A water pump water turbine leakage fault intelligent diagnosis method based on multi-dimensional rotor dynamic characteristic fusion

Abstract

This invention proposes an intelligent diagnostic method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion, belonging to the field of hydraulic machinery condition monitoring and fault diagnosis technology. The method includes: S1: System calibration, collecting vibration displacement data under dry conditions; S2: Collecting vibration displacement data and pressure pulsation data under wet conditions; calculating pure fluid excitation displacement; S3: Multi-dimensional extraction of multi-source signals within the current time window under wet conditions; S4: Using a random forest algorithm to filter the parameters in S3, selecting parameters with importance greater than a set threshold; S5: Establishing samples using the filtered parameters; S6: Training a neural network based on the samples, obtaining a trained neural network, and using the trained neural network to continue identifying turbine faults and outputting the results. This method can filter technical features from multiple dimensions and use them as samples for the neural network, improving the accuracy of water turbine leakage fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic machinery condition monitoring and fault diagnosis technology, specifically involving an intelligent diagnosis method for water pump turbine leakage faults based on the fusion of multi-dimensional rotor dynamic characteristics. Background Technology

[0002] Leaks in water pumps and turbines can directly threaten the safe and stable operation of the unit, causing equipment damage or even safety accidents.

[0003] Existing leak detection technologies (such as top cover water level sensor monitoring, acoustic / ultrasonic detection, guide vane leakage calculation models, etc.) all have inherent limitations: they are either passive response modes (requiring the alarm to be triggered after leakage accumulates), can only cover specific local areas, or heavily rely on shutdown conditions, and cannot achieve global, early proactive detection and accurate location of leakage risks when the unit is running.

[0004] Moreover, the above methods fail to take into account the fundamental physical mechanism of fluid-fluid interaction (FSI) and ignore the core key point—water leakage as the excitation source of asymmetric flow field. Its most direct and sensitive global response is concentrated in the changes of the dynamic behavior (vibration, deflection, and running trajectory) of the rotor system.

[0005] Therefore, this paper proposes an intelligent diagnostic method for water pump turbine leakage faults based on the fusion of multi-dimensional rotor dynamic characteristics. This method relies on the fluid-structure interaction mechanism and realizes leakage fault diagnosis by decoding the rotor dynamic response, which has both important engineering application value and theoretical innovation significance. Summary of the Invention

[0006] This invention proposes an intelligent diagnostic method for water pump turbine leakage faults based on the fusion of multi-dimensional rotor dynamic characteristics, which identifies water turbine leakage faults based on multi-dimensional rotor dynamic characteristics.

[0007] To achieve the above objectives, the present invention proposes the following technical content:

[0008] A method for intelligent diagnosis of water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion includes the following steps:

[0009] S1: System calibration, collecting vibration displacement data under dry operating conditions;

[0010] S2: Collect vibration displacement data and pressure pulsation data under wet conditions, ensuring that the window length is the same for dry and wet conditions; calculate the pure fluid excitation displacement;

[0011] S3: Perform multi-dimensional extraction of the pure fluid excitation signal within the current time window under wet operating conditions; including peak values. skewness Root mean square value , frequency amplitude of blade passage Baseline amplitude Amplitude ratio Spectrum centroid Harmonic distortion Phase difference Trajectory drift Pressure-vibration correlation coefficient Reference fluid dynamics Dimensionless fluid forces Dimensionless coupling stiffness Coupling stiffness coefficient in dynamic coupling characteristics ;

[0012] S4: Use the random forest algorithm to filter the parameters in S3, and select the parameters whose importance is greater than the set threshold;

[0013] S5: Use the filtered parameters to build a sample; the sample includes parameter values, fault type, fault location, severity, and confidence level;

[0014] S6: Train the neural network based on the samples to obtain the trained neural network. Use the trained neural network to continue to identify the faults of the water turbine and output the results.

[0015] Furthermore, in step S2, the displacement is purely fluid-induced. The formula is:

[0016]

[0017] In the formula, This is a wet mode signal; It is a dry modal signal; This is a purely fluid excitation signal;

[0018] In step S3, the peak value The calculation formula is:

[0019]

[0020] In the formula, and These represent the maximum and minimum values ​​of the pure fluid excitation displacement within the current time window, respectively.

[0021] Root mean square value The calculation formula is:

[0022]

[0023] In the formula, N is the number of sampling points within the current time window; The displacement is the pure fluid excitation displacement at the i-th sampling point;

[0024] Skewness The calculation formula is:

[0025]

[0026] In the formula, This is the average value of the fluid excitation displacement at all sampling points within the current window; This represents the standard deviation of the fluid excitation displacement at all sampling points within the current window.

[0027] peak skewness Root mean square value Using time-domain characteristics, we analyze turbine leakage faults from a time-domain perspective.

[0028] Furthermore, in step S3, the pure fluid excitation displacement sequence is... Perform a Fast Fourier Transform to generate a real-time spectrogram.

[0029] Amplitude ratio The calculation formula is:

[0030]

[0031] In the formula, Indicated in the real-time spectrum graph In the figure, the amplitude at the blade passing frequency corresponding to the rated speed; express After fast Fourier transform, the amplitude at the blade passing frequency corresponding to the rated speed is extracted;

[0032] Spectral centroid The calculation formula is:

[0033]

[0034] In the formula, Represents real-time spectrum Frequency values ​​in; Real-time spectrum diagram medium frequency The mold at the location; Maximum analysis frequency;

[0035] Harmonic distortion The calculation formula is:

[0036]

[0037] In the formula, Real-time spectrum diagram The frequency corresponds to the fundamental frequency of rotor rotation. Representing the spectrum The amplitude of vibration at frequencies that are integer multiples of the fundamental frequency (n ≥ 2) is the amplitude ratio. Spectrum centroid Harmonic distortion The frequency domain analysis of water turbine leakage faults.

[0038] Furthermore, in step S3,

[0039] Phase difference The calculation formula is:

[0040]

[0041] In the formula, This indicates the time lag of the maximum values ​​of the X and Y displacement sensor signals under wet conditions. Indicates the real-time rotation frequency. This indicates the real-time rotational speed obtained from the key phase signal;

[0042] Trajectory drift The calculation formula is:

[0043]

[0044] In the formula, Indicates the real-time center coordinates of the rotor; Indicates the center coordinates of the rotor under no-load conditions; phase difference and trajectory drift The water turbine leakage fault was analyzed from the trajectory characteristics.

[0045] Furthermore, in step S3,

[0046] Pressure-vibration correlation coefficient The calculation formula is:

[0047]

[0048] In the formula, Describing covariance, This represents the standard deviation of the pressure signal. Indicates the standard deviation of the displacement signal; The time-series data representing the pressure signal; the pressure-vibration correlation coefficient R, used to analyze turbine leakage faults from the perspective of fluid-structure interaction.

[0049] Furthermore, in step S3,

[0050] Reference fluid dynamics The formula is:

[0051]

[0052] In the formula, The density of water; This refers to the real-time impeller outlet circumferential speed within the current time window.

[0053] Dimensionless fluid forces The formula is:

[0054]

[0055] In the formula, It is the fluid excitation force; For reference hydrodynamics;

[0056] Dimensionless coupling stiffness The formula is:

[0057]

[0058] In the formula, It is a dimensionless hydrodynamic peak value. This represents the dimensionless peak value of the vibration displacement.

[0059] Coupling stiffness coefficient in dynamic coupling characteristics The formula is:

[0060]

[0061] In the formula, The peak value of the hydrodynamic force with dimensions. The peak displacement of the dimensional fluid vibration; reference hydrodynamics. Dimensionless fluid forces and dimensionless coupling stiffness The coupling stiffness coefficient in the dynamic coupling characteristics is used to analyze the water turbine leakage fault from the perspective of coupling correlation characteristics.

[0062] The beneficial effects that can be achieved by adopting the above technologies are:

[0063] This study analyzes turbine failures from multiple dimensions and uses a random forest algorithm to select highly important features. While the selected highly important features differ for each turbine type, the chosen parameters include peak value. Root mean square value skewness The ratio of the frequency amplitude of the blade Spectral centroid in frequency domain characteristics Pressure-vibration correlation coefficient Reference fluid dynamics Dimensionless fluid forces Dimensionless coupling stiffness Coupling stiffness coefficient in dynamic coupling characteristics These are all highly important parameters, and using them to predict turbine leakage faults has high accuracy and reliability. Attached Figure Description

[0064] Figure 1 This is the flowchart of this solution;

[0065] Figure 2 These are diagrams of dry modal signals, wet modal reference signals, and pure fluid excitation signals;

[0066] Figure 3 It is a fault spectrum diagram of displacement due to pure fluid excitation;

[0067] Figure 4 It is a rating chart of each parameter. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] like Figure 1 As shown, an intelligent diagnostic method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion includes the following steps:

[0070] S1: System calibration. This includes the following steps:

[0071] S1.1: Calibration of displacement and pressure sensors.

[0072] Calibration of the displacement sensor: First, select a standard test block made of the same material as the water pump turbine rotor and place the test block within the measurement range of the displacement sensor; then, move the test block continuously at 0.1mm intervals and record the output voltage of the displacement sensor. The distance between the actual displacement of the test block and the actual displacement of the test block. Establish a linear regression model: ,in This is the sensitivity coefficient. Set the zero drift value; then set more than ten calibration points to ensure the goodness of fit of the aforementioned linear regression model. The absolute value of the residual is ≤0.005mm; finally, repeat the calibration three times, and take the average value of each parameter three times as the final calibration parameter, and enter it into the system database to facilitate the subsequent use of the calibrated displacement sensor to collect data.

[0073] Pressure sensor calibration: First, connect the high-precision pressure calibrator to the pressure sensor; set 7 calibration points on the high-precision pressure calibrator in 0~3 bar, 0.5 bar intervals, and monitor them with the pressure sensor; after the signal at each calibration point stabilizes for 30 seconds, record the output value of the pressure sensor; calculate the linearity error, and when the linearity error... And return error ( When the pressure sensor is at full scale (meaning full scale), it meets the calibration requirements. If it does not meet either of the above two requirements, the pressure sensor parameters should be replaced and recalibrated. The calibrated pressure sensor parameters are entered into the system database for subsequent data acquisition using the calibrated pressure sensor.

[0074] S1.2: System parameter configuration.

[0075] Install each sensor in its corresponding position on the water pump and turbine according to the standard. The relevant installation positions are as follows:

[0076] Pressure sensor: installed at the inlet of the volute or the sealing area of ​​the top cover of the water pump turbine.

[0077] Two orthogonally arranged displacement sensors (denoted as X and Y displacement sensors respectively): installed in the X and Y directions of the rotor journal respectively;

[0078] Key phase sensor: mounted on the side of the spindle.

[0079] Manually rotate the turbine rotor to the angular position of the turbine rotor. At this point, the key phase sensor is triggered, and the output values ​​of the X and Y displacement sensors, calibrated in step S1.1, are monitored respectively. If the deviation of a displacement sensor from the theoretical zero point is greater than 0.02mm, the installation angle of the X and Y displacement sensors is adjusted. After the installation angle is adjusted, the rotor is rotated again to... Trigger the key phase sensor and repeat the above operation until the deviations of the X and Y displacement sensors from the theoretical zero point are both less than or equal to 0.02 mm. The theoretical zero point deviation means the deviation between the displacement values ​​output by the X and Y displacement sensors and the corresponding preset X and Y displacement values ​​in the system.

[0080] S1.3: Obtain dry mode signals .

[0081] First, drain the water pump and turbine flow channels to put the unit into a waterless idling state. Set the speed to... Traffic volume And continue to run This solution defines the waterless operation state as a dry condition. Under this condition, the raw vibration displacement signals of the X and Y displacement sensors are acquired using a NIPXIe-6363 high-speed data acquisition card (2048 sampling points for each sensor), and these signals are labeled as dry modal signals. And enter it into the system database; Vibration displacement signal output from X-displacement sensor and the vibration displacement signal output by the Y displacement sensor. .

[0082] S2: Multi-source dynamic signal acquisition and preprocessing based on wet operating conditions. Specifically, it includes the following steps:

[0083] S2.1: Establish wet operating conditions and perform synchronous acquisition of multi-source signals.

[0084] First, the pump-turbine flow channel is filled with water, and the unit is controlled to enter the actual load-bearing operation state. This scheme defines this water-filled operation state as "wet condition". Under this condition, the vibration displacement signals of the X and Y displacement sensors are collected using the NIPXIe-6363 high-speed data acquisition card in the same way as above (2048 sampling points for each). The vibration displacement signal collected at this time is the sum of pure fluid excitation and mechanical vibration, and it is marked as wet modal signal. .

[0085] Vibration displacement signal output from X-displacement sensor and the vibration displacement signal output by the Y displacement sensor. In addition, the NIPXIe-6363 high-speed data acquisition card also acquires pressure pulsation data from the pressure sensor.

[0086] S2.2: The Butterworth filter and DrY-Wet correction method are used to process the multi-source signals in the current time window.

[0087] The multi-source signals within the current time window undergo signal cleaning and physical field separation (DrY-Wet correction). First, a phase-shift-free Butterworth low-pass filter (cutoff frequency 500Hz) is used to denoise the dry and wet mode signals acquired in step S2.1. Then, to eliminate vibration interference caused by purely mechanical factors such as rotor mass imbalance and shaft misalignment, and to accurately extract the vibration component excited by the leaking water fluid, a "dry-wet condition differential (DrY-Wet Correction)" calculation is performed. Taking X-axis vibration displacement as an example, the specific calculation formula is as follows:

[0088]

[0089] in, Represented as the wet modal signal in the X direction; Represented as X-direction dry mode signal; This is represented as a pure fluid excitation signal. The signal results for the three are shown below. Figure 2 .

[0090] S3: Extract the pure fluid excitation signal within the current time window in multiple dimensions.

[0091] In this embodiment, the multi-dimensional approach encompasses four dimensions: time domain, frequency domain, trajectory, and fluid-structure interaction. The core characteristic variables involved in this step are as follows:

[0092] (1) Time-domain characteristics, including peak value ( ), skewness ( ) and root mean square value ( Among them, the peak value ( ) is used to characterize the vibration amplitude of the rotor; root mean square value ( () is used to characterize the magnitude of rotor vibration energy.

[0093] (2) Frequency domain characteristics, including the amplitude of the frequency at which the blade passes ( ), baseline amplitude ( ), amplitude ratio ( Harmonic distortion ( ) and spectral centroid ( ).

[0094] (3) Trajectory characteristics, including phase difference ( ) and trajectory drift ( ).

[0095] (4) Fluid-structure interaction characteristics, including dynamic pressure ( Pressure-vibration correlation coefficient ( ) and dimensionless fluid forces ( ).

[0096] Specifically, the following steps are included:

[0097] S3.1: Extract the time-domain features of the pure fluid excitation signal in the current time window in S2.

[0098] Time-domain characteristics include: peak value ( ), root mean square value ( ) and skewness ( ).

[0099] Peak ( This reflects the vibration amplitude of the rotor. If there is water leakage, the flow field will be asymmetrical, leading to an increase in excitation force and a significant increase in the peak value. Taking the X direction as an example again:

[0100]

[0101] In the formula, and These represent the maximum and minimum values ​​of the pure fluid excitation displacement within the current time window, respectively. The time window is the time series composed of the aforementioned 2048 sampling points. This represents the peak displacement of pure fluid excitation within the current time window. See [link to pure fluid excitation displacement]. Figure 3 .

[0102] Root mean square value ( This reflects the magnitude of rotor vibration energy; the periodic excitation caused by water leakage will... It grows linearly. Let's take the X direction as an example again:

[0103]

[0104] In the formula, N is the number of sampling points in the current time window, i.e., 2048; The displacement is the pure fluid excitation displacement at the i-th sampling point in the current time window.

[0105] Skewness ( This describes the symmetry of the vibration signal distribution; if water leaks on one side, it will cause the signal to be skewed.

[0106]

[0107] In the formula, This is the mean of the pure fluid excitation displacements of all sampling points within the current window; This represents the standard deviation of the pure fluid excitation displacement of all sampling points within the current window. This represents the pure fluid excitation displacement at the i-th sampling point within the current time window. Indicates the skewness of the current time window.

[0108] S3.2: Extract the frequency domain features of the pure fluid excitation signal in the current window.

[0109] Specifically, for the pure fluid excitation displacement sequence Perform a Fast Fourier Transform (FFT) to generate a real-time spectrogram. ;

[0110] The ratio of the frequency amplitude of the blades ( This characterizes the characteristic frequency of rotor fluid excitation; leakage will increase the amplitude of this frequency. The formula is:

[0111]

[0112] In the formula, To in real-time spectrum In the figure, the amplitude at the blade passing frequency corresponding to the rated speed is expressed in mm. For the results obtained in step S1.3 After performing the FFT transformation, the amplitude at the blade passing frequency corresponding to the rated speed is extracted, in mm.

[0113] Spectral centroid ( This reflects the location of concentrated spectral energy. Water leakage will introduce high-frequency disturbances, causing the value to shift towards higher frequencies. The formula is:

[0114]

[0115] In the formula, Real-time spectrum diagram The frequency value in the image, in Hz. Maximum analysis frequency, in Hz; Real-time spectrum diagram medium frequency The mold at the location; Represents real-time spectrum The spectral centroid.

[0116] Harmonic distortion ( This measures the harmonic content of the signal. If the flow field is distorted, additional harmonics will be generated, leading to... Increase; the formula is:

[0117]

[0118] In the formula, For the spectrum The value corresponds to the vibration amplitude at the rotor's fundamental frequency (1 × rotational frequency), in mm. Representing the spectrum The vibration amplitude corresponds to an integer multiple of the fundamental frequency (i.e., n × rotational frequency, such as 2 ×, 3 ×, 4 ×...), where n ≥ 2 and is an integer. In this embodiment, it is set as follows: ; Representing the spectrum Harmonic distortion in the medium.

[0119] S3.3: Extract trajectory features for pure fluid excitation signals in the X and Y directions.

[0120] Trajectory characteristics include: phase difference ( ) and trajectory drift ( );

[0121] Phase difference ( This characterizes the spatiotemporal relationship of vibration; if leakage occurs, it will change the fluid-structure interaction phase. The deviation will be from 90° to 60° or 120°, which is determined by calculating the phase difference of the X and Y displacement sensor signals and the peak position of the cross-correlation function;

[0122]

[0123] In the formula, The time lag of the pure fluid excitation signals in the X and Y directions within the same time window represents the maximum value of the signals in the same time window. Indicates the real-time rotation frequency. , This indicates the real-time rotational speed obtained from the key phase signal.

[0124] Trajectory drift ( This reflects the average position offset of the rotor. If water leakage occurs, the lateral force generated by the leaking seal will cause the trajectory center to deviate from the origin.

[0125]

[0126] In the formula, Indicates the real-time center coordinates of the rotor, in mm. This indicates the center coordinates of the rotor under no-load conditions, in mm, and is derived from the average coordinate values ​​recorded in the no-load baseline data.

[0127] S3.4: For the pure fluid excitation signal in the current window, extract the fluid excitation features; mainly the pressure-vibration correlation coefficient ( ).

[0128] This parameter verifies the fluid-structure interaction relationship. If leakage occurs, the resulting flow field disturbance will cause a strong correlation between pressure and vibration signals. If the value is greater than or equal to a set threshold (0.6 in this embodiment), the excitation source can be determined to be fluid disturbance. Defined as:

[0129]

[0130] In the formula, This represents the covariance of the signal within the same time window; This indicates the standard deviation of the pressure signal, in bar. This represents the standard deviation of the displacement signal from pure fluid excitation in the X direction; Time-series data representing pressure signals; This represents the pressure-vibration correlation coefficient.

[0131] S3.5: Extract coupling correlation features, mainly including static coupling features and dynamic coupling features. Static coupling features specifically include: reference hydrodynamics (…). ) and dimensionless fluid forces ( The dynamic coupling characteristics include dimensionless coupling stiffness (), while dynamic coupling characteristics include dimensionless coupling stiffness (). ) and coupling stiffness coefficient ( ).

[0132] in,

[0133]

[0134] In the formula, This refers to the density of water, expressed in kg / m³. This represents the real-time impeller outlet circumferential velocity within the current time window. , This refers to the real-time speed of the generator unit. The blade height is in meters (m). The impeller outlet radius is in meters (m).

[0135] Dimensionless hydrodynamics in static coupling characteristics It directly reflects the degree of flow field distortion, and is specifically defined as:

[0136]

[0137] In the formula, The fluid excitation force, in units of N, is the resultant force in the X and Y directions, specifically calculated from the vibration signal inversion. For reference fluid dynamics, the unit is (N).

[0138] Dimensionless coupling stiffness in dynamic coupling characteristics This reflects the relative stiffness characteristics of the fluid support system. If leakage occurs, the flow field support capacity will weaken, leading to… It decreases as the leakage increases, and its specific definition is:

[0139]

[0140] In the formula, The peak value of the dimensionless hydrodynamic force, from the dimensionless hydrodynamic force Obtained in; The dimensionless peak value of the vibration displacement is derived from the fluid vibration displacement. get;

[0141] The coupling stiffness coefficient in dynamic coupling characteristics ( The equivalent stiffness of the fluid-structure interaction system is quantified. If leakage occurs, the flow field support stiffness will be disrupted, and its value will decrease linearly with the increase of leakage. Specifically, it is defined as:

[0142]

[0143] In the formula, The value represents the dimensional peak fluid force, in N. The value represents the peak displacement of the fluid vibration with dimensions, expressed in meters (m).

[0144] S3.6: The above multi-dimensional features are filtered using the random forest algorithm to retain important and effective features. The random forest algorithm is an existing technology.

[0145] Specifically, the processing steps of the random forest algorithm mainly include:

[0146] 1. Feature mapping and input association: The features to be evaluated in the above formula In physical terms, this specifically refers to the feature set calculated in steps S3.1 to S3.5. This set specifically includes time-domain features, frequency-domain features, axisymmetric trajectory features, and fluid-structure interaction features. All of these features together constitute a high-dimensional candidate feature vector, which serves as the input variable for the random forest algorithm.

[0147] 2. Screening and Processing Method (Feature Importance Evaluation): The algorithm traverses all decision trees in the random forest (the number of trees is...). ), quantify each feature The contribution of this feature in distinguishing between "leakage faults" and "normal operating conditions". The specific processing logic is as follows: if a certain feature is used... (e.g., pressure-vibration correlation coefficient) After splitting the node, the impurity of the left and right child nodes can be reduced ( and ) relative to the impurity of the parent node ( If the value of a feature significantly decreases (meaning it can clearly distinguish between "leaking" and "non-leaking" data), then the value of the numerator in the formula increases, and the importance score of that feature increases. The score increases accordingly; conversely, if a certain characteristic (such as some higher-order harmonic distortion) cannot effectively reduce impurity before and after splitting, its score will decrease. Approaching 0;

[0148] 3. Final screening results: The system summarizes the importance scores of all features. And set a filtering threshold (in this embodiment, the threshold is set to 0.05). The specific execution logic is: retain Highly sensitive features and elimination The algorithm removes redundant or noisy features. Finally, it outputs a dimensionality-reduced optimized feature vector that contains only the key indicators most sensitive to the fault. This vector is directly used as the input to the AI ​​diagnostic model in step S4, thereby significantly improving computational efficiency while ensuring diagnostic accuracy.

[0149] S4: AI diagnostic model training and inference. This includes the following steps:

[0150] S4.1: Construct an intelligent diagnostic sample library based on screening features.

[0151] The "samples" constructed in this step are not the original sensor data collected in step S2, but rather the high-importance feature vectors retained after being filtered by the random forest in step S3.6.

[0152] Data Acquisition: Three different models of generator sets were selected, covering various operating conditions including no-load, partial load, and full load. For each generator set, four typical leakage scenarios were artificially simulated (including top cover seal failure, main shaft seal failure, guide vane end face seal failure, and combined failure), and the evolution of different degrees of failure was simulated by adjusting the test bench parameters. Finally, a large-scale vibration and pressure dataset containing normal conditions and various failure states was collected and constructed.

[0153] Feature generation and filtering: For each set of raw data collected above, steps S2 (signal preprocessing), S3.1~S3.5 (full feature calculation), and S3.6 (feature filtering) are executed sequentially. During this process, redundant features with an importance score below 0.05 are removed, and only high-importance features that are sensitive to faults are retained.

[0154] Vector Construction and Storage: Arrange the key features selected and retained in step S3.6 in a set order, and perform normalization processing to form feature vectors. (i.e., the input layer data of the model); at the same time, based on the actual operating status of the unit and the severity of water leakage, fault level labels are defined. (like: This indicates that it is normal. This represents a level one leak. This indicates a level 2 leak. (Representing a level three leak). Ultimately, all of them... The data pairs are stored in the database to complete the construction of the intelligent diagnostic sample library.

[0155] S4.2: The sample set was divided into a training set (16800 groups), a validation set (4800 groups), and a test set (2400 groups) in a 7:2:1 ratio, and stratified sampling was used to ensure a consistent proportion of fault types. The model optimizer used was AdamW with an initial learning rate of 0.001 and a weight decay of 0.001. Furthermore, a cosine annealing strategy was employed, with a maximum number of training iterations. Set to 100 rounds, with a minimum learning rate of 0.00001; the multi-task loss function is defined as follows:

[0156] ,in, Using "first-level label (fault presence or absence) and second-level label (fault location)" as true labels, the supervised model learns to classify fault types. Using "three-level labeling (fault level labeling)" ")" is used as the true value to supervise the quantitative fitting of the leakage level by the monitoring model; in this embodiment, , This is to achieve synchronous convergence of the model for classification and regression tasks.

[0157] During training, a phased fine-tuning mechanism is adopted: the parameters of the feature extraction layers (CNN and RNN) are frozen in the first 50 rounds, and only the back-end classification and regression layers are trained; then, in rounds 51 to 200, all parameters are unfrozen and fine-tuned together, and the model is validated every 10 rounds; the model is saved when the accuracy on the validation set is not less than 95%; finally, an early stopping strategy is adopted, that is, training is stopped when the loss does not decrease after 15 consecutive rounds of validation.

[0158] In the online diagnosis process after the calculation is completed, firstly, a set of 28-dimensional feature vectors is extracted every 0.1 seconds to form a feature stream with a 50ms interval; then, TensorRT is used to accelerate inference and call the pre-trained model weight file; finally, the fault type, location, severity and confidence level are output (a confidence level ≥ 0.8 is set as a valid result).

[0159] To verify the reliability of the AI ​​model results, the system will automatically calculate the pressure-vibration cross-correlation coefficient when the AI ​​model outputs fault results. and the ratio of the frequency amplitude of the blade If satisfied and If it is confirmed as a fault, it is confirmed as a fault; otherwise, it is marked as a "suspected fault" and monitored continuously.

[0160] S5: Output of diagnostic results.

[0161] After inputting the sample to be judged into the model constructed by S4, the system outputs visualized diagnostic results. The display interface specifically includes: axisymmetric trajectory diagrams in the X / Y directions, time-domain waveform diagrams, and spectrum diagrams; it also intuitively displays the fault type, confidence score, key feature outliers, and fault occurrence time. Furthermore, it can overlay and display real-time trajectories with historical baseline trajectories, and real-time spectra with baseline spectra to highlight fault evolution characteristics. Fault severity (corresponding label) The alarm levels are divided into three levels: Level 1 alarm (alert / attention), Level 2 alarm (handling / early warning), and Level 3 alarm (shutdown / danger).

[0162] The automatically generated diagnostic report consists of three parts: first, basic information, covering the unit number, diagnostic period, and real-time operating parameters; second, in-depth fault analysis, including feature evolution trend charts, quantitative indicators of abnormal features, and inferred fault mechanisms; and third, maintenance decision recommendations, such as recommended replacement seal models, standard maintenance procedures, and estimated man-hours. Through the combination of this visual interface and the detailed report, maintenance personnel can comprehensively understand the leakage status of the pumps and turbines and take timely and targeted measures.

[0163] In summary, the method and system proposed in this application have the following beneficial effects:

[0164] 1. High-precision diagnosis: It integrates the local feature extraction capability of CNN with the temporal modeling advantage of RNN to achieve accurate identification of complex water leakage conditions in water pumps and turbines;

[0165] 2. Real-time performance and lightweight design: Through random forest feature selection and TensorRT model quantization acceleration technology, computational redundancy is greatly reduced, meeting the real-time monitoring needs of industrial sites.

[0166] 3. Explainability and Applicability: It provides a complete closed loop from feature anomalies to maintenance suggestions, solving the problem of traditional AI models being "black boxes" and difficult to use, and has significant engineering application value.

[0167] Example:

[0168] To demonstrate the calculation logic and physical meaning of each key parameter in this scheme, the following detailed calculations are performed using measured data from a reversible unit of a pumped storage power station.

[0169] The preset operating conditions of this unit include: unit speed 1500r / min, frequency 25Hz, number of impeller blades 7, blade passing frequency 175Hz, sampling rate 4096Hz, and time window length 2048 points.

[0170] Data processing mainly includes the following parts:

[0171] First, perform dry-wet difference calculation and The acquisition of, specifically includes:

[0172] For dry mode signals ( In obtaining the value of the displacement voltage in the X direction at a certain moment obtained in step S1.3, it is 1.200V (corresponding to a physical displacement of 0.05mm).

[0173] For wet mode real-time signals ( In obtaining the value of the displacement voltage in the X direction at the same phase moment in step S2.1, the value is 2.640V (corresponding to a physical displacement of 0.110 mm).

[0174] Using the difference formula in step S2.2: The displacement component caused by pure fluid excitation at that moment can be calculated: By performing the above calculations on each of the 2048 sampling points within the time window, a complete fluid excitation displacement sequence can be obtained. .

[0175] Secondly, the specific calculation of multi-dimensional feature parameters includes:

[0176] For peak values ​​in time-domain features ( The result can be obtained through calculation using the formula:

[0177] Unit: mm;

[0178] For the root mean square value in the time domain features ( The result can be obtained through calculation using the formula:

[0179] Unit: mm;

[0180] For the skewness in time-domain features ( The formula can be used to calculate:

[0181] ;

[0182] For the frequency domain characteristics of the blade passing frequency amplitude ratio ( The formula can be used to calculate:

[0183] Unit: mm, where To perform the following steps under the dry operating conditions in step S1.3: Perform an FFT transform, read the amplitude at 175Hz (the blade's passing frequency), and measure... Unit: mm; In step S3.2, for Perform an FFT transform, read the amplitude at 175Hz (the blade's passing frequency), and measure... Unit: mm;

[0184] For the spectral centroid in the frequency domain characteristics ( The formula can be used to calculate:

[0185] ,in , ; , ; , ;

[0186] For the pressure-vibration correlation coefficient in fluid-structure interaction characteristics ( The formula can be used to calculate:

[0187] The standard deviation of pressure Statistical calculations yielded a value of 0.5 (in bar); standard deviation of displacement. Statistical calculations yielded a value of 0.02 (unit: mm). The covariance of the pressure and displacement sequences within the current time window, in units. ; , dimensionless.

[0188] For the phase difference in trajectory features ( The result can be obtained through calculation using the formula.

[0189] ;

[0190] For trajectory drift in trajectory features ( The formula can be used to calculate:

[0191] ,in For step S1.3, measure the coordinates of the unloaded axle center. ; For step S3.3, measure the current... mean of the sequence ;

[0192] For dimensionless fluid forces in static coupling characteristics ( The result can be obtained through calculation using the formula:

[0193] ;

[0194] For the dimensionless coupling stiffness in dynamic coupling characteristics ( The result can be obtained through calculation using the formula:

[0195] ;

[0196] For the coupling stiffness coefficient in dynamic coupling characteristics ( The result can be obtained through calculation using the formula:

[0197] .

[0198] Secondly, based on the top cover sealing leakage condition data in the aforementioned embodiments, the above feature parameters are screened using the formula in the random forest algorithm. The specific process includes:

[0199] First, the system gathers all the feature parameters calculated in steps S3.1 to S3.5 and constructs a multidimensional original feature vector under the current time window (t);

[0200] 2. Input the above multidimensional feature vectors into the pre-trained random forest model. This algorithm traverses each tree and calculates the contribution of each feature value in the binary classification of "leaking samples" and "normal samples".

[0201] Specifically, this process includes:

[0202] If the algorithm detects that a certain feature parameter splits at a set value, since the current input value is much larger than that set value, this feature can clearly classify more than 95% of faulty samples into the same leaf node, thus significantly reducing the Gini impurity of the node. This can then be achieved through the formula:

[0203] The feature parameter is assigned a relatively large value, such as 0.45; T represents the total number of decision trees in the random forest; t represents the current t-th decision tree; Indicates the Gini impurity of the parent node; This represents the Gini impurity of the left child node after splitting; This represents the Gini impurity of the right child node after splitting; This represents the total number of samples in the training set; This represents the number of samples with feature z at node v in the t-th tree.

[0204] If the algorithm detects that a certain feature parameter splits at a set value, and the current input value is extremely close to the standard value of a normal distribution, then under both the "normal" and "early leakage" conditions, this value has a very high degree of overlap. Regardless of the splitting threshold, the sample classes of the left and right child nodes remain mixed, and the Gini impurity hardly decreases. This can then be addressed using the formula:

[0205] The feature parameter is assigned a small value, such as 0.01.

[0206] III. The system makes automated screening decisions based on thresholds; feature screening thresholds are set for the system. The system will compare the calculated score Z with the threshold. Each comparison is performed, and a binary judgment is applied. For example, regarding the ratio of the frequency amplitude of the blade passage... Through comparison, Greater than (0.05) can make the blade pass through the frequency amplitude ratio This feature is identified as highly sensitive and is retained.

[0207] IV. After the system has undergone variable processing and filtering using the random forest algorithm described above, it can optimize the feature vectors and output them. For the content in this embodiment, the main features retained and output include: the ratio of blade passage frequency amplitude (...). Pressure-vibration correlation coefficient ( ), trajectory drift ( ), spectral centroid ( ), dimensionless fluid mechanics ( ), phase difference ( Harmonic distortion ( ) and peak ( The eight-dimensional preferred feature vectors for these feature parameters. Other feature parameters were discarded because their importance scores in the unit's operating conditions were below the threshold, and were therefore judged as redundant or noise. Importance scores are as follows: Figure 4 .

[0208] Secondly, to achieve intelligent diagnosis under physical constraints, a dual verification mechanism of "AI inference + physical verification" will be adopted. Specifically, this includes: firstly, inputting the aforementioned eight optimized feature vectors into a pre-trained CNN-RNN deep learning model. The model outputs a prediction result: the fault type is "top cover seal leakage," with a probability of 96% and a severity index of 0.65; secondly, the system enforces physical criterion checks. For example, regarding the blade passage frequency amplitude ratio (K) and the pressure-vibration correlation coefficient (R), if... and If the double verification passes, the diagnosis is confirmed to be effective, and the possibility of sensor false alarms is ruled out.

[0209] Finally, based on the above diagnostic results, the system performs the following operations: First, an alarm is triggered. Since the severity index falls within the range of [0.4, 0.7), a level two alarm is triggered (requiring maintenance). Second, a diagnostic report containing the "physical traceability evidence chain" is generated, clearly demonstrating... (Frequency domain anomaly) and (Mechanism correlation) provides maintenance personnel with an interpretable basis for maintenance.

[0210] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for intelligent diagnosis of water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion, characterized in that, Includes the following steps: S1: System calibration, collecting vibration displacement data under dry operating conditions; S2: Collect vibration displacement data and pressure pulsation data under wet conditions, ensuring that the window length is the same for dry and wet conditions; calculate the pure fluid excitation displacement; S3: Extract the pure fluid excitation signal within the current time window of the wet working condition from multiple dimensions; Including peak skewness Root mean square value , frequency amplitude of blade passage Baseline amplitude Amplitude ratio Spectrum centroid Harmonic distortion Phase difference Trajectory drift Pressure-vibration correlation coefficient Reference fluid dynamics Dimensionless fluid forces Dimensionless coupling stiffness Coupling stiffness coefficient in dynamic coupling characteristics ; S4: Use the random forest algorithm to filter the parameters in S3, and select the parameters whose importance is greater than the set threshold; S5: Use the filtered parameters to build a sample; the sample includes parameter values, fault type, fault location, severity, and confidence level; S6: Train the neural network based on the samples to obtain the trained neural network. Use the trained neural network to continue to identify the faults of the water turbine and output the results.

2. The intelligent diagnosis method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion as described in claim 1, characterized in that, In step S2, the displacement is purely fluid-induced. The formula is: ； In the formula, This is a wet mode signal; It is a dry modal signal; This is a purely fluid excitation signal; In step S3, the peak value The calculation formula is: ； In the formula, and These represent the maximum and minimum values ​​of the pure fluid excitation displacement within the current time window, respectively. Root mean square value The calculation formula is: ； In the formula, N is the number of sampling points within the current time window; The displacement of the i-th sampling point is the pure fluid excitation displacement. Skewness The calculation formula is: ； In the formula, This is the average value of the fluid excitation displacement at all sampling points within the current window; This represents the standard deviation of the fluid excitation displacement at all sampling points within the current window. peak skewness Root mean square value Using time-domain characteristics, we analyze turbine leakage faults from a time-domain perspective.

3. The intelligent diagnosis method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion according to claim 2, characterized in that, In step S3, the pure fluid excitation displacement sequence is... Perform a Fast Fourier Transform to generate a real-time spectrogram. Amplitude ratio The calculation formula is: ； In the formula, Indicated in the real-time spectrum graph In the figure, the amplitude at the blade passing frequency corresponding to the rated speed; express After fast Fourier transform, the amplitude at the blade passing frequency corresponding to the rated speed is extracted; Spectral centroid The calculation formula is: ； In the formula, Represents real-time spectrum The frequency value in; Real-time spectrum medium frequency The mold at the location; This represents the maximum analysis frequency. Harmonic distortion The calculation formula is: ； In the formula, Real-time spectrum The frequency corresponds to the fundamental frequency of rotor rotation. Representing the spectrum The amplitude of vibration at frequencies that are integer multiples of the fundamental frequency (n ≥ 2) is the amplitude ratio. Spectrum centroid Harmonic distortion The frequency domain analysis of water turbine leakage faults.

4. The intelligent diagnosis method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion as described in claim 1, characterized in that, In step S3, Phase difference The calculation formula is: ； In the formula, This indicates the time lag of the maximum values ​​of the X and Y displacement sensor signals under wet conditions. Indicates the real-time rotation frequency. This indicates the real-time rotational speed obtained from the key phase signal; Trajectory drift The calculation formula is: ； In the formula, Indicates the real-time center coordinates of the rotor; Indicates the center coordinates of the rotor under no-load conditions; phase difference and trajectory drift The water turbine leakage fault was analyzed from the trajectory characteristics.

5. The intelligent diagnosis method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion according to claim 1, characterized in that, In step S3, Pressure-vibration correlation coefficient The calculation formula is: ； In the formula, Describing covariance, This represents the standard deviation of the pressure signal. Indicates the standard deviation of the displacement signal; The time-series data representing the pressure signal; the pressure-vibration correlation coefficient R, used to analyze turbine leakage faults from the perspective of fluid-structure interaction.

6. The intelligent diagnosis method for water pump turbine leakage faults based on multi-dimensional rotor dynamic characteristic fusion according to claim 1, characterized in that, In step S3, Reference fluid dynamics The formula is: ； In the formula, The density of water; This refers to the real-time impeller outlet circumferential speed within the current time window. Dimensionless fluid forces The formula is: ； In the formula, It is the fluid excitation force; For reference hydrodynamics; Dimensionless coupling stiffness The formula is: ； In the formula, The peak value of the dimensionless hydrodynamic force. This represents the dimensionless peak value of the vibration displacement. Coupling stiffness coefficient in dynamic coupling characteristics The formula is: ； In the formula, The peak value of the hydrodynamic force with dimensions. The peak displacement of the dimensional fluid vibration; reference hydrodynamics. Dimensionless fluid forces and dimensionless coupling stiffness The coupling stiffness coefficient in the dynamic coupling characteristics is used to analyze the water turbine leakage fault from the perspective of coupling correlation characteristics.

Images