Hydraulic machinery stall vortex online identification method based on nonlinear dynamic characteristics

Abstract

The invention discloses a hydraulic machinery stall vortex online identification method based on nonlinear dynamic characteristics, and relates to the field of hydraulic machinery operation monitoring and fault diagnose.The hydraulic machinery stall vortex online identification method can distinguish loading and unloading processes by introducing an empirical formula and a model of a guide vane change rate, accurately identifies a vortex strip state on a hysteresis loop, and improves the accuracy of stall vortex identification. The dynamic error is reduced by more than 80%; secondly, through an empirical formula, the input of the neural network has a clear hydraulics meaning (such as a velocity triangle relation), and the unreliability of a pure black box model is avoided; wherein the complex fluid physical law is concentrated in an empirical formula, the neural network only needs to process residual mapping, the operation speed is extremely high, and the method can be deployed in an in-situ control unit.

Description

A Method for Online Identification of Stall Vortex in Hydraulic Machinery Based on Nonlinear Dynamic Characteristics Technical Field

[0001] This invention belongs to the field of hydraulic machinery operation monitoring and fault diagnosis, and specifically relates to an online identification method for hydraulic machinery stall vortex based on nonlinear dynamic characteristics. Background Technology

[0002] When hydraulic machinery (such as water turbines and pump-turbines) operates outside their optimal operating range, the water flowing out of the runner outlet often exhibits significant residual vortex flow. When the vortex intensity exceeds a critical value, vortex rupture occurs within the draft tube, forming a spiral stall vortex band. This vortex band induces low-frequency, high-amplitude pressure pulsations, leading to severe unit vibration and even power oscillations throughout the entire hydraulic system.

[0003] In existing technologies, the identification of such vortices mainly relies on two approaches: 1. Computational Fluid Dynamics (CFD) simulation: Although highly accurate, it is computationally time-consuming and cannot meet the needs of real-time online monitoring; 2. Traditional Frequency Domain Analysis (FFT): Performing Fourier transform on the pressure sensor signal. However, when a hydroelectric turbine participates in grid frequency regulation and peak shaving, its operating conditions change extremely rapidly (such as during load shedding), at which point the flow field exhibits strong non-stationarity and hysteresis effects. That is, at the same position when the guide vane opening is open and closed, the state of the vortex band is completely different due to fluid inertia and the historical memory of the flow. Traditional methods cannot identify this dynamic difference, leading to misjudgments by the control system. Summary of the Invention

[0004] To address the problems in related technologies, this invention proposes an online identification method for hydraulic machinery stall vortices based on nonlinear dynamic characteristics, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0005] To solve the above technical problems, the present invention is achieved through the following technical solution: The present invention is an online identification method for stall vortex of hydraulic machinery based on nonlinear dynamic characteristics, comprising the following steps: S1, real-time synchronous acquisition of the operating parameters of the hydraulic machinery, including rotational speed, operating head, guide vane opening, runner blade angle β, and dynamic pressure signal of the tailrace pipe monitoring section. S2. Preprocess the collected parameters. Based on the similarity law of fluid dynamics, construct a high-dimensional feature vector that can characterize the transient characteristics of the flow field using the proxy variable of swirling intensity and dynamic trend term. S3. The high-dimensional feature vector The input is fed into a pre-trained recurrent neural network model with temporal memory; S4, the neural network model outputs the dimensionless coefficient of the stall vortex intensity at the current moment. and vortex belt precession frequency .

[0006] Preferably, the high-dimensional feature vector in S2 It includes proxy variables for characterizing the swirl intensity of the flow field at the turbine runner outlet. The calculation uses the following formula: ;in: per unit rotational speed, H represents the rotational speed, and H represents the working head. Unit flow rate; For guide vane opening, Where A is the diameter of the wheel; B is the diameter of the wheel. All of these are geometric correction constants determined through model test fitting for specific impeller blade profiles; Q is the actual flow rate; this formula is used to approximately quantify the ratio of tangential velocity to axial velocity at the tailrace pipe inlet section without performing full flow field measurements, serving as a physical criterion for neural networks to determine whether a stall vortex has been generated.

[0007] Preferably, the high-dimensional feature vector in S2 It also includes dynamic trend terms for capturing the hysteresis effect of the flow field. Its mathematical expression includes the first derivative of the guide vane opening with respect to time and higher-order coupling terms, as follows: ;in: The item indicates whether the unit is in a loading (opening degree increased) or unloading (opening degree decreased) state, where t represents time; For sign functions, and The product term is used to describe the nonlinear offset characteristics of the boundary layer separation point under different adjustment directions.

[0008] By introducing this set of empirical terms, the neural network is able to distinguish the different flow patterns caused by different historical paths under the same guide vane opening.

[0009] Preferably, the preprocessing and construction of high-dimensional feature vectors in S2 include the following steps: S21, identifying outliers in the rotational speed, working head, guide vane opening, runner blade angle, and tailrace pressure signal data using the 3σ criterion or box plot method, and then supplementing or removing them; S22, synchronizing the sampling time with the pressure pulsation signal as the reference time axis, and unifying the sampling frequency through linear resampling or cubic spline interpolation; S23, performing bandpass filtering on the dynamic pressure signal to remove low-frequency drift and high-frequency electrical noise; for slow variables such as operating parameters, smoothing is performed using a moving average window; S24, converting the original physical quantities into dimensionless features based on the principle of fluid dynamics similarity; the original physical quantities include rotational speed, flow coefficient, and pressure coefficient; S25, constructing transient feature vectors using a moving time window; S26, normalizing or Z-score standardizing all features.

[0010] Preferably, the neural network model in S3 adopts a long short-term memory network or a gated recurrent unit architecture, and the state update equation of its hidden layer embeds a regularization term of physical constraints, as follows: ;in For data fitting error, The surrogate variable for the swirling intensity is a function of time t. The critical swirl number threshold. is the regularization coefficient; the regularization term is used to force the network to suppress the output of stall vortex intensity when the swirling intensity is below the critical value, thereby incorporating the stability criterion of fluid dynamics into the network training process.

[0011] Preferably, the dimensionless coefficient of the stall vortex intensity The definition adopts a comprehensive empirical formula that includes pressure pulsation energy and resonance risk weights: ;in: The root mean square value of pressure pulsation predicted by the neural network; The identified vortex frequency. The natural frequency of the hydraulic system; As a resonance-sensitive factor, Where is the fluid density and g is the gravitational acceleration; this formula not only quantifies the pressure amplitude of the vortex belt, but also weights the degree of harm when the vortex belt frequency is close to the system's natural frequency through an exponential term, thus achieving a comprehensive evaluation of the destructiveness of the stall vortex.

[0012] Preferably, the constant term in the empirical formula is determined through a virtual-physical fusion method, specifically including: S41, obtaining the velocity triangle distribution at the runner exit using unsteady CFD calculation; S42, integrating the velocity triangle distribution and performing backfitting. Coefficients of the formula and S43. Correct the weighting coefficient of the dynamic trend term using real machine load shedding test data; S41 includes the following steps: S411. Establish a three-dimensional geometric model of the entire hydraulic machinery flow channel, the model including the inlet flow channel, guide vanes, impeller, and tailrace flow channel; then mesh the three-dimensional geometric model to obtain a computational mesh model; S412. Set boundary conditions and impeller rotation angular velocity, use the separated eddy simulation (DES) model or the scale adaptive simulation (SAS) model for unsteady solution and unsteady calculation to obtain the three-dimensional transient velocity field under stable operating conditions. Data; S413, Set up a monitoring section at the runner outlet, extract the axial velocity component and tangential velocity component at each time step on the monitoring section; then calculate the circumferential velocity based on the runner angular velocity and the radius of the monitoring section, and construct the absolute velocity triangle and relative velocity triangle by combining the axial velocity component and the tangential velocity component; S414, Perform time statistical processing on multiple runner cycles after stabilization to obtain the instantaneous distribution and periodic average distribution of the runner outlet velocity triangle; S42 includes the following steps: S421, Establish a three-dimensional full-channel numerical calculation model of the hydraulic machinery, using Unsteady numerical calculations are performed using either the Detached Eddy Simulation (DES) model or the Scale Adaptive Simulation (SAS) model to obtain three-dimensional transient flow field data under stable operating conditions; S422, a monitoring section is set at the runner outlet, and the instantaneous velocity components on the section are extracted, including axial velocity and tangential velocity, and the circumferential velocity is calculated based on the runner angular velocity to construct a velocity triangle distribution; S423, area-weighted integration is performed on the runner outlet section to obtain overall flow characteristic parameters, including mass flow rate weighted average tangential velocity, mass flow rate weighted average axial velocity, and circulation or swirl integral. S424. Based on the integral results in S423, construct a vortex surrogate expression, which includes two undetermined coefficients A and B; and based on the numerical calculation results under different operating conditions, calculate the actual vortex intensity parameters or pressure pulsation characteristic parameters, and construct a sample dataset; S425. Use the least squares method or regularized regression method to perform backfitting on the coefficients A and B in the vortex surrogate expression to obtain the optimal coefficient combination; S426. Use the fitted vortex surrogate expression in S425 for hydraulic machinery operating status assessment or subsequent data-driven model input.

[0013] Preferably, S43 specifically includes: constructing a measured hysteresis loop based on experimental data and calculating the hysteresis loop area; constructing a model hysteresis loop under given initial weighting coefficients; establishing an objective function composed of hysteresis area error and trajectory error; optimizing the pressure change rate weighting coefficient and pressure fluctuation intensity weighting coefficient by minimizing the objective function to obtain the optimal weighting coefficients, so that the model hysteresis loop matches the measured hysteresis loop in terms of area, shape and evolution trend, thereby improving stall identification accuracy.

[0014] Preferably, the controller receives the identified Value, when When the preset threshold is exceeded, the opening of the air supply valve is adjusted according to the following nonlinear control law. ; ;in It is a proportionality coefficient, corresponding to the direct contribution of the current vorticity state to aerodynamics and torque; These are differential coefficients, corresponding to the contribution of the vortex state change rate to aerodynamics and torque; the control law is used to utilize the identified vortex band intensity and its change rate, and in... As a gain scheduling factor, it enables adaptive suppression of strong stall vortex conditions.

[0015] An online stall vortex identification system for hydraulic machinery based on nonlinear dynamic characteristics includes a hydraulic machinery operating parameter acquisition module, a high-dimensional feature vector construction module, and a stall vortex intensity-vortex precession frequency output module. The hydraulic machinery operating parameter acquisition module is used to synchronously acquire the operating parameters of the hydraulic machinery in real time, including rotational speed, operating head, guide vane opening, runner blade angle, and dynamic pressure signal of the tailrace monitoring section. The high-dimensional feature vector construction module is used to construct a high-dimensional feature vector based on the data acquired by the hydraulic machinery operating parameter acquisition module. The stall vortex intensity-vortex precession frequency output module is used to input the high-dimensional feature vector into a pre-trained recurrent neural network model with time-series memory function, and output the dimensionless coefficient of stall vortex intensity and vortex precession frequency at the current moment.

[0016] The present invention has the following beneficial effects: 1. By introducing an empirical formula for the guide vane change rate, the model can distinguish between the "loading" and "unloading" processes, accurately identify the vortex band state on the hysteresis loop, and reduce the dynamic error by more than 80%; secondly, the empirical formula makes the input of the neural network have clear hydraulic meaning (such as velocity triangle relationship), avoiding the unreliability of pure black box models; the complex fluid physics laws have been condensed into the empirical formula, and the neural network only needs to process the residual mapping, which has a very fast calculation speed and can be deployed in the local control unit.

[0017] 2. Compared with ordinary BP neural networks that do not incorporate empirical formulas, the method of this invention significantly reduces the recognition error in the dynamic transition process and successfully reproduces the figure-eight hysteresis loop of pressure pulsation as the guide vane opening changes.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 is a schematic flowchart of an online identification method for stall vortices in hydraulic machinery based on nonlinear dynamic characteristics according to the present invention. Detailed Implementation

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1 (see Figure 1) illustrates an online identification method for stall vortices in hydraulic machinery based on nonlinear dynamic characteristics. The method includes the following steps: S1. Real-time synchronous acquisition of the hydraulic machinery's operating parameters, including rotational speed, operating head, guide vane opening, runner blade angle, and dynamic pressure signals from the tailrace monitoring section. S2. Preprocess the collected parameters. Based on the similarity law of fluid dynamics, construct a high-dimensional feature vector that can characterize the transient characteristics of the flow field using the proxy variable of swirling intensity and dynamic trend term. The preprocessing and high-dimensional feature vector construction described in S2 include the following steps: S21, processing the real-time acquired rotational speed, operating head, guide vane opening, runner blade angle, and tailrace pressure signals.

[0024] Initial screening is performed; specifically, outliers are identified using the 3σ criterion or box plot method; when the continuous missing time is less than 5 sampling periods, linear interpolation is used to fill in the gaps, and if it exceeds this threshold, the data segment is discarded. A typical setting is: outlier threshold coefficient k=3; S22, due to differences in sampling frequencies among different sensors, synchronization alignment is first performed using the pressure pulsation signal as the reference time axis; the sampling frequency is then unified through linear resampling or cubic spline interpolation. For hydraulic machinery pressure pulsation analysis, a typical setup is... =10~20 ;in =n / 60 is the rotational frequency to ensure sufficient capture of the blade frequency and its harmonics; S23, bandpass filtering is applied to the dynamic pressure signal, with the filtering range generally set to 0.2. ~20 To remove low-frequency drift and high-frequency electrical noise, a fourth-order Butterworth filter can be used. For slow variables such as operating parameters, a moving average window smoothing is used, with a recommended window length of L = 5 to 10 time steps; S24. Based on the principle of fluid dynamics similarity, the original physical quantities are converted into dimensionless features to reduce the scale difference between different operating conditions, for example: unit rotational speed: Flow coefficient: Pressure coefficient: Where D is the diameter of the wheel, Indicates rotational speed. Indicates the working head. Let p be the fluid density and p be the fluid pressure; if the flow rate Q cannot be obtained directly, it can be estimated using empirical characteristic curves; S25, construct the transient eigenvector using a sliding time window; let the window length be... Step size is Then the input at time t is: Commonly used in engineering =1 to 3 wheel cycles to ensure complete dynamic and static interference information is included; Indicates the angle of the impeller blades; This represents a dynamic signal, which is dynamic data reflecting the operating status of the equipment. t represents the current time. The length of the sliding time window is typically 1 to 3 wheel cycles in engineering to cover the complete interference process between the moving and stationary components; Formula Indicates from A dynamic signal sequence within a continuous time range from time t to time t; S26. To improve the training stability of the recurrent neural network, all features are subjected to Min-Max normalization or Z-score standardization. ;in, The result represents the normalization or standardization process, μ is the mean of the training set, and σ is the standard deviation of the training set; μ and σ are obtained from the training set statistics and remain consistent during the validation and testing phases; the high-dimensional feature vector mentioned in S2 It includes proxy variables for characterizing the swirl intensity of the flow field at the turbine runner outlet. The calculation uses the following formula: ;in: per unit rotational speed, H represents the rotational speed, and H represents the working head. Unit flow rate; Where A is the diameter of the wheel; B is the diameter of the wheel. All are geometric correction constants determined through model test fitting for specific impeller blade profiles; Q is the actual flow rate; this formula is used to approximately quantify the ratio of tangential velocity to axial velocity at the tailrace pipe inlet section without performing full flow field measurements, serving as a physical criterion for neural networks to determine whether a stall vortex has been generated; the high-dimensional feature vector mentioned in S2 It also includes dynamic trend terms for capturing the hysteresis effect of the flow field. Its mathematical expression includes the first derivative of the guide vane opening with respect to time and higher-order coupling terms, as follows: ;in: The item indicates whether the unit is in a loading (opening degree increased) or unloading (opening degree decreased) state, where t represents time; For sign functions, and The product term is used to describe the nonlinear offset characteristics of the boundary layer separation point under different adjustment directions.

[0025] By introducing this set of empirical terms, the neural network can distinguish the differentiated flow patterns caused by different historical paths under the same guide vane opening; S3, the high-dimensional feature vector The high-dimensional feature vector is input into a pre-trained recurrent neural network model with temporal memory; After being input into a pre-built recurrent neural network model, the sample data is first normalized, and training and validation sets are constructed in time series order. The hidden state of the network at time t is determined by the current input vector and the hidden state at the previous time step, thus achieving the memorization of historical information and dynamic feature extraction. During the training phase, backpropagation is used to optimize the network parameters through a time-series algorithm: first, forward propagation is performed to obtain the predicted output, then a loss function is constructed based on the error between the predicted value and the true label, and finally, the gradient is calculated by expanding the network structure along the time dimension.

[0026] To improve model convergence stability, adaptive optimization algorithms such as Adam are typically used to update weight parameters, and a learning rate decay strategy is implemented to avoid oscillations and overfitting. Simultaneously, Dropout regularization is introduced during training to enhance the model's generalization ability, and an early stopping mechanism monitors the validation set loss change, terminating training when model performance no longer improves. Finally, the trained RNN model is used to predict new time-series inputs, achieving dynamic estimation of target flow field features or performance parameters. The neural network model described in S3 employs a Long Short-Term Memory (LSTM) network or gated recurrent unit (GRU) architecture, and its hidden layer state update equations embed regularization terms based on physical constraints, as follows: ;in For data fitting error, The surrogate variable for the swirling intensity is a function of time t. The critical swirl number threshold. S4 is the regularization coefficient; the regularization term is used to force the network to suppress the output of stall vortex intensity when the swirling intensity is below the critical value, thereby incorporating the stability criterion of fluid dynamics into the network training process; S5, the neural network model outputs the dimensionless coefficient of stall vortex intensity at the current moment. and vortex belt precession frequency This enables real-time identification of stall vortices; the dimensionless coefficient of stall vortex intensity... The definition adopts a comprehensive empirical formula that includes pressure pulsation energy and resonance risk weights: ;in: The root mean square value of pressure pulsation predicted by the neural network; The identified vortex frequency. The natural frequency of the hydraulic system; As a resonance-sensitive factor, Let g be the fluid density and g be the gravitational acceleration. This formula not only quantifies the pressure amplitude of the vortex belt but also weights the degree of harm when the vortex belt frequency approaches the system's natural frequency through an exponential term, achieving a comprehensive evaluation of the destructiveness of the stall vortex. The constant term in the empirical formula is calibrated through a virtual-physical fusion method, specifically including: S41, obtaining the velocity triangle distribution at the runner outlet using unsteady CFD calculation; S42, integrating the velocity triangle distribution and backfitting. Coefficients of the formula and S43. Correct the weighting coefficient of the dynamic trend term using real machine load shedding test data; S41 includes the following steps: S411. Establish a three-dimensional geometric model of the entire flow channel of the hydraulic machinery, the model including the inlet flow channel, guide vanes, impeller, and tailrace flow channel; then mesh the three-dimensional geometric model to obtain a computational mesh model; S412. Set boundary conditions and impeller rotational angular velocity, use the separated eddy simulation (DES) model or the scale adaptive simulation (SAS) model to perform unsteady solution and unsteady calculation, and obtain three-dimensional transient velocity field data under stable operating conditions; S413. Set a monitoring section at the impeller outlet, and extract the axial velocity components and shear velocity components at each time step on the monitoring section. The axial velocity component is calculated; then, the circumferential velocity is calculated based on the angular velocity of the runner and the radius of the monitoring section, and the absolute velocity triangle and relative velocity triangle are constructed by combining the axial velocity component and the tangential velocity component; S414, the time statistics of multiple runner cycles after stabilization are performed to obtain the instantaneous distribution and periodic average distribution of the runner outlet velocity triangle; S42 includes the following steps: S421, a three-dimensional full-channel numerical calculation model of the hydraulic machinery is established, and unsteady numerical calculation is performed using the separated eddy simulation (DES) model or the scale adaptive simulation (SAS) model to obtain three-dimensional transient flow field data under stable operating conditions; S422, a monitoring section is set at the runner outlet, and the instantaneous velocity components on the section are extracted, including The process involves: S423, calculating the axial and tangential velocities and the circumferential velocity based on the runner angular velocity to construct a velocity triangle distribution; S424, performing area-weighted integration on the runner outlet section to obtain overall flow characteristic parameters, including mass flow rate weighted average tangential velocity, mass flow rate weighted average axial velocity, and circulation or swirl integral; S425, constructing a swirl proxy expression based on the integration results in S423, the expression including two undetermined coefficients A and B; and calculating the actual swirl intensity parameters or pressure pulsation characteristic parameters based on numerical calculation results under different operating conditions to construct a sample dataset; S426, using the least squares method or regularized regression method to refine the coefficients A and B in the swirl proxy expression. The process involves: performing a reverse fitting solution to obtain the optimal coefficient combination; S426, using the fitted vortex proxy expression from S425 for hydraulic machinery operating status assessment or subsequent data-driven model input; S43 specifically includes: constructing a measured hysteresis loop based on experimental data and calculating the hysteresis loop area; constructing a model hysteresis loop under given initial weighting coefficients; establishing an objective function composed of hysteresis area error and trajectory error; optimizing the pressure change rate weighting coefficient and pressure fluctuation intensity weighting coefficient by minimizing the objective function to obtain the optimal weighting coefficients, so that the model hysteresis loop matches the measured hysteresis loop in area, shape, and evolution trend, thereby improving stall identification accuracy; the controller receives the identified... Value, when When the preset threshold is exceeded, the opening of the air supply valve is adjusted according to the following nonlinear control law. ; ;in It is a proportionality coefficient, corresponding to the direct contribution of the current vorticity state to aerodynamics and torque; These are differential coefficients, corresponding to the contribution of the vortex state change rate to aerodynamics and torque; the control law is used to utilize the identified vortex band intensity and its change rate, and in... As a gain scheduling factor, it enables adaptive suppression of strong stall vortex conditions; for example, a mixed-flow turbine with a single unit capacity of 200MW is used.

[0027] SCADA data acquisition system with 100 Frequency acquisition guide vane opening Rotation speed The pressure pulsation signals at the inlet of the volute and the tailrace cone section.

[0028] To enable neural networks to understand the physical mechanisms of hydraulics, instead of directly using raw data, the following empirical features are calculated: vortex number surrogate quantity. Theoretically, the formation of vortex bands in the tailrace depends on the number of swirls. ; ;in For the projected length, From a basic perspective, this formula reflects that: the higher the unit rotational speed, the smaller the unit flow rate, and the larger the guide vane angle (under a specific geometry), the larger the tangential component of the fluid relative to the axial component, and the easier it is to form vortices.

[0029] Hysteresis indicator: The turbine is shedding load ( When ), the guide vanes close rapidly. At this point, the fluid inertia is extremely high, and the collapse of the vortex band lags behind the guide vane movement. During the start-up process ( The formation of vortex zones follows different pathways; among them, The rate of change of the guide vane opening (time derivative) reflects whether the guide vane is open or closed, and how fast it changes; therefore, it has the following structural features: These two features clearly inform the neural network of its current adjustment direction and nonlinearity, enabling it to accurately predict the upper and lower branches of the hysteresis loop. The neural network model design employs a 3-layer LSTM network structure. The input layer contains the values ​​calculated using the above empirical formula. and basic parameters and according to ; The calculated dynamic trend term The dynamic trend item It consists of changes in operating parameters and is used to characterize the direction of change in the operating path, thereby depicting the hysteresis effect of the flow field.

[0030] Hidden Layers: Two LSTM layers are set up, with 32 nodes in each layer. The "gating" mechanism of the LSTM cells can effectively store the historical state information of the flow (time constant of about 2-5 seconds), which perfectly matches the hysteresis effect caused by the inertia of water flow.

[0031] Output layer: Output vortex band intensity (Linked to pressure pulsation amplitude normalization) and frequency .

[0032] This project utilizes CFD transient simulation data (using the SAS turbulence model to simulate the load shedding process) for pre-training, and then combines it with three months of on-site measured data from the power plant for fine-tuning. ).

[0033] Results Comparison: Compared with the ordinary BP neural network without the introduction of empirical formulas, the recognition error of the method of the present invention in the dynamic transition process is reduced from 18% to less than 3%, and the figure-eight hysteresis loop of pressure pulsation with the change of guide vane opening is successfully reproduced.

[0034] Example 2 This example discloses an online identification system for stall vortices in hydraulic machinery based on nonlinear dynamic characteristics. The system can implement the methods of the above examples and includes a hydraulic machinery operating parameter acquisition module, a high-dimensional feature vector construction module, and a stall vortex intensity-vortex precession frequency output module. The hydraulic machinery operating parameter acquisition module is used to synchronously acquire the operating parameters of the hydraulic machinery in real time, including rotational speed, operating head, guide vane opening, runner blade angle, and dynamic pressure signal of the tailrace monitoring section. The high-dimensional feature vector construction module is used to construct a high-dimensional feature vector based on the data acquired by the hydraulic machinery operating parameter acquisition module. The stall vortex intensity-vortex precession frequency output module is used to input the high-dimensional feature vector into a pre-trained recurrent neural network model with time-series memory function and output the dimensionless coefficient of stall vortex intensity and vortex precession frequency at the current moment.

[0035] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0036] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

Claims

1. A method for online identification of stall vortices in hydraulic machinery based on nonlinear dynamic characteristics, characterized in that, Includes the following steps: S1. Real-time synchronous acquisition of hydraulic machinery operating parameters, including rotational speed, operating head, guide vane opening, runner blade angle, and dynamic pressure signal at the tailrace monitoring section. ； S2. Preprocess the collected parameters. Based on the similarity law of fluid dynamics, construct a high-dimensional feature vector that can characterize the transient characteristics of the flow field using the proxy variable of swirling intensity and dynamic trend term. ； S3, convert the high-dimensional feature vector The input is fed into a pre-trained recurrent neural network model with temporal memory; S4, the neural network model outputs the dimensionless coefficient of the stall vortex intensity at the current moment. and vortex belt precession frequency 。 2. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 1, characterized in that, The high-dimensional feature vectors mentioned in S2 It includes proxy variables for characterizing the swirl intensity of the flow field at the turbine runner outlet. The calculation uses the following formula: ;in: per unit rotational speed, H represents the rotational speed, and H represents the working head. Unit flow rate; For guide vane opening, Where A is the diameter of the wheel; B is the diameter of the wheel. All of these are geometric correction constants determined through model tests for specific impeller blade profiles; Q is the actual flow rate.

3. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 2, characterized in that, The high-dimensional feature vectors mentioned in S2 It also includes dynamic trend terms for capturing the hysteresis effect of the flow field. Its mathematical expression includes the first derivative of the guide vane opening with respect to time and higher-order coupling terms, as follows: ;in: The item indicates whether the unit is in a loading or unloading state, where t represents time; For sign functions, and The product term is used to describe the nonlinear offset characteristics of the boundary layer separation point under different adjustment directions.

4. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 1, characterized in that, The preprocessing and high-dimensional feature vector construction described in S2 include the following steps: S21, using the 3σ criterion or box plot method to identify outliers in the rotational speed, working head, guide vane opening, runner blade angle, and tailrace pressure signal data, and then supplementing or removing them; S22, synchronizing the sampling time with the pressure pulsation signal as the reference time axis, and unifying the sampling frequency through linear resampling or cubic spline interpolation; S23, performing bandpass filtering on the dynamic pressure signal to remove low-frequency drift and high-frequency electrical noise; for slow variables such as operating parameters, a moving average window is used for smoothing; S24, based on the principle of fluid dynamics similarity, converting the original physical quantities into dimensionless features; the original physical quantities include rotational speed, flow coefficient, and pressure coefficient; S25, constructing transient feature vectors using a moving time window; S26, normalizing or Z-score standardizing all features.

5. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 2, characterized in that: The neural network model described in S3 employs a long short-term memory network or a gated recurrent unit architecture, and its hidden layer state update equations embed regularization terms based on physical constraints, as follows: in For data fitting error, The surrogate variable for the swirling intensity is a function of time t. The critical swirl number threshold. is the regularization coefficient; the regularization term is used to force the network to suppress the output of stall vortex intensity when the vortex intensity is below the critical value.

6. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 3, characterized in that, The dimensionless coefficient of stall vortex intensity The definition adopts a comprehensive empirical formula that includes pressure pulsation energy and resonance risk weights: in: The root mean square value of pressure pulsation predicted by the neural network; The identified vortex frequency. The natural frequency of the hydraulic system; As a resonance-sensitive factor, Let g be the fluid density and g be the acceleration due to gravity.

7. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 6, characterized in that, The constant term in the empirical formula is determined through a virtual-physical fusion method, specifically including: S41, obtaining the velocity triangle distribution at the runner exit using unsteady CFD calculation; S42, integrating the velocity triangle distribution and backfitting. Coefficients of the formula and S43. Correct the weighting coefficient of the dynamic trend term using real machine load shedding test data; S41 includes the following steps: S411. Establish a three-dimensional geometric model of the entire hydraulic machinery flow channel, the model including the inlet flow channel, guide vanes, impeller, and tailrace flow channel; then mesh the three-dimensional geometric model to obtain a computational mesh model; S412. Set boundary conditions and impeller rotation angular velocity, use the separated eddy simulation (DES) model or the scale adaptive simulation (SAS) model for unsteady solution and unsteady calculation to obtain the three-dimensional transient velocity field under stable operating conditions. Data; S413, Set up a monitoring section at the runner outlet, extract the axial velocity component and tangential velocity component at each time step on the monitoring section; then calculate the circumferential velocity based on the runner angular velocity and the radius of the monitoring section, and construct the absolute velocity triangle and relative velocity triangle by combining the axial velocity component and the tangential velocity component; S414, Perform time statistical processing on multiple runner cycles after stabilization to obtain the instantaneous distribution and periodic average distribution of the runner outlet velocity triangle; S42 includes the following steps: S421, Establish a three-dimensional full-channel numerical calculation model of the hydraulic machinery, using Unsteady numerical calculations are performed using either the Detached Eddy Simulation (DES) model or the Scale Adaptive Simulation (SAS) model to obtain three-dimensional transient flow field data under stable operating conditions; S422, a monitoring section is set at the runner outlet, and the instantaneous velocity components on the section are extracted, including axial velocity and tangential velocity, and the circumferential velocity is calculated based on the runner angular velocity to construct a velocity triangle distribution; S423, area-weighted integration is performed on the runner outlet section to obtain overall flow characteristic parameters, including mass flow rate weighted average tangential velocity, mass flow rate weighted average axial velocity, and circulation or swirl integral. S424. Based on the integral results in S423, construct a vortex surrogate expression, which includes two undetermined coefficients A and B; and based on the numerical calculation results under different operating conditions, calculate the actual vortex intensity parameters or pressure pulsation characteristic parameters, and construct a sample dataset; S425. Use the least squares method or regularized regression method to perform backfitting on the coefficients A and B in the vortex surrogate expression to obtain the optimal coefficient combination; S426. Use the fitted vortex surrogate expression in S425 for hydraulic machinery operating status assessment or subsequent data-driven model input.

8. The method for online identification of stall vortices in hydraulic machinery based on nonlinear dynamic characteristics according to claim 7, characterized in that, S43 specifically includes: constructing a measured hysteresis loop based on experimental data and calculating the hysteresis loop area; constructing a model hysteresis loop under given initial weighting coefficients; establishing an objective function composed of hysteresis area error and trajectory error; optimizing the pressure change rate weighting coefficient and pressure fluctuation intensity weighting coefficient by minimizing the objective function to obtain the optimal weighting coefficients, so that the model hysteresis loop matches the measured hysteresis loop in terms of area, shape and evolution trend.

9. The method for online identification of hydraulic machinery stall vortex based on nonlinear dynamic characteristics according to claim 6, characterized in that, The controller receives the identified Value, when When the preset threshold is exceeded, the opening of the air supply valve is adjusted according to the following nonlinear control law. ； ;in It is a proportionality coefficient, corresponding to the direct contribution of the current vorticity state to aerodynamics and torque; It is a differential coefficient, corresponding to the contribution of the vorticity state change rate to aerodynamics and torque; The control law is used to utilize the identified vortex intensity and its rate of change, and in accordance with... As a gain scheduling factor, it enables adaptive suppression of strong stall vortex conditions.

10. A system for implementing the online identification method for hydraulic machinery stall vortex based on nonlinear dynamic characteristics as described in any one of claims 1-9, characterized in that: The system includes a hydraulic machinery operating parameter acquisition module, a high-dimensional feature vector construction module, and a stall vortex intensity-vortex precession frequency output module. The hydraulic machinery operating parameter acquisition module is used to synchronously acquire the operating parameters of the hydraulic machinery in real time, including rotational speed, operating head, guide vane opening, runner blade angle, and dynamic pressure signal of the tailrace monitoring section. The high-dimensional feature vector construction module is used to construct a high-dimensional feature vector based on the data acquired by the hydraulic machinery operating parameter acquisition module. The stall vortex intensity-vortex precession frequency output module is used to input the high-dimensional feature vector into a pre-trained recurrent neural network model with time-series memory function, and output the dimensionless coefficient of stall vortex intensity and vortex precession frequency at the current moment.

Images