Reconstruction method of cascade flow field based on limited measurement points

Abstract

The present invention discloses a method for reconstructing a cascade flow field based on limited measuring points, comprising the following steps: S1, obtaining a cascade flow field model: using the first-stage stator blade of a high-pressure turbine as a geometric sample, using CFD to calculate the pressure distribution on the blade surface, changing the blade inlet angle of attack and outlet back pressure, obtaining data under different working conditions, and using CFD data as a basic database to train a deep neural network. In the present invention, by obtaining a cascade flow field model, obtaining data, and processing data, a machine is enabled to parse data, extract features, and learn rules, thereby realizing an artificial intelligence method, and then reconstructing the cascade flow field. The operation is simple, no experimental equipment is required, and the use cost is low. Using CFD data as a basic database to train a deep neural network, the trained model is fine-tuned and optimized to enhance data accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of flow field reconstruction methods, and in particular to a method for reconstructing a cascade flow field based on limited measuring points. Background Art

[0002] With the continuous development of aero-engine technology, the requirements for refined measurement of the internal flow field of the engine are becoming increasingly higher. There are usually two traditional methods for analyzing the flow field of an aero-engine. The first is experimental measurement. Although experimental measurement can measure the flow field more accurately, the experimental cost is high, the time consumption is long, and the accuracy is highly dependent on the equipment. The second is numerical simulation. The cost of numerical simulation is lower, but the accuracy of numerical simulation is lower.

[0003] In recent years, with the rapid development of computing science, it seems that making machines intelligent is no longer so far-fetched. Machine learning is a method based on big data that uses specific algorithms to allow machines to parse data, extract features, learn rules, and realize artificial intelligence. Therefore, the present invention proposes a method for reconstructing the blade flow field based on limited measuring points. Summary of the invention

[0004] In order to solve the technical problems mentioned in the above background technology, a method for reconstructing the cascade flow field based on limited measuring points is proposed.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The reconstruction method of the cascade flow field based on limited measuring points includes the following steps:

[0007] S1. Obtain the cascade flow field model:

[0008] The first-stage stator blade of a high-pressure turbine is used as a geometric sample, and the pressure distribution on the blade surface is calculated using CFD. By changing the values ​​of the blade inlet angle of attack and outlet back pressure, the deep neural network is trained using CFD data as the basic database, and the trained model is fine-tuned and optimized.

[0009] S2. Get data:

[0010] The input data is the pressure measurement point data at several blade height section positions, and the output data is the dimensionless pressure distribution on the entire blade surface, and the data is obtained through a pyramid structure; data interfaces one to four are set, and different interfaces are different numbers of cross-sectional data input ports, and losses one to three are used to match the predicted values ​​and actual values ​​of interfaces two to four, and loss four is used to match the neural network predicted flow field and the actual flow field; after the neural network training is completed, when the input data meets one of the multiple pre-set cross-sectional numbers, the neural network will automatically match the interface to obtain the corresponding flow field data;

[0011] S3. Data preprocessing:

[0012] The original data is standardized and scaled to a new distribution with a mean of 0 and a variance of 1 without changing the probability density of the original data. The calculation formula is as follows:

[0013]

[0014] S4. Data processing of cascade flow field:

[0015] The structured grid of the cascade flow field model can always be represented as a regular matrix form, and the initial flow field data Field n,j,g According to the location information Grid n,m,g Sort by grid position and convert to readable matrix format Field n,j,k,l,h , the corresponding position information is transformed into Grid n,m,k,l,h ; where k = 1, 2, 3..., K, K is the total number of nodes in the x direction, l = 1, 2, 3..., L, L is the total number of nodes in the y direction, h = 1, 2, 3..., H, H is the total number of nodes in the z direction. If the model is a two-dimensional model, then H = 1. The processed flow field data Field n,j,k,l,h Degenerates into a four-dimensional array; for all known information data Input n,i With the processed flow field data eld n,j,k,l,h Perform normalization processing;

[0016] S5. Divide the training set and the validation set

[0017] S6. Build a flow field reconstruction network

[0018] Use 2D deconvolution / 3D deconvolution to build a generator G Net, and build a generator from known data Input n,i To 2D / 3D flow field dataField n,j,k,l,h process; secondly, a discriminator DNet is constructed using two-dimensional convolution / three-dimensional convolution to distinguish the real original flow field from the reconstructed flow field generated by the generator G Net;

[0019] S7, training the GAN-based flow field reconstruction network;

[0020] S8. Flow field reconstruction of flow heat transfer model.

[0021] As a further description of the above technical solution:

[0022] In step S1: a finite number of discrete sampling points are used to describe the flow field. X sampling points are evenly arranged along the blade height line in the span direction, which is equivalent to X blade height sections; Y sampling points are evenly arranged along the curve from the leading edge point to the trailing edge point of each blade height section in the flow direction, which is equivalent to Y blade profile arc length sections. A dimensionless uniform grid of size X×Y is formed to meet the computational requirements of the neural network.

[0023] As a further description of the above technical solution:

[0024] In step S2, it also includes: obtaining initial flow field data Field n,j,g And structured grid node location information Grid n,m,g , Input n,i and Field n,j,g , Grid n,m,g One to one correspondence;

[0025] Where n=1,2,3...,N, N is the total number of measured fluid channel models, i=1,2,3...,I, I is the total number of known parameters, j=1,2,3...,J, respectively representing the grid node temperature field, pressure field and velocity field, J is the total number of recorded flow field parameters, g=1,2,3...,G, represents the grid node, G is the total number of grid nodes, m=1,2,3, respectively representing the grid node x-coordinate, y-coordinate and z-coordinate.

[0026] As a further description of the above technical solution:

[0027] Step S5 also includes the following steps: Input n,i , input n,i Randomly shuffled and divided into training set (Input n ×0.8,i)train and validation set (Input n ×0.2,i)validate, the corresponding flow field data is divided into (Fieldn×0.8,j,k,l,h)train and (Fieldn×0.2,j,k,l,h)validate, and the corresponding grid data is divided into (Gridn×0.8,m,k,l,h)train and (Gridn×0.2,m,k,l,h)validate.

[0028] As a further description of the above technical solution:

[0029] In step S6, the flow field reconstruction GAN network uses the weighted average of the binary classification loss function BCE_loss and the weighted average loss function WMSE_loss as the overall loss function of the network; the formula of the binary classification loss function BCE_loss is as follows:

[0030] BCE_loss=minGmaxDV(D,G)=

[0031] EField[logD(Field n,j,k,l,h ))]+EInput[log(1-D(G(Input n,i )))]

[0032] Among them, D(·) means that the discriminator DNet is used to discriminate the sample, G(·) means that the generator GNet is used to generate the reconstructed flow field, EField[·] and EInput[·] represent the expected values ​​of the variables in brackets; the weighted average loss function is as follows:

[0033]

[0034] Where Wi is the weight of the absolute value of the reconstructed flow field data and the original flow field data at each grid node. This weight can be appropriately increased at the boundary layer grid to improve the accuracy at the boundary layer; K×H×L is the total number of grid nodes; Field′ n,j,k,l,h To generate the reconstructed flow field data generated by the network GNet, Field n,j,k,l,h is the original real flow field data;

[0035] The total loss function of the flow field reconstruction GAN network is as follows: Loss = W1×BCE_loss+W2×WMSE_loss, where W1 is the weight of the binary classification loss function BCE_loss, and W2 is the weight of the weighted average loss function MSE_loss.

[0036] As a further description of the above technical solution:

[0037] In step S7, the following steps are also included: training starts from the DNet network, and DNet hopes to improve the ability to distinguish between the real flow field and the reconstructed flow field; following the training of the D Net network, the G Net network is trained, and at this time, it is hoped that the reconstructed flow field can deceive the D Net network so that the resolution network cannot distinguish the reconstructed flow field; the successful generation network obtained through the mutual game between the G Net generation network and the D Net resolution network is used to reconstruct the flow field.

[0038] As a further description of the above technical solution:

[0039] In step S8, the following steps are also included: according to step n,i ) test After normalization, it is passed into the generator GNet to obtain the reconstructed flow field (Field n,j,k,l,h ) test, according to the reconstructed flow field n,j,k,l,h ) test and the corresponding grid position information (Grid n,m,k,l,h ) test Calculate the thermodynamic and fluid dynamic parameters related to the cascade flow field.

[0040] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0041] 1. In the present invention, by obtaining the blade flow field model, obtaining data, and processing data, the machine can analyze data, extract features, learn rules, and realize artificial intelligence methods, thereby reconstructing the blade flow field. The operation is simple, no experimental equipment is required, and the cost of use is low. CFD data is used as the basic database to train the deep neural network, and the trained model is fine-tuned and optimized to enhance data accuracy.

[0042] 2. In the present invention, the original intention of designing the pyramid neural network is to expand the available data of the neural network model. Data with different numbers of cross sections can use the same set of neural network models to reconstruct accurate flow fields. From the results, the pyramid neural network has achieved this function. Data with different numbers of cross sections can be reconstructed into flow fields with similar accuracy through the proxy model.

[0043] 3. In the present invention, the network performance inputted by interface four is significantly degraded compared with the first three interface inputs, which means that the subsequent network matches the interface four data poorly. The problem should lie in the middle module three, which can be improved in a targeted manner. This is also the advantage of the pyramid network, which can be optimized in blocks and then matched as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of a module of a pyramid neural network provided according to an embodiment of the present invention is shown;

[0045] Figure 2 A schematic diagram of a neural network convolution according to an embodiment of the present invention is shown;

[0046] Figure 3 A schematic diagram of a transposed convolution provided according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Example

[0049] See also Figure 1-3 The present invention provides a technical solution: a method for reconstructing a cascade flow field based on limited measuring points, comprising the following steps:

[0050] S1. Obtain the cascade flow field model:

[0051] The first-stage stator blade of a high-pressure turbine is used as a geometric sample, and the pressure distribution on the blade surface is calculated using CFD. By changing the values ​​of the blade inlet angle of attack and outlet back pressure, the deep neural network is trained using CFD data as the basic database, and the trained model is fine-tuned and optimized.

[0052] S2. Get data:

[0053] The input data is the pressure measurement point data at several blade height cross-section positions, and the output data is the dimensionless pressure distribution on the entire blade surface. The data is obtained through a pyramid structure, as follows:

[0054] Set data interfaces one to four, with different interfaces being input ports for different numbers of cross-section data. Losses one to three are used to match the predicted values ​​and actual values ​​of interfaces two to four, and loss four is used to match the flow field predicted by the neural network with the actual flow field. After the neural network training is completed, when the input data meets one of the multiple pre-set cross-section numbers, the neural network will automatically match the interface to obtain the corresponding flow field data.

[0055] The original intention of designing the pyramid neural network is to expand the available data of the neural network model. Data with different numbers of sections can use the same set of neural network models to reconstruct accurate flow fields. From the results, the pyramid neural network has achieved this function. Data with different numbers of sections can reconstruct flow fields with similar accuracy through the proxy model.

[0056] S3. Data preprocessing:

[0057] The original data is standardized and scaled to a new distribution with a mean of 0 and a variance of 1 without changing the probability density of the original data. The calculation formula is as follows:

[0058]

[0059] S4. Data processing of cascade flow field:

[0060] The structured grid of the cascade flow field model can always be represented as a regular matrix form, and the initial flow field data Field n,j,g According to the location information Grid n,m,g Sort by grid position and convert to readable matrix format Field n,j,k,l,h , the corresponding position information is transformed into Gridn,m,k,l,h ; where k = 1, 2, 3..., K, K is the total number of nodes in the x direction, l = 1, 2, 3..., L, L is the total number of nodes in the y direction, h = 1, 2, 3..., H, H is the total number of nodes in the z direction. If the model is a two-dimensional model, then H = 1. The processed flow field data Field n,j,k,l,h Degenerates into a four-dimensional array; for all known information data Input n,i With the processed flow field data eld n,j,k,l,h Perform normalization processing;

[0061] S5. Divide the training set and the validation set

[0062] S6. Build a flow field reconstruction network

[0063] Use 2D deconvolution / 3D deconvolution to build a generator GNet, and build a generator from known data Input n,i To 2D / 3D flow field dataField n,j,k,l,h process; secondly, a discriminator DNet is constructed using two-dimensional convolution / three-dimensional convolution to distinguish the real original flow field from the reconstructed flow field generated by the generator GNet;

[0064] S7, training the GAN-based flow field reconstruction network;

[0065] S8. Flow field reconstruction of flow heat transfer model.

[0066] In step S1: a finite number of discrete sampling points are used to describe the flow field. X sampling points are evenly arranged along the blade height line in the span direction, which is equivalent to X blade height sections; Y sampling points are evenly arranged along the curve from the leading edge point to the trailing edge point of each blade height section in the flow direction, which is equivalent to Y blade profile arc length sections. A dimensionless uniform grid of size X×Y is formed to meet the computational requirements of the neural network.

[0067] In step S2, it also includes: obtaining initial flow field data Field n,j,g And structured grid node location information Grid n,m,g , Input n,i and Field n,j,g , Grid n,m,g One to one correspondence;

[0068] Where n=1,2,3...,N, N is the total number of measured fluid channel models, i=1,2,3...,I, I is the total number of known parameters, j=1,2,3...,J, respectively representing the grid node temperature field, pressure field and velocity field, J is the total number of recorded flow field parameters, g=1,2,3...,G, represents the grid node, G is the total number of grid nodes, m=1,2,3, respectively representing the grid node x-coordinate, y-coordinate and z-coordinate.

[0069] Step S5 also includes the following steps: Input n,i , input n,i Randomly shuffled and divided into training set (Inputn×0.8,i)train and validation set (Input n ×0.2,i)validate, the corresponding flow field data is divided into (Field n ×0.8,j,k,l,h)train and (Field n ×0.2,j,k,l,h)validate, the corresponding grid data is divided into (Grid n ×0.8,m,k,l,h)train and (Gridn×0.2,m,k,l,h)validate.

[0070] In step S6, the flow field reconstruction GAN network uses the weighted average of the binary classification loss function BCE_loss and the weighted average loss function WMSE_loss as the overall loss function of the network; the formula of the binary classification loss function BCE_loss is as follows:

[0071] BCE_loss=minGmaxDV(D,G)=

[0072] E Field [logD(Field n,j,k,l,h ))]+EInput[log(1-D(G(Input n,i )))]

[0073] Among them, D(·) means that the discriminator DNet is used to discriminate the sample, G(·) means that the generator GNet is used to generate the reconstructed flow field, EField[·] and EInput[·] represent the expected values ​​of the variables in brackets; the weighted average loss function is as follows:

[0074]

[0075] Where Wi is the weight of the absolute value of the reconstructed flow field data and the original flow field data at each grid node. This weight can be appropriately increased at the boundary layer grid to improve the accuracy at the boundary layer; K×H×L is the total number of grid nodes; Field′ n,j,k,l,h To generate the reconstructed flow field data generated by the network GNet, Field n,j,k,l,h is the original real flow field data;

[0076] The total loss function of the flow field reconstruction GAN network is as follows: Loss = W1×BCE_loss+W2×WMSE_loss, where W1 is the weight of the binary classification loss function BCE_loss, and W2 is the weight of the weighted average loss function MSE_loss.

[0077] In step S7, the following steps are also included: training starts from the DNet network, DNet hopes to improve the ability to distinguish between the real flow field and the reconstructed flow field; following the training of the DNet network, the GNet network is trained, at this time it is hoped that the reconstructed flow field can deceive the DNet network, so that the resolution network cannot distinguish the reconstructed flow field; the successful generation network obtained through the mutual game between the GNet generation network and the DNet resolution network is used to reconstruct the flow field.

[0078] In step S8, the following steps are also included: according to step n,i ) test After normalization, it is passed into the generator GNet to obtain the reconstructed flow field (Field n,j,k,l,h ) test , according to the reconstructed flow field n,j,k,l,h ) test and the corresponding grid position information (Grid n,m,k,l,h ) test Calculate the thermodynamic and fluid dynamic parameters related to the cascade flow field.

[0079] The present invention, through the steps of obtaining a blade flow field model, obtaining data, and digital processing, enables a machine to parse data, extract features, and learn rules, thereby realizing an artificial intelligence method and reconstructing the blade flow field. The operation is simple, no experimental equipment is required, and the use cost is low. CFD data is used as a basic database to train a deep neural network, and the trained model is fine-tuned and optimized to enhance data accuracy.

[0080] The data set of the basic model is provided by the CFD results. After the model is trained, the experimental data can be input and the original experimental data and the model reconstruction results can be stored in the database as new data samples. Alternatively, the CFD results corresponding to the experimental data can be used at the same time to fine-tune the existing model and achieve continuous updating.

[0081] The performance of the network inputted by interface four has been significantly degraded compared with the inputs of the first three interfaces, which means that the subsequent network matches the data of interface four poorly. The problem should lie in the middle module three, which can be improved in a targeted manner. This is also the advantage of the pyramid network, which can be optimized in blocks and then matched as a whole.

[0082] Using a deep neural network and CFD results as the basic database for training, a mapping model from a small number of measurement point data on a given section to the flow field on the entire blade surface was obtained. The model reconstructed the overall contour of the flow field well, but locally, the reconstruction results of the model with one section data as input showed a large range of blur and noise, while the reconstruction results of the model with three section data as input performed better.

[0083] By comparing the performance of models with different input interfaces of the same neural network, it is found that as the data interface gets closer to the network output layer, the number of neuron layers that the data actually passes through after entering the neural network decreases, and the accuracy improves. This shows that the more neuron layers there are, the more likely the error generated by the front neurons will be transmitted backwards with the calculation, resulting in an increase in the error of the final output.

[0084] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for reconstructing the cascade flow field based on finite measurement points, characterized in that: The following steps are involved: S1. Obtaining the cascade flow field model The first-stage stator blade of a high-pressure turbine is used as a geometric sample. The pressure distribution on the blade surface is calculated using CFD. Data under different working conditions are obtained by changing the values ​​of the blade inlet angle of attack and outlet back pressure. The deep neural network is trained using CFD data as the basic database, and the trained model is fine-tuned and optimized. S2. Obtain flow field data The input data is the pressure measurement point data at several blade height cross-section positions, and the output data is the dimensionless pressure distribution on the entire blade surface. The data is obtained through a pyramid structure, as follows: Set data interfaces one to four, with different interfaces being input ports for different numbers of cross-section data. Losses one to three are used to match the predicted values ​​and actual values ​​of interfaces two to four, and loss four is used to match the flow field predicted by the neural network with the actual flow field. After the neural network training is completed, when the input data meets one of the multiple pre-set cross-section numbers, the neural network will automatically match the interface to obtain the corresponding flow field data. S3. Data preprocessing The original data is standardized and scaled to a new distribution with a mean of 0 and a variance of 1 without changing the probability density of the original data. The calculation formula is as follows: S4. Data processing of blade flow field The structured grid of the cascade flow field model is represented as a regular matrix, and the initial flow field data Field n,j,g According to the location information Grid n,m,g Sort by grid position and convert to readable matrix format Field n,j,k,l,h , the corresponding position information is transformed into Grid n,m,k,l,h ; Where k = 1, 2, 3..., K, K is the total number of nodes in the x direction, l = 1, 2, 3..., L, L is the total number of nodes in the y direction, h = 1, 2, 3..., H, H is the total number of nodes in the z direction. If the model is a two-dimensional model, then H = 1. The processed flow field data Field n,j,k,l,h Degenerates into a four-dimensional array; for all known information data Input n,i With the processed flow field data eld n,j,k,l,h Perform normalization processing; S5, divide the training set and the validation set; S6. Build a flow field reconstruction network Use 2D deconvolution / 3D deconvolution to build a generator G Net, and build a generator from known data Input n,i To 2D / 3D flow field dataField n,j,k,l,h process; secondly, a discriminator DNet is constructed using two-dimensional convolution / three-dimensional convolution to distinguish the real original flow field from the reconstructed flow field generated by the generator G Net; S7, training the GAN-based flow field reconstruction network; S8, flow field reconstruction of flow heat transfer model; In step S2, the following steps are also included: Get the initial flow field data Field n,j,g And structured grid node location information Grid n,m,g , Input n,i and Field n,j,g , Grid n,m,g One to one correspondence; Where n=1,2,3...,N, N is the total number of measured fluid channel models, i=1,2,3...,I, I is the total number of known parameters, j=1,2,3...,J, respectively represent the grid node temperature field, pressure field and velocity field, J is the total number of recorded flow field parameters, g=1,2,3...,G, represents the grid node, G is the total number of grid nodes, m=1,2,3, respectively represent the grid node x coordinate, y coordinate, z coordinate; Step S5 also includes the following steps: Input n,i , input n,i Randomly shuffle and divide them into training set and validation set in a ratio of 4:

1. Also divide the corresponding flow field data and grid data into training set and validation set in a ratio of 4:

1. In step S6, the flow field reconstruction GAN network uses the weighted average of the binary classification loss function BCE_loss and the weighted average loss function WMSE_loss as the overall loss function of the network; the formula of the binary classification loss function BCE_loss is as follows: BCE_loss=minGmaxDV(D,G)=EField[logD(Field n,j,k,l,h ))] +EInput[log(1-D(G(Input n,i )))] Among them, D(·) means that the discriminator DNet is used to discriminate the sample, G(·) means that the generator GNet is used to generate the reconstructed flow field, EField[·] and EInput[·] represent the expected values ​​of the variables in the brackets, and the weighted average loss function is as follows: Where Wi is the weight of the absolute value of the reconstructed flow field data and the original flow field data at each grid node. The weight is appropriately increased at the boundary layer grid to improve the accuracy at the boundary layer; K×H×L is the total number of grid nodes; Field′ n,j,k,l,h To generate the reconstructed flow field data generated by the network GNet, Field n,j,k,l,h is the original real flow field data; The total loss function of the flow field reconstruction GAN network is as follows: Loss = W1×BCE_loss+W2×WMSE_loss, where W1 is the weight of the binary classification loss function BCE_loss, and W2 is the weight of the weighted average loss function MSE_loss.

2. The method for reconstructing the cascade flow field based on limited measurement points according to claim 1 is characterized in that: In step S1, the following steps are also included: using a finite number of discrete sampling points to describe the flow field, evenly arranging X sampling points along the blade height line in the span direction, which is equivalent to X blade height sections; evenly arranging Y sampling points along the curve from the leading edge point to the trailing edge point of each blade height section in the flow direction, which is equivalent to Y blade profile line arc length sections; forming a dimensionless uniform grid of size X×Y to meet the calculation requirements of the neural network.

3. The method for reconstructing the cascade flow field based on limited measurement points according to claim 2 is characterized in that: In step S7, the following steps are also included: training starts from the DNet network, DNet hopes to improve the ability to distinguish between the real flow field and the reconstructed flow field; following the training of the DNet network, the GNet network is trained, at this time it is hoped that the reconstructed flow field can deceive the DNet network, so that the resolution network cannot distinguish the reconstructed flow field; the successful generation network obtained through the mutual game between the GNet generation network and the DNet resolution network is used to reconstruct the flow field.

4. The method for reconstructing the cascade flow field based on limited measurement points according to claim 3 is characterized in that: In step S8, the following steps are also included: according to the steps, the input known information is normalized and then passed into the generator to obtain the reconstructed flow field, and the thermodynamic and fluid dynamic parameters related to the blade flow field are calculated according to the reconstructed flow field and the corresponding grid position information.

Images