Five-hole probe calibration and data processing method based on deep learning

Abstract

The invention discloses a five-hole probe calibration and data processing method based on deep learning, relates to experimental measurement of aerodynamic performance of a gas turbine, and is used for quickly and accurately processing data acquired by a five-hole probe in the gas turbine. The method comprises the following steps: carrying out a five-hole probe calibration experiment; constructing a five-hole probe neural network dimensionless calibration data set; constructing and training a five-hole probe neural network; carrying out turbine blade grid pneumatic experiment and five-hole probe experiment data acquisition; experimental data processing based on a neural network; and inverting aerodynamic parameters. The internal flow of a gas turbine is complicated and accompanied by a large airflow angle and a strong pressure gradient, fine measurement is often adopted for fully capturing flow characteristics, and the five-hole probe calibration and data processing method based on deep learning is constructed to be used for realizing rapid and high-applicability five-hole probe data processing.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of gas turbine aerodynamic performance test, and particularly relates to a five-hole probe calibration and data processing method based on deep learning, which is used for realizing rapid and accurate processing of five-hole probe collected data in a gas turbine. BACKGROUND

[0002] There are complex aerodynamic characteristics in a gas turbine. In order to develop an advanced gas turbine, the aerodynamic characteristics in the gas turbine must be fully measured and evaluated in the design stage, thereby guiding the cooling layout and optimization design of the gas turbine. Therefore, a five-hole probe is usually used to perform fine scanning measurement on a key two-dimensional section. In order to realize accurate processing of the five-hole probe data, researchers have proposed a series of methods, but these methods have the following advantages and disadvantages:

[0003] (1) The yaw zero method is used for five-hole probe calibration. The advantage of this method is that the probe calibration and data processing are relatively fast, but the experimental measurement operation is complex, and it is necessary to ensure that the pressure measurement value is stable, which is undoubtedly not suitable for fine measurement of a gas turbine with complex flow characteristics;

[0004] (2) The non-zero position measurement method is used for five-hole probe calibration, and the calibration graph is processed by curve fitting and direct interpolation method, and the aerodynamic characteristics are processed by iteration. Since the calibration surface of the five-hole probe usually has obvious nonlinear characteristics, the prediction accuracy of the curve fitting and direct interpolation method is limited, and it is usually necessary to adjust according to the calibration surface;

[0005] (3) Some scholars consider that the internal flow of a gas turbine has a high deflection angle characteristic, and develop a method based on a neural network, such as patent CN118857655A. However, it ignores the fact that the flow velocity in the same gas turbine changes from subsonic to transonic. This method constructs a neural network at a specific Mach number, and the complexity of the flow velocity change of the gas turbine directly limits the applicability of this method. Moreover, the neural network itself is based on the specific Mach number characteristics, which limits the operation of iterative correction by assuming the Mach number. SUMMARY

[0006] In order to overcome the disadvantages of the above-mentioned prior art, the purpose of the present application is to provide a five-hole probe calibration and data processing method based on deep learning, which is a gas turbine data processing means, effectively realizes rapid and high-applicability five-hole probe data processing, and is more in line with the experimental and engineering requirements of a gas turbine.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present application is:

[0008] The five-hole probe calibration and data processing method based on deep learning comprises the following steps:

[0009] Step 1, five-hole probe data calibration is carried out to obtain a calibration data set;

[0010] Step 2, construct the dimensionless calibration data set required for the five-hole probe neural network;

[0011] Step 3, construct and train the five-hole probe neural network;

[0012] Step 4, carry out turbine cascade aerodynamic experiment and five-hole probe experimental data acquisition;

[0013] Step 5, experimental data processing based on neural network;

[0014] Step 6, inverse normalization aerodynamic parameter inversion.

[0015] Compared with the prior art, the beneficial effects of the present application are:

[0016] (1) Considering the demand for fine measurement of gas turbines, the present application carries out five-hole probe calibration based on non-zero position measurement method, and uses neural network nonlinear characteristic fitting to calibrate the surface, thereby reducing the operation difficulty of experimental measurement personnel;

[0017] (2) The network architecture based on KANs is adopted, so that the present application has more accurate and flexible calibration graph fitting capability compared with curve fitting, direct interpolation method and BP neural network based on traditional multilayer perception;

[0018] (3) Based on the network architecture of KANs, the present application can adopt a more complex calibration scheme, specifically, for the aerodynamic parameter range (such as Mach number, pitch angle, yaw angle, etc.) specially concerned for gas turbine flow, a local refined non-uniform calibration supplement point scheme can be adopted, thereby improving the accurate capture and processing precision for key aerodynamic characteristics;

[0019] (4) Based on the characteristics of large flow angle, strong pressure gradient and wide Mach number range in gas turbine, the expression form of dimensionless coefficient is designed specifically, which ensures the accuracy of data processing in the range of large flow angle, strong pressure gradient and wide Mach number. In addition, the designed dimensionless coefficient can directly predict the aerodynamic parameters without iterative correction by Mach number, thereby greatly improving the data processing efficiency in fine scanning measurement. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flow chart of a five-hole probe calibration and data processing method based on deep learning.

[0021] Figure 2 is the geometric structure and hole position definition of the five-hole probe.

[0022] Figure 3 is a five-hole probe to flow parameter definition.

[0023] Figure 4 is a calibration map of a preferred embodiment of the present application.

[0024] Figure 5 is a schematic diagram of a neural network architecture of the present application.

[0025] Figure 6 is a Mach number prediction effect diagram of a preferred embodiment of the present application. DETAILED DESCRIPTION

[0026] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and examples.

[0027] As Figure 1 shown, the five-hole probe calibration and data processing method based on deep learning includes the following steps:

[0028] Step 1: Perform five-hole probe calibration.

[0029] The geometric structure and hole positions of the five-hole probe are shown in Figure 2 , including one central hole and four holes evenly surrounding the central hole. The central hole is numbered 5, and the remaining four holes are numbered 1, 2, 3, and 4, respectively. Figure 2 The xyz three axes are shown in Figure 3 , which can be used to represent the orientation and distance relationship of the holes. The flow parameters of the five-hole probe can be referred to Figure 3 , where the flow velocity V can be decomposed according to

[0030]

[0031] where u, v, and w are the flow velocity components in the x, y, and z directions, respectively.

[0032] This embodiment uses the NASA open-source five-hole probe calibration dataset for verification to prove the applicability of the present application. The NASA dataset uses the non-nulling method and non-uniform calibration layout for five-hole probe data calibration, and its calibration map is shown in Figure 4 . Figure 4 The setting range and resolution of the pitch angle and yaw angle at a certain Mach number are given. The range of the pitch angle is [-30°, 30°], and the range of the yaw angle is [-25°, 25°]. In the calibration range of 0° to ±4° of the pitch angle and yaw angle, a resolution of 1° is used, in the range of ±4° to ±10°, a resolution of 2° is used, and in the remaining calibration range, a resolution of 5° is used. According to the above non-uniform calibration map, the flow characteristics in the low yaw angle and pitch angle range are accurately characterized.

[0033] The non-zero position measurement method adopted by the present application is that, in the calibration wind tunnel, according to the set calibration map, the five-hole probe is adjusted to the corresponding state and the pressure values measured by the five holes of the five-hole probe are obtained in sequence, and the calibration data set is composed of the state (pitch angle β, yaw angle α) of the five-hole probe and the measured aerodynamic parameters (Mach number Ma, aerodynamic total pressure P t , aerodynamic static pressure P s and the pressure values P j of the five pressure holes of the five-hole probe, j = 1, 2, 3, 4, 5.

[0034] The arrangement of the calibration map of the present application is determined according to the range and accuracy of the aerodynamic characteristics to be measured in the specific turbine aerodynamic experiment. The range of the aerodynamic characteristics includes the range of the pitch angle, the range of the yaw angle and the range of the Mach number, and the range of the aerodynamic characteristics of the calibration map includes the range of the two-dimensional cross-sectional aerodynamic characteristics concerned in the turbine aerodynamic experiment; the accuracy of the calibration map depends on the resolution or traversal step of the aerodynamic characteristics, and the higher the resolution / traversal step, the higher the accuracy of the obtained calibration map, but the calibration cost will increase. The uniform arrangement of the calibration map scheme is that the traversal step of each aerodynamic parameter is consistent in the range of the aerodynamic parameter.

[0035] It should be noted that the arrangement of the calibration map supports uniform arrangement and non-uniform arrangement (local encryption). The present application locally refines the resolution of the calibration map according to the specific aerodynamic phenomenon concerned in the aerodynamic experiment and according to the corresponding range of the specific aerodynamic phenomenon, so as to improve the measurement accuracy of the specific aerodynamic characteristic phenomenon. The above local refinement operation can achieve a balance between local accuracy and calibration cost, that is, the non-uniform arrangement of the calibration map scheme.

[0036] Step 2: Construct the dimensionless calibration data set required for the five-hole probe neural network.

[0037] The NASA open-source five-hole probe calibration data set is dimensionless to obtain the dimensionless calibration data set required for the five-hole probe neural network. The present application realizes the mapping relationship between the input characteristics and the output characteristics by constructing the five-hole probe neural network.

[0038] The dimensionless calibration data set adopts the dimensionless format as follows:

[0039]

[0040] In the formula, P i represents the pressure value of the i-th hole of the four holes of the five-hole probe except the center hole, i = 1, 2, 3, 4, P avg is used to represent the local static pressure.

[0041]

[0042] where P5 represents the pressure value of the center hole of the five-hole probe, P MAE The parameter is used to characterize the dynamic pressure differential, which has a continuous linear variation feature to expand the applicability in the environment of large atmospheric flow angle.

[0043]

[0044] The above four dimensionless parameters cp pd_1,s , cp pd_2,s , cp pd_3,s , cp pd_4,s are the dimensionless parameters of the pressure difference of the remaining four holes relative to the static pressure, used to characterize the pressure difference of each hole relative to the static pressure.

[0045]

[0046] The above four dimensionless parameters cp pd_1,t , cp pd_2,t , cp pd_3,t , cp pd_4,t are the dimensionless parameters of the pressure variation characteristics of the remaining four holes in the pitch and yaw directions, used to characterize the pressure variation characteristics in the pitch and yaw directions.

[0047]

[0048] cp m is a dimensionless parameter used to characterize the Mach number, avoiding the problem of determining the Mach number by iteration in the traditional calibration method, thereby improving the data processing speed in the fine measurement experiment.

[0049] The above nine parameters are used as input features of the five-hole probe neural network

[0050] In addition to the pitch angle β, the yaw angle α, and the Mach number Ma, the total pressure coefficient cp t , the static pressure coefficient cp s characterize the local pressure characteristics. Among them, the total pressure coefficient cp t , the static pressure coefficient cp s are expressed as follows:

[0051]

[0052] where P t represents the aerodynamic total pressure at the outlet of the wind tunnel, and P s represents the aerodynamic static pressure at the outlet of the wind tunnel.

[0053] The pitch angle β, the yaw angle α, the Mach number Ma, the total pressure coefficient cp t , and the static pressure coefficient cp sThe five parameters are taken as output features of the five-hole probe neural network

[0054] It should be noted that, in order to facilitate the five-hole probe neural network to train its actual input parameters and output parameters, the above parameters are normalized results, that is:

[0055] The actual input parameters are:

[0056] The actual output parameters are:

[0057] In the formula, each parameter is the normalized result of the corresponding parameter.

[0058] Step three, constructing and training the five-hole probe neural network.

[0059] According to the characteristics of the output parameters of the five-hole probe neural network, when the input parameters are consistent, two groups of neural networks are divided, one group of output parameters is the pitch angle β and the yaw angle α, and the other group of output parameters is the Mach number Ma, the total pressure coefficient cp t , and the static pressure coefficient cp s , which are referred to as neural network one and neural network two. Neural network one is used to predict the pitch angle β and the yaw angle α, and neural network two is used to predict the Mach number Ma, the total pressure coefficient cp t , and the static pressure coefficient cp s .

[0060] Both groups of neural networks (neural network one and neural network two) of the application adopt Kolmogorov-Arnold Networks (KANs) architecture, the initial hidden layer is 6 layers, the number of neural units of each hidden layer is 16, 32, 64, 64, 32, 16 in turn, the training optimization method adopts AdamW method, the loss function adopts mean square error loss (MSE Loss), and the optimizer adjustment method adopts ReduceLROnPlateau scheme. The neural network structure diagram is shown in Figure 5 .

[0061] The input parameters and output parameters of the neural network are linearly normalized. The normalization range of neural network one is [-1, 1], and the normalization range of neural network two is [0, 1].

[0062] Taking the normalization of the pitch angle β as an example:

[0063]

[0064] In the formula, based on the calibration range [-30°, 30°] of the pitch angle β, the normalized pitch angle parameter

[0065] The dimensionless calibration data set obtained in step two is shuffled using the Fisher-Yates shuffle algorithm, and is divided into a training set and a validation set, with an allocation ratio of 80% and 20%. The mean square error loss of the validation set is used to fine-tune the neural network hidden layer structure to achieve the best model precision under different calibration maps.

[0066] Step four: carry out turbine cascade aerodynamic experiment and five-hole probe experimental data acquisition.

[0067] Carry out turbine aerodynamic experiment, and use the calibrated five-hole probe to scan and measure the two-dimensional plane of interest to obtain five-hole probe five-hole pressure data.

[0068] In this embodiment, the validation set of the dimensionless calibration data set is used instead of the turbine cascade aerodynamic experiment and five-hole probe experimental data acquisition to verify the data processing precision of the method.

[0069] Step five: experimental data processing based on neural network.

[0070] The data in the validation set are structured according to the nine neural network dimensionless input parameter formats described in step two to obtain a dimensionless experimental data set.

[0071] The dimensionless experimental data set is linearly normalized according to the normalization features of the two groups of neural networks described in step three to obtain a normalized dimensionless experimental data set, and is input into the two groups of five-hole probe neural networks.

[0072] The output parameters of the two groups of five-hole probe neural networks are obtained

[0073] Step six: inverse normalization aerodynamic parameter inversion.

[0074] The actual output parameters obtained in step five are inverse normalized to obtain the pitch angle β, the yaw angle α, the Mach number Ma, the total pressure coefficient cp t , and the static pressure coefficient cp s . The inverse normalized characteristic parameters use the corresponding parameters used in the pitch angle β, the yaw angle α, the Mach number Ma, the total pressure coefficient cp t , and the static pressure coefficient cp s used in the normalization process, i.e., the maximum and minimum values used in the normalization process.

[0075] Taking the inverse normalization of the pitch angle β as an example:

[0076]

[0077] where, based on the calibration range [-30°, 30°] of the pitch angle β, the normalized value is The inverse normalization of the predicted β gives the predicted Mach number.

[0078] Figure 6 The predicted Mach number deviation of the present application is given in the following table. The predicted Mach number deviation is calculated by Figure 6 It can be seen that the average predicted Mach number deviation is between [-1.5%, 1%] in the range of the pitch and yaw angles concerned in the present example, which proves the prediction accuracy of the present application.

Claims

1. A five-hole probe calibration and data processing method based on deep learning, characterized in that: The following steps are involved: Step 1: calibrate the five-hole probe data to obtain a calibration data set; Step 2: construct the dimensionless calibration data set required by the five-hole probe neural network; Step 3, construct and train the five-hole probe neural network; Step 4: Conduct turbine blade cascade aerodynamic experiments and five-hole probe experimental data collection; Step 5, experimental data processing based on neural network; Step 6: Perform inversion of the denormalized aerodynamic parameters.

2. The five-hole probe calibration and data processing method based on deep learning according to claim 1 is characterized in that: In step 1, the five-hole probe data is calibrated using a non-nulling method as follows: In the calibration wind tunnel, according to the set calibration diagram, the five-hole probe is adjusted to the corresponding state, and the pressure values ​​measured at the five holes of the five-hole probe are obtained in sequence. The calibration data set is composed of the five-hole probe state and the measured aerodynamic parameters; the five-hole probe state includes the pitch angle β and the yaw angle α, and the measured aerodynamic parameters include the Mach number Ma, the total aerodynamic pressure P at the wind tunnel outlet t , wind tunnel outlet aerodynamic and static pressure P s And the pressure values ​​P of the five pressure measuring holes of the five-hole probe j , j=1,2,3,4,5.

3. The five-hole probe calibration and data processing method based on deep learning according to claim 2 is characterized in that: The setting of the calibration map is based on the aerodynamic characteristic range and accuracy required to be measured in the turbine aerodynamic experiment; the aerodynamic characteristic range includes the pitch angle range, the yaw angle range and the Mach number range, and the aerodynamic characteristic range of the calibration map includes the two-dimensional cross-sectional aerodynamic characteristic range of interest in the turbine aerodynamic experiment; the accuracy of the calibration map depends on the resolution or traversal step of the aerodynamic characteristics, the higher the resolution / the smaller the traversal step, the higher the accuracy of the obtained calibration map, and each aerodynamic parameter has a consistent traversal step within its aerodynamic parameter range, which is a uniformly arranged calibration map scheme; the resolution of the calibration map is locally refined according to the aerodynamic characteristic range corresponding to the specific aerodynamic phenomenon of interest in the aerodynamic experiment.

4. The five-hole probe calibration and data processing method based on deep learning according to claim 1, characterized in that: The step 2 is to convert the calibration data set into a dimensionless calibration data set; The mapping relationship between input features and output features is realized by constructing a five-hole probe neural network; Input features of the five-hole probe neural network for: The dimensionless parameter cp characterizes the difference between the pressure values ​​of the four holes except the central hole and the static pressure pd_1,s 、cp pd_2,s 、cp pd_3,s 、cp pd_4,s ; The dimensionless parameter cp that characterizes the pressure variation characteristics of the other four holes except the central hole in the pitch and yaw directions pd_1,t 、cp pd_2,t 、cp pd_3,t 、cp pd_4,t ; as well as The dimensionless parameter cp that characterizes the Mach number m ; The output characteristics of the five-hole probe neural network for: Pitch angle β, yaw angle α, Mach number Ma, total pressure coefficient cp t and the static pressure coefficient cp s ; Each parameter is normalized and used as the actual input parameter and actual output parameter of the five-hole probe neural network.

5. The five-hole probe calibration and data processing method based on deep learning according to claim 4 is characterized in that: The expressions of the dimensionless parameters are as follows: Where, P avg Indicates the local static pressure, P MAE represents the dynamic pressure difference, which can be expressed as: Where, P i represents the pressure value of the i-th hole among the four holes of the five-hole probe excluding the center hole, and P5 represents the pressure value of the center hole; The total pressure coefficient cp t , static pressure coefficient cp s The expression is as follows: Where P t represents the total aerodynamic pressure at the wind tunnel outlet, P s represents the aerodynamic and static pressure at the wind tunnel outlet.

6. The five-hole probe calibration and data processing method based on deep learning according to claim 4 or 5, characterized in that: In step 3, the five-hole probe neural network is divided into a neural network 1 and a neural network 2, wherein the neural network 1 is used to predict the pitch angle β and the yaw angle α, and the neural network 2 is used to predict the Mach number Ma and the total pressure coefficient cp. t and static pressure coefficient cp s ; The input parameters and output parameters are linearly normalized, where the normalization range of neural network 1 is [-1, 1] and the normalization range of neural network 2 is [0, 1]; The dimensionless calibration dataset is shuffled and divided into a training set and a validation set. The hidden layer structure of the neural network is fine-tuned by the mean square error loss of the validation set to achieve the best model accuracy under different calibration maps.

7. The five-hole probe calibration and data processing method based on deep learning according to claim 6, characterized in that: Both neural network one and neural network two adopt the Kolmogorov-Arnold Networks architecture, with an initial hidden layer of 6 layers, and the number of neural units in each hidden layer is 16, 32, 64, 64, 32, and 16 respectively. The training optimization method adopts the AdamW method, the loss function adopts the mean square error loss, and the optimizer adjustment method adopts the ReduceLROnPlateau scheme.

8. The five-hole probe calibration and data processing method based on deep learning according to claim 1, characterized in that: The step 4 is to carry out a turbine blade cascade aerodynamic experiment, use the calibrated five-hole probe to perform scanning measurement on the two-dimensional plane of interest, and obtain five hole pressure data of the five-hole probe.

9. The five-hole probe calibration and data processing method based on deep learning according to claim 6, characterized in that: In step 5, the collected experimental data are constructed according to the format of the input parameters to obtain a dimensionless experimental data set; the dimensionless experimental data set is linearly normalized according to the normalized characteristics of the neural network 1 and the neural network 2 to obtain a normalized dimensionless experimental data set, and the data is input into the neural network 1 and the neural network 2 to obtain actual output parameters.

10. The five-hole probe calibration and data processing method based on deep learning according to claim 6, characterized in that: In step 6, the actual output parameters obtained in step 5 are denormalized to obtain the pitch angle β, yaw angle α, Mach number Ma, and total pressure coefficient cp. t , static pressure coefficient cp s The characteristic parameters of the denormalization process are the pitch angle β, yaw angle α, Mach number Ma, total pressure coefficient cp in step 3 normalization process t , static pressure coefficient cp s The corresponding parameters used.

Images