A method for reconstructing physical fields on the surface of a variable input turbine

Abstract

The present invention belongs to the technical field of gas turbine turbines, and specifically designs a variable input turbine surface physical field reconstruction method based on a pyramid neural network. A data set is obtained, data in the data set is extracted and preprocessed, and the data set is divided into training data and verification data; a pyramid network is constructed based on a convolutional network; a distributed training strategy is used to train the constructed pyramid network using the training data; the performance of the trained pyramid network is verified using the verification data; the normalized experimental data is input into the verified pyramid network, and the complete blade surface physical field can be reconstructed. The present invention uses a supervised learning method to estimate the complete blade surface pressure and temperature field from sparse data obtained from a limited number of fixed measurement points distributed at a given number of blade height sections, which can improve the information density of blade grid experiments and reduce the cost of experiments and numerical simulations.

Description

Technical Field

[0001] The invention belongs to the technical field of gas turbine turbines, and specifically designs a method for reconstructing the physical field of a variable input turbine surface. Background Art

[0002] Gas turbines play a key role in aviation power, ship propulsion, ground power generation and other industrial fields due to their compact structure, high specific power and rapid start-up. The compressor, combustion chamber and gas turbine are collectively known as the three major components of a gas turbine. The turbine, as the main working component, converts the internal energy of the high-temperature and high-pressure gas at the combustion chamber outlet into kinetic energy and outputs mechanical work. The aerodynamic shape of the turbine determines the organization of the internal flow and is an important factor affecting the performance and efficiency of the gas turbine.

[0003] The technical bottleneck of the measurement method also limits the temporal and spatial resolution of the experimental data that can be obtained. Whether it is the inherent discrete measurement properties of pressure probes, thermocouples, etc., or the missing PIV data in local areas due to light source occlusion or reflection, it shows that the experimental data of turbine aerodynamics and heat transfer have typical sparse characteristics. Usually, the corresponding numerical simulation is supplemented to verify each other and obtain more detailed flow field information. However, when faced with complex blade geometry, this work often becomes complicated and time-consuming. In order to overcome these problems, many methods have been proposed for flow field reconstruction for limited measurement data, which are used as fast data processing methods to enhance experiments. In the prior art, the PIV measurement results of gas turbine combustors containing noise and missing data are processed, and the complex swirl field in the combustor is reconstructed at a data missing ratio of up to 50%. However, the original Gappy POD implies a premise that the missing positions of different samples should be different to retain sufficient flow field coherence information. When all samples have missing data at the same position, the Gappy POD will degenerate into a POD reconstruction of the flow field excluding the missing spatial position, and the missing data cannot be filled. The prior art proposes an improved GappyPOD method, which first obtains the POD orthogonal basis through a batch of complete high-fidelity data, finds the spatial position of the measurement point and extracts the corresponding basis vector elements to form a simplified basis vector, fuses the measurement data to solve the coefficient, successfully reconstructs the steam turbine casing temperature field and realizes online monitoring, with an average error of less than 4% over time. However, it should be pointed out that the method based on (Gappy)POD may face three difficulties in practical application: (1) The amount of calculation required to execute POD in a big data scenario is unbearable, and it also limits the space for its continuous update and evolution; (2) The process of extracting simplified basis vectors requires that the measurement point must be included in the data set for constructing the reference basis vector, which limits the flexibility of the measurement location; (3) The mathematical nature of the POD method is linear operation, and its accuracy will be challenged when the application involves strong nonlinear dynamics.

[0004] Machine learning methods are also widely used to improve the model's ability to handle nonlinear problems. In the prior art, deep neural networks were first used for super-resolution reconstruction of flow fields. They proposed a hybrid downsampling jump connection / multi-scale model to achieve reconstruction from low-resolution flow fields to high-resolution flow fields. They showed good performance in two typical problems, two-dimensional cylindrical wake and two-dimensional attenuated isotropic turbulence. In some cases, only 50 training data are needed to learn the model. In the prior art, a physical information embedded neural network is proposed, which embeds the control equation into the loss function, and only a small amount of observation data is needed to reconstruct the complete flow field. In the prior art, a variety of methods are also used to try to sparsely reconstruct the flow on the stall airfoil. Compared with various POD-based linear methods, shallow neural networks have shown better reconstruction effects and the results do not depend on the arrangement of the probe. In the prior art, an MLP-CNN model is developed to capture the vortex impact airfoil wake from a small amount of sensor data arranged on the airfoil surface, and the introduction of long short-term memory network assisted transfer learning enhances the model's perception of the dynamic system. Only 10 flow field snapshots are needed to complete the super-resolution reconstruction of unsteady flow. A multi-channel convolutional neural network is proposed in the prior art, which realizes the reconstruction of high-resolution flow fields of supersonic flow fields and temperature fields in scramjet engine isolators from sparse flow data. The spatial resolution is improved by up to 82 times. Compared with bicubic interpolation, the mean square error of the prediction results is reduced by 90%, and the peak signal-to-noise ratio is improved by 50%.

[0005] Although machine learning methods have become a new trend in related research fields, less attention has been paid to the physical fields of turbine surfaces. Compared with other fluid mechanics application scenarios, the super-resolution reconstruction task of turbomachinery faces three major difficulties: (1) Turbine blades usually have complex three-dimensional geometries, and their surface physical fields have typical three-dimensional characteristics, which significantly increases the difficulty of sparse reconstruction; (2) The data obtained from experiments usually only account for a very small part of the complete surface field, which leads to a large information expansion ratio; (3) The complex flow form inside the blade leads to complex pressure and temperature distribution on the blade surface, which imposes more challenges on the reconstruction task. Summary of the invention

[0006] The present invention provides a variable input turbine surface physical field reconstruction method, which uses a supervised learning method to estimate the complete blade surface pressure and temperature field from sparse data obtained from a finite number of fixed measurement points distributed at a given number of blade height sections, thereby improving the information density of blade cascade experiments and reducing the costs of experiments and numerical simulations.

[0007] The present invention is achieved through the following technical solutions:

[0008] A variable input turbine surface physical field reconstruction method based on pyramid neural network, the reconstruction method comprises the following steps:

[0009] Step 1: Obtain a data set, extract and preprocess the data in the data set, and divide the data set into training data and verification data, where the data in the data set are experimental data and corresponding CFD data;

[0010] Step 2: Build a pyramid network based on the convolutional network;

[0011] Step 3: Use the distributed training strategy to train the pyramid network constructed in step 2 using the training data in step 1;

[0012] Step 4: Use the validation data from step 1 to verify the performance of the pyramid network trained in step 3.

[0013] Step 5: Input the normalized experimental data into the pyramid network verified in step 4 to reconstruct the complete blade surface physical field.

[0014] Furthermore, the step 1 specifically includes arranging uniform sampling points on the surface of the turbine blade. The surface of the turbine blade is uniformly cut into 128 sections along the span direction, and 256 sampling points are uniformly arranged on the suction side and the pressure side along the blade arc respectively;

[0015] Starting from the pressure side trailing edge, across the leading edge line to the suction side trailing edge, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane;

[0016] Finally, each complete blade surface field is constructed into an array of shape 2×128×512 to store the pressure and temperature data of the blade surface; then, the data is normalized according to the statistical characteristics of the CFD data:

[0017] The experimental data are normalized according to the CFD statistical characteristics; the experimental and CFD data sets together form a composite data set.

[0018] Furthermore, the pyramid network constructed in step 2 includes a data alignment module, an upsampling module and a reconstruction module;

[0019] The data alignment module includes three submodules and four arrangements;

[0020] The upsampling module consists of several transposed convolutional layers;

[0021] The reconstruction module includes an input convolution layer, a residual block, a transposed convolution layer and an output convolution layer.

[0022] Further, the four arrangements include: That is, 50% leaf height, That is, 10%, 50%, 90% leaf height, i.e. 10%, 30%, 50%, 70%, 90% and That is, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%;

[0023] The three submodules are: data as input, and The data is taken as output, and the remaining sub-modules are connected in series in the same way.

[0024] Each submodule simulates an experimental setup with a specific sensor arrangement and generates data to be passed to downstream modules;

[0025] When the network receives sparse data from one of the four arrangements above, the corresponding preceding module no longer functions; the missing experimental data at other leaf height sections are created.

[0026] Furthermore, the pyramid network includes a convolution computing unit, a transposed convolution computing unit and a ResNet unit;

[0027] The output of the convolution unit of the lth layer of the convolution calculation unit can be written as the following filtering formula:

[0028]

[0029] in Represents the corresponding position element of the output matrix of the convolution unit in the lth layer; represents the corresponding position element of the matrix of the output of the l-1th layer and the input of the lth layer; w pqrk represents a convolution kernel of size P×Q×R, b k is the bias; δ(·) is the activation function.

[0030] Furthermore, the transposed convolution unit first fills the input data internally and then performs a conventional convolution operation; assuming a transposed convolution unit parameter, a convolution kernel with a size of P×Q, and a step size of [S h ,S w ] and fill [G h ,G w ], the input is a matrix of size H×W, then the output data size should satisfy:

[0031]

[0032] Furthermore, the ResNet unit introduces a direct path to simplify the identity mapping task to f'(x)=f(x)-x=0.

[0033] Furthermore, regarding the loss function, the pyramid network has a total of 4 output ports, including the local outputs of the three submodules and the global output of the network. Each output uses the mean square error to calculate the corresponding residual:

[0034]

[0035] Among them, L loss is the composite loss function, L global is the global loss function, L sub is the submodule loss function, F is the neural network, q CFD It is the CFD label data.

[0036] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the above method is implemented.

[0037] A computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0038] The beneficial effects of the present invention are:

[0039] The present invention can be applied to aeroengine and gas turbine blades, and can perform geometric super-resolution on sparse experimental data to reconstruct the complete blade surface pressure and temperature field. The method can effectively reduce the cost of obtaining high-quality data in blade cascade experiments.

[0040] On the reference data set, the present invention achieves high-precision sparse reconstruction. The reconstruction error of more than 94% of the turbine surface area is less than 1%, and the average reconstruction relative error is 0.25%, which meets the accuracy requirements of engineering applications. In addition, the present invention realizes the reconstruction of input data with different amounts of information, which improves the flexibility in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the pyramid network structure of the present invention.

[0042] Figure 2 Schematic diagram of the ResNet unit of the present invention.

[0043] Figure 3 It is a schematic diagram of the distributed training strategy of the present invention.

[0044] Figure 4 It is a data preprocessing schematic diagram of the present invention.

[0045] Figure 5 This is the network reconstruction effect diagram of this application under different inputs, where (a) The reconstruction effect diagram of the input pressure prediction, (b) The reconstruction effect diagram of the input temperature prediction, (c) The reconstruction effect diagram of the input pressure prediction, (d) Reconstructed effect diagram for input temperature prediction, (e) The reconstruction effect diagram of the input pressure prediction, (f) The reconstruction effect diagram for the input temperature prediction, (g) The reconstruction effect diagram of the input pressure prediction, (h) Reconstruct the effect graph for the input temperature prediction.

[0046] Figure 6 A schematic diagram showing the PSNR and SSIM results of the model on the entire dataset.

[0047] Figure 7 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0048] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0049] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0050] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.

[0051] The following is attached to this application specification Figure 1-7 , the technical solutions in the embodiments of the present application are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present application, but the present application may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0053] Implementation Method 1

[0054] This embodiment provides a variable input turbine surface physical field reconstruction method based on a pyramid neural network, and the reconstruction method includes the following steps:

[0055] Step 1: Obtain a data set, extract and preprocess the data in the data set, and divide the data set into training data and verification data, where the data in the data set are experimental data and corresponding CFD data;

[0056] Step 2: Build a pyramid network based on the convolutional network;

[0057] Step 3: Use the distributed training strategy to train the pyramid network constructed in step 2 using the training data in step 1; the network and its associated code are 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. Using the distributed training strategy, the network is split into 4 sub-networks, which are trained independently and then merged into a complete network after completion;

[0058] The distributed training strategy is a distributed training strategy proposed for the special structure of the pyramid network. The distributed training uses a composite database, including different types of sparse data and complete surface fields. According to the data input port, the three submodules and the fourth module consisting of the upsampling block and the reconstruction module are trained separately, and then the local networks are spliced ​​together to form a reconstruction network that is exactly the same as the overall training. Figure 3 Demonstrated the distributed training strategy process;

[0059] Step 4: Use the validation data from step 1 to verify the performance of the pyramid network trained in step 3. After the network is completed, evaluate the performance of the network on the data set using mean square error, relative error, peak signal-to-noise ratio, and structural similarity criterion.

[0060] Step 5: Simply input the normalized experimental data into the pyramid network verified in step 4 to reconstruct the complete blade surface physical field; the normalization index should adopt the statistical characteristics of the CFD data in the training set.

[0061] Furthermore, the step 1 specifically includes arranging uniform sampling points on the surface of the turbine blade. The surface of the turbine blade is uniformly cut into 128 sections along the span direction, and 256 sampling points are uniformly arranged on the suction side and the pressure side along the blade arc respectively;

[0062] Starting from the pressure side trailing edge, across the leading edge line to the suction side trailing edge, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane;

[0063] Finally, each complete blade surface field is constructed into an array of shape 2×128×512, storing the pressure and temperature data of the blade surface; Figure 4 The data extraction process is shown. Then, according to the statistical characteristics of CFD data, the data is normalized based on the following formula:

[0064]

[0065] Where μ is the mean and σ is the standard deviation; the experimental data are normalized according to the CFD statistical characteristics; the experimental and CFD data sets together constitute a composite data set.

[0066] Furthermore, the structure of the pyramid network constructed in step 2 is as follows: Figure 1 As shown, in engineering practice, depending on the experimental objectives, it may be necessary to arrange measurement points at different leaf height sections, which results in the sparse data used for reconstruction having different numbers of sections and data sizes. In order to improve the versatility of the network, a data alignment module is designed. The pyramid network includes a data alignment module, an upsampling module and a reconstruction module.

[0067] The data alignment module includes three submodules and four arrangements;

[0068] The upsampling module is composed of several transposed convolutional layers; the sampling module receives the last set of sparse data, i.e. Expand the low spatial resolution data to match the reconstructed surface field data of the target;

[0069] The reconstruction module comprises an input convolution layer, a residual block, a transposed convolution layer and an output convolution layer; the reconstruction module extracts key features from the upsampling module and predicts the reconstructed blade surface field.

[0070] Further, the four arrangements include: and

[0071] Said At 50% leaf height, 10%, 50%, 90% leaf height; At 10%, 30%, 50%, 70%, 90% leaf height, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90% leaf height;

[0072] The three submodules are: data as input, and The data is taken as output, and submodule 2 receives data as input, and The data is taken as output; submodule three receives data as input, and The data is taken as output; the rest of the sub-modules are connected in series in the same way.

[0073] Each submodule simulates an experimental setup with a specific sensor (e.g., pressure probe, thermocouple, etc.) placement and generates data to pass to downstream modules; specifically, a submodule outputs Three cross-section arrangements were simulated, namely the blade root (10% blade height), blade middle (50% blade height) and blade tip (90% blade height);

[0074] When the network receives sparse data from one of the four arrangements mentioned above, the corresponding front module no longer functions; therefore, the submodule can be regarded as a virtual experimental platform to create missing experimental data at other leaf height sections.

[0075] Furthermore, the pyramid network includes a convolution computing unit, a transposed convolution computing unit and a ResNet unit;

[0076] The convolution calculation unit is essentially a linear operation based on a set of small matrices called convolution kernels; the output of the l-th layer convolution unit can be written as the following filtering formula:

[0077]

[0078] in Represents the corresponding position element of the output matrix of the convolution unit in the lth layer; represents the corresponding position element of the l-1th layer output and the lth layer input matrix; w pqrk represents a convolution kernel of size P×Q×R, b k is the bias; δ(·) is the activation function. It is used to improve the nonlinear modeling ability of the neural network. Here, the leaky ReLU function is selected to satisfy the mathematical relationship described below:

[0079]

[0080] Where γ is the slope of the negative component and is set to 0.2.

[0081] Furthermore, the transposed convolution unit can be regarded as a special form of convolution unit, which first fills the input data internally and then performs a conventional convolution operation; in the present invention, the transposed convolution is used as an upsampling method to expand the size of the input data; assuming a transposed convolution unit parameter, a convolution kernel of size P×Q, and a step size [S h ,S w ] and fill [G h ,G w ], the input is a matrix of size H×W, then the output data size should satisfy:

[0082]

[0083] Furthermore, the ResNet unit achieves effective training of deep neural networks through a unique direct connection design; for a problem to be determined, assume that there is an optimal network architecture with L layers of neurons Able to fit the objective function; in order to better obtain the best network, a deeper network is usually set where L'>L; it can be imagined that only when the additional layers are trained as the identity mapping f(x)≡x, will eventually converge to In other words, improving the ability of network components to fit the identity mapping can effectively train deep neural networks. ResNet introduces a direct path to simplify the identity mapping task to f'(x) = f(x) - x = 0, which is easier to implement for neural networks. The basic structure of ResNet is as follows: Figure 2 shown.

[0084] Furthermore, regarding the loss function, the pyramid network has a total of 4 output ports, including the local outputs of the three submodules and the global output of the network. Each output uses the mean square error to calculate the corresponding residual:

[0085]

[0086]

[0087] Among them, L loss is the composite loss function, L global is the global loss function, L sub is the submodule loss function, F is the neural network, q CFD It is the CFD label data.

[0088] Specifically, the main goal of the present invention is to use supervised learning methods to estimate the complete blade surface pressure and temperature field from sparse data obtained from a limited number of fixed measurement points distributed at a given number of blade height sections. This can be regarded as a method of experimental data enhancement, which can improve the information density of cascade experiments and reduce the cost of experiments and numerical simulations.

[0089] based on Figure 6 As shown, the PSNR of all samples is higher than 44dB and the SSIM is higher than 0.996, which means that the reconstruction effect of the model is very close to the reference value and the reconstruction quality is very high.

[0090] The peak signal-to-noise ratio and structural similarity criterion are introduced to quantitatively evaluate the quality of flow field reconstruction. In order to be closer to the application background of these two evaluation indicators, the flow field will be normalized to [0,255] to simulate grayscale image data. The peak signal-to-noise ratio is defined as:

[0091]

[0092] The higher the PSNR, the better the reconstruction effect. The structural similarity criterion evaluates the similarity between the reconstructed data and the reference data:

[0093]

[0094] C 1 =(K 1 *255) 2 ,K 1 =0.01

[0095] C 2 =(K 2 *255) 2 ,K 2 =0.03

[0096] where μ CFD and σ CFD represents the mean and variance of CFD data, and represents the mean and variance of the reconstructed data, It represents the covariance between two sets of data, C 1 and C 2 They are all constants to ensure the stability of the calculation. SSIM∈[0,1], the higher the value, the better the flow field reconstruction effect.

[0097] Implementation Method 2

[0098] This embodiment provides a variable input turbine surface physical field reconstruction system based on a pyramid neural network. The reconstruction system uses the variable input turbine surface physical field reconstruction method based on a pyramid neural network as described in the first embodiment. The reconstruction system includes:

[0099] The data set acquisition module is used to acquire the data set, extract and preprocess the data in the data set, and divide the data set into training data and verification data, where the data in the data set are experimental data and corresponding CFD data;

[0100] Pyramid network building module, building a pyramid network based on the convolutional network;

[0101] The network training module uses a distributed training strategy to train the constructed pyramid network using training data;

[0102] Performance verification module, which uses verification data to verify the performance of the trained pyramid network;

[0103] The reconstruction module inputs the normalized experimental data into the verified pyramid network to reconstruct the complete physical field of the blade surface.

[0104] Implementation Method 3

[0105] An embodiment of the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory is used to store software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and the processor are connected via a bus. Specifically, the processor implements any step in the above-mentioned embodiment 1 by running the computer program stored in the memory.

[0106] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0107] The memory may include a read-only memory, a flash memory, and a random access memory, and provides instructions and data to the processor. A part or all of the memory may also include a nonvolatile random access memory.

[0108] It should be understood that if the above-mentioned integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in source code form, object code form, executable file or some intermediate form. The above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the above-mentioned computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0109] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0110] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0111] It should be noted that the methods and detailed examples provided in the above embodiments can be combined with the devices and equipment provided in the embodiments, and references can be made to each other, and no further details will be given.

[0112] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0113] In the embodiments provided by the present invention, it should be understood that the disclosed apparatus / terminal equipment and method can be implemented in other ways. For example, the apparatus / equipment embodiments described above are only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, other division methods can be used, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0114] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A variable input turbine surface physical field reconstruction method based on pyramid neural network, characterized in that: The reconstruction method comprises the following steps: Step 1: Obtain a data set, extract and preprocess the data in the data set, and divide the data set into training data and verification data, where the data in the data set are experimental data and corresponding CFD data; Step 2: Build a pyramid network based on the convolutional network; Step 3: Use the distributed training strategy to train the pyramid network constructed in step 2 using the training data in step 1; Step 4: Use the validation data from step 1 to verify the performance of the pyramid network trained in step 3. Step 5: Input the normalized experimental data into the pyramid network verified in step 4 to reconstruct the complete blade surface physical field; The pyramid network constructed in step 2 includes a data alignment module, an upsampling module and a reconstruction module; The data alignment module includes three submodules and four arrangements; The upsampling module consists of several transposed convolutional layers; The reconstruction module includes an input convolution layer, a residual block, a transposed convolution layer and an output convolution layer; The pyramid network includes a convolution computing unit, a transposed convolution computing unit and a ResNet unit; The output of the convolution unit of the lth layer of the convolution calculation unit is written as the following filtering formula: in Represents the corresponding position element of the output matrix of the l-th layer convolution unit; represents the corresponding position element of the l-1th layer output and the lth layer input matrix; w pqrk represents a convolution kernel of size P×Q×R, b k is the bias; δ(·) is the activation function.

2. The variable input turbine surface physical field reconstruction method according to claim 1, characterized in that: Specifically, the step 1 is to arrange uniform sampling points on the surface of the turbine blade; the surface of the turbine blade is uniformly cut into 128 sections along the span direction, and 256 sampling points are uniformly arranged on the suction side and the pressure side along the blade arc respectively; Starting from the pressure side trailing edge, across the leading edge line to the suction side trailing edge, 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 shape 2×128×512 to store the pressure and temperature data of the blade surface; then, the data is normalized according to the statistical characteristics of the CFD data: The experimental data are normalized according to the CFD statistical characteristics; the experimental and CFD data sets together form a composite data set.

3. The variable input turbine surface physical field reconstruction method according to claim 1, characterized in that: The four arrangements include, That is, 50% leaf height, That is, 10%, 50%, 90% leaf height, That is, 10%, 30%, 50%, 70%, 90% leaf height and That is, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90% leaf height; The three submodules are: data as input, and The data is taken as output, and the rest of the sub-modules are connected in series in the same way; Each submodule simulates an experimental setup with a specific sensor arrangement and generates data to be passed to downstream modules; When the network receives sparse data from one of the four arrangements above, the corresponding preceding module no longer functions; the missing experimental data at other leaf height sections are created.

4. The variable input turbine surface physical field reconstruction method according to claim 1, characterized in that: The transposed convolution calculation unit first fills the input data internally, and then performs a conventional convolution operation; assuming a transposed convolution unit parameter, a convolution kernel with a size of P×Q, and a step size [S h ,S w ] and fill [G h ,G w ], the input is a matrix of size H×W, then the output data size should satisfy:

5. The variable input turbine surface physical field reconstruction method according to claim 1, characterized in that: The ResNet unit introduces a direct path to simplify the identity mapping task to f'(x)=f(x)-x=0.

6. The variable input turbine surface physical field reconstruction method according to claim 1, characterized in that: Regarding the loss function, the pyramid network has a total of 4 output ports, including the local outputs of the three submodules and the global output of the model. Each output uses the mean square error to calculate the corresponding residual: Among them, L loss is the composite loss function, L global is the global loss function, L sub1 is the loss function of the first submodule, L sub2 is the loss function of the fourth submodule, L sub3 is the loss function of the third submodule, is a neural network, q CFD It is the CFD label data.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images