temperature fluctuations, real-time temperature monitoring is crucial for subsequent data processing. The entire sensor module should be designed to ensure a sufficiently high sampling frequency, preferably between 1000Hz and 10000Hz, to capture subtle changes in the pressure signal and avoid data distortion caused by a too low sampling frequency. In addition, the sensor module should have good anti-interference capabilities and be able to operate stably under complex working conditions to ensure the reliability of the system.

[0075] The data processing module is responsible for analyzing and processing the pressure and temperature data collected by the sensors to extract valuable information and determine whether a water leak is present. First, the data processing module performs pressure calibration on the sensor output signal to correct for thermal drift errors caused by temperature fluctuations. The data processing module then uses spectral analysis to perform FFT (Fast Fourier Transform) processing on the collected pressure signal, converting the time-domain signal into the frequency domain. This process extracts the amplitude of the dominant frequency and the energy proportion of the high-frequency components. By analyzing the amplitude spectrum, it identifies abnormal periodic fluctuations, which may be caused by water flow disturbances caused by a water leak. The data processing module also calculates the pressure differential inside and outside the sealed chamber in real time and compares it with the pressure differential range during normal operation. If the pressure differential decreases significantly or exhibits abnormal fluctuations, the system flags the condition as abnormal, indicating a potential water leak risk. Through this series of data processing steps, the module effectively provides accurate baseline data for subsequent anomaly detection.

[0076] The anomaly detection module's primary function is to determine whether the turbine guide vanes are leaking, based on the pressure and spectrum information provided by the data processing module. This module employs a multi-level decision strategy, first performing primary filtering to identify preliminary signs of leakage. The system sets a dynamic threshold to monitor changes in pressure differential. When the pressure differential falls below the set dynamic threshold, the system triggers a primary warning signal, alerting the operator to a possible anomaly. If the initial determination is an anomaly, the system enters a secondary verification phase. During this phase, the module analyzes changes in high-frequency component energy. If the proportion of high-frequency component energy exceeds a baseline value, a further warning is issued. This baseline value is derived by fitting historical data to ensure it aligns with the equipment's operating characteristics. If an abnormal high-frequency component energy is detected, the module issues an orange intermediate warning signal, indicating the need for further attention to the seal chamber pressure changes. The final decision phase combines multiple factors (such as pressure differential, frequency component, load, and speed) and uses a classification model to comprehensively score and determine whether immediate personnel deployment is necessary. This module's multi-level decision-making mechanism aims to improve the accuracy and reliability of leak detection.

[0077] The model training module is the intelligent core of the entire system. By learning from historical data and training the model, it continuously optimizes the accuracy and response speed of water leak detection. This module first organizes the collected data and divides the dataset into training, validation, and test sets, with a 70%-15%-15% split. The training set is used to learn the parameters of the LSTM (Long Short-Term Memory) model, optimize weights, and fit the data pattern. The validation set is used to adjust hyperparameters such as the learning rate and the number of hidden units to prevent overfitting. The test set is used to ultimately evaluate the model's generalization ability and ensure good performance on unseen data. During model training, a sliding window technique is used to construct time series samples, each containing data from multiple time steps in the past, to ensure that the model can learn both short-term and long-term temporal dependencies. The LSTM model, through its forget gate, input gate, and output gate mechanisms, effectively captures important features in the time series. This enhances the model's predictive capabilities, particularly when influenced by features such as the rate of change of pressure fluctuations. Simultaneously, the SVM (Support Vector Machine) model is trained during this phase to extract static information based on the current data. Ultimately, the model training module combines the outputs of LSTM and SVM, integrating time series features and classification probabilities through logistic regression to generate the final probability of leak prediction. This module design ensures that the system can adapt to the operational requirements of different working conditions and improves the intelligent level of detection.

[0078] The early warning and feedback module provides operators with real-time monitoring information and early warning signals so that they can take timely measures to deal with potential water leakage risks. This module sets different levels of alarm signals based on the prediction results of the model to ensure that operators can respond quickly. When the system determines that there is a risk of water leakage, it will output alarm signals of different colors to facilitate operators to judge and take corresponding measures, as shown in Table 1: The yellow primary warning indicates that the pressure difference is lower than the dynamic threshold, prompting the operator to pay attention; the orange intermediate alarm indicates that the energy ratio of the high-frequency component is abnormal, prompting further attention; the red alarm signal indicates that the final prediction is that there is a risk of water leakage, and it is recommended to immediately dispatch staff for on-site inspection. In order to enhance the practicality of the system, the early warning and feedback module should also have a user interface, such as Figure 5 As shown, the module displays real-time pressure data, spectrum analysis results, and warning status, helping operators quickly obtain key information. In addition, the module supports historical data query, allowing operators to analyze past operating status and warning records to further optimize maintenance strategies.

[0079] Table 1 Three-level early warning mechanism

[0080] The system integration and maintenance module is responsible for effectively integrating the various sub-modules to ensure the efficient operation of the entire turbine guide vane leakage detection system. This module integrates the system's hardware and software to ensure smooth data transmission and coordination between modules such as sensors, data processing, anomaly detection, model training, and early warning feedback. At the same time, system integration also needs to take into account the interface design between different modules to ensure the consistency and compatibility of data formats. To ensure the long-term stability and accuracy of the system, the maintenance module regularly formulates calibration and maintenance plans, including sensor calibration, system software updates, and hardware inspections. In addition, the maintenance module also has a fault diagnosis function that can monitor the system's operating status in real time, promptly detect and eliminate faults, and ensure the reliability and stability of the system.

[0081] In this example, 2000 historical operating data points were collected from a certain model of hydroturbine unit at a power plant from January to December of the past year. The specific model parameters are shown in Table 2. The model was trained using these 2000 historical operating data points. The batch size was set to 64, the learning rate was 0.001, the number of hidden units was 200, and the training rounds were 100. The mean squared error (MSE) loss function was used to ensure that the model could effectively capture the complex nonlinear relationship between input features and the output leakage probability value.

[0082] Table 2 Turbine parameter settings

[0083] Compared with other models, the accuracy of the probability prediction data of turbine leakage between January and December is as follows: Figure 6 shown.

[0084] In the preferred solution, in the spectrum analysis step of Step 3, the discrete pressure signal is subjected to FFT transformation, the amplitude spectrum of the FFT result is calculated, the amplitudes of different frequency components are obtained, and the main frequency amplitude and the energy proportion of the high-frequency component are determined; the above settings are intended to accurately identify signal characteristics, determine the main vibration source by the main frequency amplitude, and evaluate the signal noise level and potential fault trend by the energy proportion of the high-frequency component, providing key data support for subsequent fault diagnosis.

[0085] In the preferred solution, the classification model adopts a feature fusion model of LSTM, SVM and logistic regression to automatically learn and adapt to data patterns under different working conditions to achieve intelligent water leakage prediction; the above settings significantly improve the accuracy and efficiency of water leakage detection; at the same time, the solution integrates real-time data monitoring and abnormal alarm functions. Once signs of water leakage are detected, the early warning mechanism is immediately triggered to ensure rapid response and processing, effectively avoiding water resource waste and potential losses.

[0086] In the preferred solution, the sampling frequency of the sensor module is between 1000Hz and 10000Hz to ensure that subtle changes in the pressure signal are captured; the above settings can greatly improve the accuracy and real-time performance of data acquisition, laying a solid foundation for subsequent signal processing and analysis; at the same time, the solution also adopts advanced filtering algorithms to effectively reduce noise interference and improve the overall performance of the system.

[0087] In the preferred solution, the data processing module uses a spectrum analysis method to perform FFT processing on the collected pressure signal, extracts the main frequency amplitude and the energy proportion of the high-frequency component, and calculates the pressure difference inside and outside the sealing cavity in real time; the above settings effectively improve the accuracy of equipment fault warning. Through the analysis of the main frequency amplitude change and the energy proportion of the high-frequency component, it can timely detect signs of sealing performance degradation, providing reliable data support for equipment maintenance.

[0088] In the preferred solution, the model training module divides the collected data set into a training set, a validation set, and a test set, uses sliding window technology to construct time series samples, trains an LSTM model to capture important features in the time series, and combines it with an SVM model to extract static information. Finally, through logistic regression, the time series features and classification probabilities are integrated to generate the final probability of water leakage prediction. The above settings greatly improve the accuracy of water leakage detection. In practical applications, this solution can effectively identify water leakage anomalies, reduce false alarm rates, and provide strong technical support for the maintenance and management of urban water supply systems.

[0089] In the preferred solution, the early warning and feedback module sets yellow primary warning, orange intermediate warning and red alarm signals according to the prediction results of the model, displays pressure data, spectrum analysis results and early warning status in real time, and supports historical data query function; the above settings ensure that managers can quickly identify potential equipment failures or overload conditions and take maintenance measures in time; at the same time, the system also supports data export function to facilitate further analysis and optimization of equipment operation efficiency.

[0090] In summary, the present invention proposes a method and system for detecting water leakage in guide vanes of turbines. The core purpose of this solution is to solve the specific technical problems existing in the field of water leakage detection in guide vanes of turbine units in hydropower stations. Currently, water leakage detection in guide vanes mainly relies on methods such as water level monitoring, flow meter detection and manual visual inspection, but these methods have exposed many limitations in practical applications. Water level monitoring has limited accuracy in the location and amount of leakage, making it difficult to capture early signs of leakage. Although flow meter detection can reflect the total flow change in the flow channel, it cannot accurately locate the leakage point. Manual visual inspection relies on the rotation detection of the main shaft after shutdown, which is not only complicated to operate, but can only indirectly judge the severity of leakage and cannot detect leakage trends in a timely manner. These limitations make it difficult for existing technologies to meet the needs of modern hydropower stations for safe and efficient operation.

[0091] To overcome these technical challenges, the present invention innovatively installs piezoelectric sensors adjacent to the seals on the upper and lower ends of the guide vanes, integrating an RTD temperature sensor with the pressure sensor to achieve real-time monitoring of pressure and temperature. This multi-sensor fusion approach, uncommon in existing technologies, significantly improves the comprehensiveness and accuracy of monitoring, providing a solid data foundation for subsequent leak detection.

[0092] In terms of signal processing, the present invention uses the FFT (Fast Fourier Transform) method to perform spectral analysis on the pressure signal, extracting the main frequency amplitude and the energy proportion of the high-frequency component to identify abnormal periodic fluctuations; this frequency domain analysis method is innovative in the field of water leakage detection and can effectively capture abnormal signals caused by water flow disturbances caused by water leakage; at the same time, by correcting the thermal drift error of the pressure data, the accuracy of pressure monitoring is further improved, ensuring the reliability of the detection results.

[0093] To improve the accuracy and reliability of leak detection, the system of this invention employs a multi-stage decision-making strategy: this strategy comprises three stages: primary filtering, secondary verification, and final decision-making. This strategy effectively reduces false positives and missed negatives by combining pressure differential, high-frequency energy content, and a comprehensive scoring system based on a classification model. In the primary filtering stage, the system monitors changes in pressure differential to make a preliminary assessment of leak detection. In the secondary verification stage, the system further analyzes changes in high-frequency energy content to verify the results of the primary filtering. In the final decision-making stage, the system integrates multiple factors and uses a classification model to make a final decision, ensuring the accuracy and reliability of the detection results.

[0094] In addition, the present invention also constructs a feature fusion model based on LSTM (long short-term memory network), SVM (support vector machine) and logistic regression. This model can automatically learn and adapt to data patterns under different working conditions to achieve intelligent water leakage prediction; by combining the time series feature extraction capability of LSTM and the static classification capability of SVM, as well as the integration effect of logistic regression, the system can more accurately predict water leakage risks and provide operators with timely early warning information.

[0095] To ensure operators can quickly respond to leak risks, the system also provides a three-level warning mechanism. Depending on the severity of the leak, the system outputs different levels of alarm signals, including a yellow primary warning, an orange intermediate warning, and a red alarm. This hierarchical alarm mechanism is innovative in leak detection systems and can effectively improve operators' response speed and coping capabilities.

[0096] In terms of system integration, this invention ensures efficient collaboration between submodules through system integration modules. Regular calibration and maintenance plans, along with the introduction of fault diagnosis capabilities, ensure long-term stable system operation while facilitating upgrades and maintenance. This modular and integrated design not only improves system reliability and stability but also reduces operational and maintenance costs.

[0097] By implementing this solution, the present invention can promptly detect water leakage problems, reduce equipment downtime and maintenance costs, and improve the operating efficiency and safety of turbines. Its significant economic and social benefits provide strong support for the safe and efficient operation of hydropower stations. In summary, through innovative features such as multi-sensor fusion monitoring, spectrum analysis and pressure calibration, multi-level judgment strategy, intelligent water leakage prediction model, three-level early warning mechanism, dynamic threshold adjustment, system integration and modular design, this invention achieves real-time, efficient, and accurate water leakage detection, providing a strong guarantee for the safe and efficient operation of hydropower stations.

Claims

1. A method for detecting water leakage of a turbine guide vane, characterized in that: The following steps are involved: Step 1: Install piezoelectric sensors on the upper and lower end seals of the guide vanes, and integrate RTD temperature sensors with the pressure sensors to monitor the operating temperature of the pressure sensors and the pressure changes inside and outside the seal cavity in real time. Step 2: Perform pressure calibration to correct the thermal drift error of the pressure data and obtain the corrected pressure data inside and outside the sealed cavity; Step 3: Use spectrum analysis method to extract the main frequency amplitude and high-frequency component energy ratio of the pressure signal; Step 4: Calculate the pressure difference between the inside and outside of the sealed cavity, and calculate the rate of change of pressure over time based on the instantaneous rate of change of the sealed cavity pressure; Step 5: A multi-level judgment strategy is adopted to determine whether the sealed cavity is leaking, including primary filtration, secondary verification and final decision-making. The comprehensive score of the pressure difference, high-frequency energy ratio and classification model is combined to output the leakage detection result.

2. A method for detecting water leakage of a turbine guide vane according to claim 1, characterized in that: In the Step 2 pressure calibration step, the corrected pressure Expressed as: （1）； （2）； Where, At the initial temperature The pressure value under At the set temperature The pressure value under is the rate of change of pressure due to temperature change, determined by experimental calibration.

3. A method for detecting water leakage of a turbine guide vane according to claim 2, characterized in that: In the spectrum analysis step of Step 3, the discrete pressure signal is subjected to FFT transformation, the amplitude spectrum of the FFT result is calculated, the amplitudes of different frequency components are obtained, and the main frequency amplitude and the energy proportion of the high-frequency components are determined.

4. A method for detecting water leakage of a turbine guide vane according to claim 3, characterized in that: The specific calculation steps for Step 4 to calculate the pressure difference between the inside and outside of the sealed cavity and the rate of change of pressure over time are as follows: Step 4.1: Calculate the pressure difference between the inside and outside of the sealed chamber : （3）； Where, 、 Respectively represent the pressure inside and outside the sealing cavity; Step 4.2: Calculate the rate of change of pressure over time based on the instantaneous rate of change of the sealed chamber pressure: （4）； （5）； Where, 、 Respectively represent the pressure inside and outside the sealing chamber The instantaneous rate of change at a moment.

5. A method for detecting water leakage of a turbine guide vane according to claim 4, characterized in that: The Step 5 multi-level determination strategy includes: Step 5.1: Primary filtration: By comparing the pressure difference with the set dynamic threshold, determine whether there are preliminary signs of water leakage; Step 5.2: Secondary verification: further verify whether there is a water leak by analyzing the changes in the high-frequency component energy; Step 5.3: Final decision is made by combining multiple factors and using a classification model to comprehensively score and decide whether to issue an alarm.

6. A method for detecting water leakage of a turbine guide vane according to claim 5, characterized in that: Dynamic threshold in the primary filtering step of Step 5.1 The formula is as follows: （6）； Where, is the turbine load; is the turbine speed; 、 、 They represent different empirical coefficients, which are obtained through historical data or regression fitting.

7. A method for detecting water leakage of a turbine guide vane according to claim 6, characterized in that: The formula for the benchmark value of the high-frequency component energy ratio in the Step 5.2 secondary verification step is as follows: （7）； Where, Indicates the instantaneous rate of change of the seal chamber pressure; It is the benchmark value of high-frequency energy proportion; Historical benchmark values ​​set for experience; 、 、 、 、 、 、 They represent different empirical coefficients, obtained by fitting historical data; is the natural base of the exponential function.

8. A method for detecting water leakage of a turbine guide vane according to claim 7, characterized in that: The final decision in Step 5.3 includes the following steps: Step 5.3.1: The pressure inside and outside the sealed cavity, the fluctuation frequency, the high-frequency component energy ratio and the instantaneous change rate of the pressure inside and outside the sealed cavity together constitute the turbine guide vane pressure data sequence, turbine load and speed As auxiliary input data sequence, the two data sequences together constitute the model input first-level parameter sequence : （8）； Where, 、 Respectively represent the fluctuation frequency of the pressure inside and outside the sealing chamber; 、 They represent the proportion of high-frequency component energy of the internal and external pressures of the sealed cavity respectively; 、 denote the load and speed of the turbine at time t respectively; Step 5.3.2: Use the sliding window technique to construct time series samples from the first-level parameter sequence data. Each sample contains the past Time step data is collected to ensure that the model learns short-term and long-term temporal dependencies and obtains the model input secondary parameter sequence : （9）； Where, 、 、 The classification representation model is 、 and Input; represents the time step that has passed; Step 5.3.3: Input into the LSTM model to extract timing information.

9. A method for detecting water leakage of a turbine guide vane according to claim 8, characterized in that: The classification model adopts a feature fusion model of LSTM, SVM and logistic regression to automatically learn and adapt to data patterns under different working conditions to achieve intelligent water leakage prediction.

10. A water turbine guide vane leakage detection system, used to implement the water turbine guide vane leakage detection method according to claim 9, characterized in that: include: The sensor module is responsible for real-time monitoring of pressure and temperature changes inside and outside the guide vane sealing cavity, including piezoelectric sensors, pressure sensors, and RTD temperature sensors; The data processing module is responsible for analyzing and processing the pressure and temperature data collected by the sensor, extracting valuable information and determining whether there is a water leak; The anomaly detection module determines whether there is water leakage in the turbine guide vanes based on the pressure and spectrum information provided by the data processing module, using a multi-level judgment strategy; The model training module continuously optimizes the accuracy and response speed of water leak detection by learning historical data and training models; The early warning and feedback module provides operators with real-time monitoring information and early warning signals, and sets alarm signals of different levels; The system integration and maintenance module is responsible for effectively integrating the various sub-modules to ensure the efficient operation of the entire turbine guide vane leakage detection system and perform regular calibration and maintenance.

11. A turbine guide vane leakage detection system according to claim 10, characterized in that: The sampling frequency of the sensor module is between 1000Hz and 10000Hz to ensure that subtle changes in the pressure signal are captured; the data processing module uses a spectrum analysis method to perform FFT processing on the collected pressure signal, extract the main frequency amplitude and the energy proportion of the high-frequency component, and calculate the pressure difference inside and outside the sealed cavity in real time.

12. The water turbine guide vane leakage detection system according to claim 10, characterized in that: The model training module divides the collected data set into a training set, a validation set, and a test set, uses sliding window technology to construct time series samples, trains an LSTM model to capture important features in the time series, and combines it with the SVM model to extract static information. Finally, it integrates time series features and classification probabilities through logistic regression to generate the final probability of water leakage prediction. The early warning and feedback module sets yellow primary warning, orange intermediate warning, and red alarm signals based on the model's prediction results, displays pressure data, spectrum analysis results, and early warning status in real time, and supports historical data query functions.

Images