Turbine cascade test boundary condition back calculation method and device based on sparse measurement data

Abstract

The invention discloses a turbine cascade test boundary condition back calculation method and device based on sparse measurement data, belongs to the technical field of gas turbine test data processing, and particularly relates to back calculation of boundary conditions of a cascade test. The problems that in the prior art, the calculated amount is large, only low Reynolds number flow or a simple three-dimensional even two-dimensional flow mode is concerned, and only a scene with the unchanged sensor number and the fixed position can be coped with are solved. The method comprises the following steps of: obtaining a space analysis network structure established based on a self-attention unit and a Transform encoder; and inputting the normalized sparse measurement data into the trained spatial analysis network structure to obtain a corresponding test boundary condition. The turbine cascade test boundary condition back calculation method and device based on sparse measurement data are suitable for back calculation of turbine cascade test boundary conditions.

Description

Technical Field

[0001] The invention relates to the technical field of gas turbine turbine test data processing, and in particular to the back calculation of boundary conditions of blade cascade tests. Background Art

[0002] Gas turbine is an important energy conversion device, which plays a very critical role in aviation power, ship propulsion, ground power generation and other industrial fields. The compressor, turbine and combustor are the three main components of gas turbines. In order to meet the performance requirements of the new generation of gas turbines, each component must be carefully designed to improve thermal efficiency and reduce the emission of greenhouse gases such as carbon dioxide. In the traditional design and optimization process, a large number of experiments and numerical simulations are carried out to evaluate candidate design schemes. Both methods have their advantages and disadvantages. On the one hand, although experimental data can provide information closer to basic physics, it is usually unable to cover the design parameter domain and the spatial domain of interest due to the limitations of measurement technology and cost. Therefore, the experimental data usually exhibits low temporal and spatial resolution. On the other hand, numerical simulation is often used to supplement unconducted experiments due to its flexibility. However, achieving high-fidelity simulation requires complex modeling, a large amount of computing resources and considerable time investment. Most importantly, accurate boundary conditions are required to match the test environment. In order to reduce the acquisition cost of high-value data, a method must be developed to back-calculate the experimental boundary conditions from sparse test data.

[0003] Existing sparse reconstruction techniques based on limited measurement data include methods based on linear theory and nonlinear methods based on machine learning.

[0004] The linear method is fast in calculation and has strong applicability for most problems. However, the linear method still has some limitations:

[0005] (1) The amount of computing required to execute POD in big data scenarios is unbearable, and it also limits the space for its continuous update and evolution;

[0006] (2) The process of extracting simplified basis vectors requires that the measurement points must be included in the data set used to construct the reference basis vectors, which limits the flexibility of the measurement locations;

[0007] (3) The mathematical nature of the POD method is linear operation, and its accuracy will be challenged when the application involves strong nonlinear dynamics.

[0008] Machine learning methods are also widely used to improve the model's ability to deal with nonlinear problems.

[0009] Most existing technologies focus on low Reynolds number flows (RE<500), or simpler three-dimensional or even two-dimensional flow patterns. However, the flow field inside the impeller machinery usually exhibits complex three-dimensional flow characteristics. The presence of complex wave structures (such as shock waves and expansion waves) leads to the formation of complex pressure and temperature distribution stripes on the surface of the turbine blades. In addition, existing technologies can usually only cope with scenarios where the number of sensors remains unchanged and the positions are fixed. Summary of the invention

[0010] The present invention proposes a method and device for back-calculating boundary conditions of a turbine blade grid test based on sparse measurement data, which solves the problems of the prior art, such as large amount of calculation, focusing only on low Reynolds number flow or relatively simple three-dimensional or even two-dimensional flow patterns, and only being able to cope with scenarios where the number of sensors remains unchanged and their positions are fixed.

[0011] The back-calculation method of turbine blade cascade test boundary conditions based on sparse measurement data of the present invention has the following technical solutions:

[0012] The method comprises the following steps:

[0013] Steps for obtaining the structure of a spatial analysis network based on self-attention units and Transformer encoders;

[0014] The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein:

[0015] The multi-head self-attention unit is an improved scaled dot-product attention:

[0016] Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d k is the number of features of the query and key; the superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution that sums to 1;

[0017]

[0018] In the self-attention calculation, Q, K, and V are the same value;

[0019] For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head:

[0020]

[0021] Steps for training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residuals using mean square error:

[0022]

[0023] Steps to obtain the structure of the spatial analysis network after training:

[0024] Mathematical description of the structure of the spatial analysis network after training:

[0025]

[0026] Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution;

[0027] A step for inputting normalized sparse measurement data into a trained spatial analysis network structure to obtain corresponding experimental boundary conditions.

[0028] Further, a preferred embodiment is provided, wherein the position encoding module is used for performing position encoding;

[0029] Position encoding implies relative position encoding and absolute position encoding:

[0030]

[0031] Relative position encoding is used to embed position information in the spatial domain; absolute position encoding is used to embed position information in the feature domain;

[0032] Relative position encoding is achieved through trigonometric matrix:

[0033]

[0034] Absolute position coding originates from the characteristics of binary computing: the frequency of alternation of higher-order values ​​is lower than that of lower-order values. Since the feature vector itself does not have position information in Euclidean space, the order of feature vectors is abstractly described by referring to the binary law.

[0035] A multilayer perceptron is a series of stacked fully connected layers, which is essentially a continuous linear operation; the parameters of each layer of neurons include the weight matrix and the bias vector Where m is the number of neurons in the current layer, and n is the input data size;

[0036] Let ζ(·) be the activation function, then the output of the neurons in the lth layer is:

[0037]

[0038] Select the leaky ReLU function to satisfy the mathematical relationship described below:

[0039]

[0040] The Transformer encoder consists of 6 identical network blocks; each network block consists of two layer regularizations, a multi-head self-attention unit, and a multi-layer perceptron;

[0041] The average pooling layer is an aggregation function.

[0042] Furthermore, a preferred implementation is provided, in which the Adam algorithm is used to train the spatial analysis network structure.

[0043] Furthermore, a preferred embodiment is provided, wherein the spatial analysis network structure is established on the basis of uniform grid division of the turbine blade surface, and the sampling grid nodes are used as sensor layout positions; at least 32 sampling points are arranged on the suction side and pressure side of the blade respectively.

[0044] Furthermore, a preferred embodiment is provided, in which 128 sampling points are arranged on the suction side of the blade, and 256 sampling points are arranged on the pressure side of the blade.

[0045] Further, a preferred implementation is provided, in the step of training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residual using the mean square error:

[0046] Adopt a random sampling strategy with training:

[0047] The sensor positions are randomly resampled in each training iteration, and each sampling process is independent of each other.

[0048] The present invention also proposes a turbine blade cascade test boundary condition back-calculation device based on sparse measurement data, and its technical solution is as follows:

[0049] The device comprises the following modules:

[0050] A module for obtaining a spatial analysis network structure based on self-attention units and Transformer encoders;

[0051] The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein:

[0052] The multi-head self-attention unit is an improved scaled dot-product attention:

[0053] Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d kis the number of features of the query and key; superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution with a sum of 1;

[0054]

[0055] In the self-attention calculation, Q, K, and V are the same value;

[0056] For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head:

[0057]

[0058] Module used to train the spatial analysis network structure, optimize model parameters, and calculate the corresponding residual using mean square error:

[0059]

[0060] Module to obtain the spatial analysis network structure after training:

[0061] Mathematical description of the structure of the spatial analysis network after training:

[0062]

[0063] Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution;

[0064] A module used to input normalized sparse measurement data into the trained spatial analysis network structure to obtain the corresponding experimental boundary conditions.

[0065] The present invention also proposes a computer device, and its technical solution is as follows:

[0066] A computer device comprises: a processor and a memory, wherein the memory is used to store executable instructions of the processor, and the processor is configured to execute the above-mentioned turbine blade cascade test boundary condition backcalculation method based on sparse measurement data by executing the executable instructions.

[0067] The present invention also proposes a computer storage medium, and its technical solution is as follows:

[0068] A computer storage medium stores a computer program, and when the computer program is run, the above-mentioned turbine blade grid test boundary condition backcalculation method based on sparse measurement data is executed.

[0069] The present invention also proposes a computer program product, and its technical solution is as follows:

[0070] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned turbine blade cascade test boundary condition backcalculation method based on sparse measurement data.

[0071] The present invention has the following beneficial effects:

[0072] 1. The method for back-calculating boundary conditions of a turbine blade array test based on sparse measurement data described in the present invention proposes a spatial analysis network structure (model) that has permutation invariance.

[0073] 2. The turbine blade test boundary condition back-calculation method based on sparse measurement data described in the present invention will effectively improve the coverage of the training data on the sample domain through a random sampling strategy accompanied by training, thereby improving the generalization and robustness of the model.

[0074] 3. The method for backcalculating boundary conditions of a turbine blade array test based on sparse measurement data described in the present invention uses a supervised learning method in machine learning and proposes a spatial analysis network structure based on a self-attention mechanism. For sparse discrete measurement data in a gas turbine turbine blade array test, it can effectively process data from any number of randomly arranged sensors, extract key features from them, and backcalculate the boundary conditions of the flow.

[0075] 4. The core contribution of the turbine blade cascade test boundary condition backcalculation method based on sparse measurement data described in the present invention is that it proposes a method that can be applied to complex flow field data of gas turbine turbines. The new architecture is suitable for the targeted design of any number of randomly arranged sensors and for obtaining the optimal model parameters θ * Training strategy

[0076] 5. The turbine blade cascade test boundary condition back-calculation method based on sparse measurement data described in the present invention can be applied to aircraft engine and gas turbine turbine blades to perform feature extraction on sparse test data and back-calculate the boundary conditions of the blade cascade test.

[0077] 6. The turbine blade cascade test boundary condition back-calculation method based on sparse measurement data described in the present invention can effectively reduce the cost of obtaining high-quality data in the blade cascade test.

[0078] 7. The turbine blade test boundary condition back-calculation method based on sparse measurement data described in the present invention can also provide solutions to key problems in the development process of the digital twin system. The scenario often faced in the operation and maintenance stage of the digital twin system is to evaluate the current operating status of the unit from limited sensor data. The method can integrate the measurement data of discretely arranged limited sensors, back-calculate the current inlet and outlet flow conditions of the unit, assist in evaluating possible operating risks, and provide numerical support for control system decision-making.

[0079] The method and device for backcalculating turbine blade cascade test boundary conditions based on sparse measurement data of the present invention are suitable for backcalculating turbine blade cascade test boundary conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0081] Figure 1 A schematic diagram of a spatial analysis network structure in one embodiment of the present invention;

[0082] Figure 2 This is a schematic diagram of the relative position encoding principle in one embodiment of the present invention;

[0083] Figure 3 A schematic diagram of the performance of a model (spatial analysis network structure) on a training set in one embodiment of the present invention;

[0084] Figure 4 A schematic diagram of the performance of a model (spatial analysis network structure) on a test set in one embodiment of the present invention;

[0085] Figure 5 In one embodiment of the present invention, the number of different sensors N s Schematic diagram of the impact on the performance of the model (spatial analysis network structure);

[0086] Figure 6 In one embodiment of the present invention, the number of different attention heads N h Schematic diagram of the performance of the model (spatial analysis network structure);

[0087] Figure 7 This is a schematic diagram of a data processing method in one embodiment of the present invention. DETAILED DESCRIPTION

[0088] In order to make the technical solutions and advantages of the present invention more clearly described, the specific implementation methods of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various implementation methods described below are only part of the preferred solutions of the present invention, rather than all implementation methods; the various implementation methods described below are intended to explain the present invention and cannot be understood as limitations on the present invention; the reasonable combination of technical features defined in the various implementation methods of the present invention, and all other implementation methods obtained by ordinary technicians in the field without creative work based on the implementation methods of the present invention, are within the scope of protection of the present invention.

[0089] In one embodiment, a back calculation method for turbine blade cascade test boundary conditions based on sparse measurement data is provided:

[0090] The method comprises the following steps:

[0091] Steps for obtaining the structure of a spatial analysis network based on self-attention units and Transformer encoders;

[0092] The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein:

[0093] The multi-head self-attention unit is an improved scaled dot-product attention:

[0094] Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d k is the number of features of the query and key; superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution with a sum of 1;

[0095]

[0096] In the self-attention calculation, Q, K, and V are the same value;

[0097] For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head:

[0098]

[0099] Steps for training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residuals using mean square error:

[0100]

[0101] Steps to obtain the structure of the spatial analysis network after training:

[0102] Mathematical description of the structure of the spatial analysis network after training:

[0103]

[0104] Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution;

[0105] A step for inputting normalized sparse measurement data into a trained spatial analysis network structure to obtain corresponding experimental boundary conditions.

[0106] It should be noted that the position codes of any two points with a spatial interval of δ are converted to each other through the sum and difference formula, that is, the position code calculated according to formula (2) naturally contains the relative spatial position information of the two points.

[0107] In this implementation, the basic structure of the spatial analysis network structure is as follows: Figure 1 shown.

[0108] In this embodiment, the absolute position code is as follows: Figure 2 shown.

[0109] In this embodiment, the multilayer perceptron is an important basic component of the neural network.

[0110] In this implementation, the structure of each network block is as follows: Figure 1 (c) as shown.

[0111] In this embodiment, the multi-head self-attention unit is the core computing component of the encoder.

[0112] One point that needs to be emphasized is that no matter how the sparse observations are arranged, the boundary conditions calculated inversely should remain unchanged, which is called the permutation invariance of the calculation results. The spatial analysis network structure (model) proposed by the method described in this embodiment has permutation invariance, and the following derivation proves this property:

[0113] The input tensors are organized as matrices Where n is the number of sensors, and d is the number of physical quantities measured;

[0114] First, the weights of the multilayer perceptron are shared by all sensors and are naturally permutation invariant;

[0115] Activation functions and regularization operations are essentially scaling operations and have nothing to do with the order of the data;

[0116] Formula (8) proves the permutation invariance of the scaled dot product attention computation, where is the observed value at a specific point;

[0117] The final mean pooling layer is an aggregation function that has nothing to do with the order of the data;

[0118] In summary, it has been proved that the proposed spatial analysis network structure (model) is permutation invariant.

[0119]

[0120] In addition, in one embodiment, the position encoding module is used to perform position encoding;

[0121] Position encoding implies relative position encoding and absolute position encoding:

[0122]

[0123] Relative position encoding is used to embed position information in the spatial domain; absolute position encoding is used to embed position information in the feature domain;

[0124] Relative position encoding is achieved through trigonometric matrix:

[0125]

[0126] Absolute position coding originates from the characteristics of binary computing: the frequency of alternation of higher-order values ​​is lower than that of lower-order values. Since the feature vector itself does not have position information in Euclidean space, the order of feature vectors is abstractly described by referring to the binary law.

[0127] A multilayer perceptron is a series of stacked fully connected layers, which is essentially a continuous linear operation; the parameters of each layer of neurons include the weight matrix and the bias vector Where m is the number of neurons in the current layer, and n is the input data size;

[0128] Let ζ(·) be the activation function, then the output of the neurons in the lth layer is:

[0129]

[0130] Select the leaky ReLU function to satisfy the mathematical relationship described below:

[0131]

[0132] The Transformer encoder consists of 6 identical network blocks; each network block consists of two layer regularizations, a multi-head self-attention unit, and a multi-layer perceptron;

[0133] The average pooling layer is an aggregation function.

[0134] In addition, in one embodiment, the Adam algorithm is used to train the spatial analysis network structure.

[0135] In addition, in one embodiment, the spatial analysis network structure is established on the basis of uniformly gridding the surface of the turbine blade, and the sampling grid nodes are used as sensor layout positions; at least 32 sampling points are arranged on the suction side and pressure side of the blade respectively.

[0136] In addition, in one embodiment, 128 sampling points are arranged on the suction side of the blade, and 256 sampling points are arranged on the pressure side of the blade.

[0137] In this implementation, a total of 128×256 sampling points are arranged. Assume that the number of sensors to be arranged is N. s , then the possible choices are This means that the problem to be solved has a very large sample domain.

[0138] In addition, in one embodiment, in the step of training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residual using the mean square error:

[0139] Adopt a random sampling strategy with training:

[0140] The sensor positions are randomly resampled in each training iteration, and each sampling process is independent of each other.

[0141] In this implementation, the random sampling strategy accompanying training will effectively improve the coverage of the sample domain by the training data, thereby improving the generalization and robustness of the model.

[0142] In this embodiment, the method can also provide solutions to key issues in the development process of the digital twin system. The scenario often faced in the operation and maintenance stage of the digital twin system is to evaluate the current operating status of the unit from limited sensor data. The method can integrate the measurement data of discretely arranged limited sensors, calculate the current inlet and outlet flow conditions of the unit, assist in evaluating possible operating risks, and provide numerical support for control system decision-making.

[0143] In this embodiment, the method uses the supervised learning method in machine learning and proposes a spatial analysis network structure based on the self-attention mechanism. For the sparse discrete measurement data in the gas turbine turbine blade grid test, it can effectively process the data from any number of randomly arranged sensors, extract key features from them, and inversely calculate the boundary conditions of the flow.

[0144] In this embodiment, the core contribution of the method is to propose a method that can be applied to complex flow field data of gas turbine turbines. The new architecture is suitable for the targeted design of any number of randomly arranged sensors and for obtaining the optimal model parameters θ * training strategy.

[0145] In addition, in one embodiment, a back-calculation device for turbine blade cascade test boundary conditions based on sparse measurement data is provided:

[0146] The device comprises the following modules:

[0147] A module for obtaining a spatial analysis network structure based on self-attention units and Transformer encoders;

[0148] The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein:

[0149] The multi-head self-attention unit is an improved scaled dot-product attention:

[0150] Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d k is the number of features of the query and key; superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution with a sum of 1;

[0151]

[0152] In the self-attention calculation, Q, K, and V are the same value;

[0153] For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head:

[0154]

[0155] Module used to train the spatial analysis network structure, optimize model parameters, and calculate the corresponding residual using mean square error:

[0156]

[0157] Module to obtain the spatial analysis network structure after training:

[0158] Mathematical description of the structure of the spatial analysis network after training:

[0159]

[0160] Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution;

[0161] A module used to input normalized sparse measurement data into the trained spatial analysis network structure to obtain the corresponding experimental boundary conditions.

[0162] In addition, in one embodiment, a specific embodiment of a method for back-calculating boundary conditions of a turbine blade cascade test based on sparse measurement data is provided:

[0163] Step 1, data preparation:

[0164] In practical applications, if targeted training of the network is required, sufficient experimental data should be prepared and data extraction and preprocessing should be performed.

[0165] First, uniform sampling points are arranged on the surface of the turbine blade. The surface of the turbine blade is evenly cut into 128 sections along the span direction, and 256 sampling points are evenly arranged on the suction side and pressure side along the blade arc. Starting from the trailing edge of the pressure side, across the leading edge line to the trailing edge of the suction side, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane. Finally, each complete blade surface field is constructed into an array of 2×128×512, which stores the pressure P and temperature data T of the blade surface.

[0166] Figure 7 The data extraction process is shown. Then, according to the statistical characteristics of the data, the data is normalized based on the following formula:

[0167]

[0168] Where μ is the mean and σ is the standard deviation.

[0169] The preparation of test data should take into account the requirements of the proposed spatial analysis network structure, and at least 32 measurement sensors should be placed on the suction side and the pressure side respectively. The digitization of sensor positions should refer to the sampling grid of 128×512 to select the closest node relative position. The test measurement data and boundary conditions are corresponding one by one to construct the training and test data sets.

[0170] Step 2, model training:

[0171] The network and its associated code can be written in Python and built on the open source deep learning framework PaddlePaddle. The network is trained on a computer platform equipped with specific hardware devices. It is usually recommended to use the Adam optimizer with default parameters.

[0172] Step 3, Performance Verification:

[0173] After the model is completed, the performance of the model on the dataset is evaluated by mean square error and relative error.

[0174] Step 4, model prediction:

[0175] When the model is trained, in the prediction stage, only the normalized test data needs to be provided and imported into the model to obtain the corresponding test boundary conditions. Among them, the normalization index should adopt the statistical characteristics of the CFD data in the training set.

[0176] In addition, in one embodiment, the performance of the turbine blade cascade test boundary condition back-calculation method based on sparse measurement data is verified:

[0177] To evaluate the performance of the proposed model, mean square error and relative error were used to assess the deviation of the predicted values ​​from the reference values.

[0178]

[0179] The scatter plot shows the overall performance of the model on the training and test data sets to more clearly demonstrate the inversion capability of the spatial analysis network structure (or spatial analysis model).

[0180] Figure 3 and Figure 4 The scatter plots of the predicted and actual values ​​of the model on the training and test datasets are shown respectively. 2s The absolute deviation and relative deviation are used to evaluate the prediction accuracy of the model. In the training set, except for a few outliers, the absolute error between the AOA prediction and the true value of most samples is less than ±1°, while Ma 2s The relative error of the prediction is less than ±2%. The model shows comparable prediction accuracy in the test set as the training set, confirming the good generalization ability of the proposed method.

[0181] Figure 5 Shows different number of sensors N s The mean square error and relative error are plotted below. s The prediction error shows a significant upward trend as N s When changing from 32 to 16, both the mean square error and the relative error increase dramatically. s >32, the mean square error is 10 -3 Below, the relative error is below 2%, and the average relative error is below 0.5%. s When <32, the mean square error increases to 10 -3The maximum mean square error is significantly increased to 10 -1 Correspondingly, the maximum and average RE increase to 3.5% and 1%, respectively. This shows that the proposed method requires at least 32 sensors to be arranged on the suction side and pressure side of the blade to perform sufficiently accurate boundary condition backcalculations.

[0182] Three models with different numbers of attention heads are further configured while other hyperparameters remain unchanged.

[0183] Figure 6 The results of the survey are shown. On the one hand, with different N h The prediction errors of the three models and N s The relationship between Figure 5 Specifically, as N s The lack of information that can be extracted from the observations will lead to a sharp increase in the reconstruction error. s What is the value of N? h =8, the model can achieve the best performance. Therefore, it can be used in N s =32 data for training h =8 model.

[0184] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation modes and equivalent substitutions based on the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A back calculation method for turbine blade cascade test boundary conditions based on sparse measurement data, characterized in that: The method comprises the following steps: Steps for obtaining the structure of a spatial analysis network based on self-attention units and Transformer encoders; The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein: The multi-head self-attention unit is an improved scaled dot-product attention: Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d k is the number of features of the query and key; superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution with a sum of 1; In the self-attention calculation, Q, K, and V are the same value; For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head: Steps for training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residuals using mean square error: Steps to obtain the structure of the spatial analysis network after training: Mathematical description of the structure of the spatial analysis network after training: Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution; A step for inputting normalized sparse measurement data into a trained spatial analysis network structure to obtain corresponding experimental boundary conditions.

2. The back-calculation method for turbine blade cascade test boundary conditions based on sparse measurement data according to claim 1 is characterized in that: The position coding module is used for performing position coding; Position encoding implies relative position encoding and absolute position encoding: Relative position encoding is used to embed position information in the spatial domain; absolute position encoding is used to embed position information in the feature domain; Relative position encoding is achieved through trigonometric matrix: Absolute position coding originates from the characteristics of binary computing: the frequency of alternation of higher-order values ​​is lower than that of lower-order values. Since the feature vector itself does not have position information in Euclidean space, the order of feature vectors is abstractly described by referring to the binary law. A multilayer perceptron is a series of stacked fully connected layers, which is essentially a continuous linear operation; the parameters of each layer of neurons include the weight matrix and the bias vector Where m is the number of neurons in the current layer, and n is the input data size; Let ζ(·) be the activation function, then the output of the neurons in the lth layer is: Select the leaky ReLU function to satisfy the mathematical relationship described below: The Transformer encoder consists of 6 identical network blocks; each network block consists of two layer regularizations, a multi-head self-attention unit, and a multi-layer perceptron; The average pooling layer is an aggregation function.

3. The back-calculation method for turbine blade cascade test boundary conditions based on sparse measurement data according to claim 2 is characterized in that: The Adam algorithm is used to train the spatial analysis network structure.

4. The back-calculation method for turbine blade cascade test boundary conditions based on sparse measurement data according to claim 2 is characterized in that: The spatial analysis network structure is based on uniform grid division of the turbine blade surface, and the sampling grid nodes are used as sensor layout locations; At least 32 sampling points are arranged on the suction side and pressure side of the blade respectively.

5. The back-calculation method for turbine blade cascade test boundary conditions based on sparse measurement data according to claim 4 is characterized in that: 128 sampling points are arranged on the suction side of the blade, and 256 sampling points are arranged on the pressure side of the blade.

6. The back-calculation method for turbine blade cascade test boundary conditions based on sparse measurement data according to claim 4, characterized in that: In the steps of training the spatial analysis network structure, optimizing the model parameters, and calculating the corresponding residual using the mean square error: Adopt a random sampling strategy with training: The sensor positions are randomly resampled in each training iteration, and each sampling process is independent of each other.

7. A back-calculation device for turbine blade cascade test boundary conditions based on sparse measurement data, characterized in that: The device comprises the following modules: A module for obtaining a spatial analysis network structure based on self-attention units and Transformer encoders; The spatial analysis network structure includes a position encoding module, a multi-layer perceptron, a Transformer encoder and an average pooling layer, wherein: The multi-head self-attention unit is an improved scaled dot-product attention: Given a series of queries Q and key-value pairs KV, the scaled dot product attention is calculated using formula (6); where d k is the number of features of the query and key; the superscript T is the matrix transpose; softmax(·) is the normalized activation function, which normalizes the calculated attention score to a probability distribution that sums to 1; In the self-attention calculation, Q, K, and V are the same value; For multi-head self-attention, multiple attention calculations are performed in parallel to capture features from different scales, and information is aggregated by connecting the calculation results, as shown in formula (7); where head i For the i-th attention head: Module used to train the spatial analysis network structure, optimize model parameters, and calculate the corresponding residual using mean square error: Module to obtain the spatial analysis network structure after training: Mathematical description of the structure of the spatial analysis network after training: Where s is the experimental observation data, is a supervised learning model, θ is a model hyperparameter, are actual boundary conditions, and the superscript * represents the optimal solution; A module used to input normalized sparse measurement data into the trained spatial analysis network structure to obtain the corresponding experimental boundary conditions.

8. A computer device comprising: A processor and a memory, characterized in that the memory is used to store executable instructions of the processor, and the processor is configured to execute the turbine blade grid test boundary condition backcalculation method based on sparse measurement data as described in any one of claims 1 to 6 by executing the executable instructions.

9. A computer storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is run, the method for backcalculating boundary conditions of a turbine blade cascade test based on sparse measurement data as described in any one of claims 1 to 6 is executed.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for backcalculating boundary conditions of a turbine blade cascade test based on sparse measurement data as described in any one of claims 1 to 6 are implemented.

Images