Hydraulic turbine blade topology optimization method combining doublet learning neural network and conformal mapping

Abstract

The application discloses a kind of water turbine blade topological optimization methods combining double-line learning neural network and conformal mapping.First, the water turbine blade model is mapped to two-dimensional plane by conformal mapping algorithm, generates data sample library and corresponding conformal factor.The fluid-structure interaction analysis is used to obtain the external force field and displacement field of the model nodes under specific flow field conditions.The method of the application introduces a global attention U-NET neural network, and uses the conformal factor, structure topology, external force field and displacement field as the input and output of the network, and trains the neural network.The optimization process reduces the dependence on traditional finite element analysis, improves the optimization efficiency, and realizes efficient topological optimization of the water turbine blade through the combination of offline learning and online learning, thereby improving the performance and reliability of the water turbine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fluid mechanical structure optimization, and particularly relates to a water turbine blade topology optimization method combining a double-line learning neural network and conformal mapping. BACKGROUND

[0002] As an important water energy conversion device, water turbines are widely used in hydropower stations worldwide, and their efficiency and reliability directly affect the energy conversion efficiency and economic benefits of the hydropower station. Water turbine blades, as one of the key components of water turbines, are crucial to improving the mechanical properties of water turbines. Topology optimization can significantly improve the mechanical properties of blades, however, traditional topology optimization methods for blades usually rely on empirical design and finite element analysis, which are time-consuming, costly, and difficult to adapt to complex working conditions and variable design requirements. Therefore, it is necessary to design a water turbine blade optimization method that reduces reliance on finite element processes and has general applicability.

[0003] The information disclosed in this BACKGROUND section is only for the purpose of increasing the understanding of the background of the present application and should not be taken as an acknowledgment or any form of suggestion that this information forms prior art with respect to the present application. SUMMARY

[0004] The present application aims to provide a water turbine blade topology optimization method combining a double-line learning neural network and conformal mapping, thereby overcoming the shortcomings of existing water turbine blade topology optimization methods, such as time-consuming, high cost, and low efficiency.

[0005] To achieve the above-mentioned purpose, the present application provides a water turbine blade topology optimization method combining a double-line learning neural network and conformal mapping, comprising the following steps:

[0006] (1) Create a data sample library composed of water turbine blade models, and use a conformal mapping algorithm for each sample to map it to a two-dimensional plane, thereby obtaining a mapped two-dimensional grid and the conformal factor corresponding to the grid nodes;

[0007] (2) Perform fluid-structure coupling analysis on each sample under different structural topologies to obtain the external force field and displacement field of the model nodes under specific flow field conditions;

[0008] (3) Replace the smooth transition layer of the U—NET neural network with a global attention layer to obtain a global attention U—NET neural network;

[0009] (4) taking the conformal factor and the structural topology of each sample as inputs of the network, and taking the external force field and the displacement field of the model node generated in step (2) as outputs of the network, integrating the inputs and the outputs of the network to form a single training sample matrix;

[0010] (5) integrating all the training sample matrices obtained in step (4) to serve as an offline learning data set for training the global attention U—NET neural network in step (3);

[0011] (6) processing a water turbine blade model to be optimized through a common mapping algorithm to obtain a mapped grid and a conformal factor, and on this basis, completing initialization of the level set topology optimization method and performing multiple optimization iterations;

[0012] (7) integrating the external force field, the displacement field, the structural topology and the conformal factor generated in step (6), and referring to the online learning data set obtained in step (5);

[0013] (8) retraining the global attention U—NET neural network trained in step (5) through the online learning data set;

[0014] (9) replacing the fluid-structure coupling analysis step of the level set method with the global attention U—NET neural network retrained in step (8) to generate a level set method combined with double-line learning;

[0015] (10) using the level set method combined with double-line learning obtained in step (9) for multiple subsequent optimization iterations of the model selected in step (5).

[0016] Preferably, in the above technical solution, the common mapping algorithm comprises the following steps:

[0017] 1) using a RICCI flow algorithm for discrete triangular grids, and simultaneously mapping to a bounded rectangular plane in two dimensions;

[0018] 2) finding the largest inscribed rectangle in the surface of the water turbine blade, and setting the four nodes of the rectangle as fixed points for mapping;

[0019] 3) converting the obtained conformal factor into a matrix consistent in size with the two-dimensional grid through a barycentric interpolation algorithm.

[0020] Preferably, in the above technical solution, the external force field and the displacement field generated in step (2) are represented in the form of a matrix, and the size of the matrix is consistent with the matrix in step 3).

[0021] Preferably, in the above technical solution, the global attention layer in step (3) is a network architecture in which matrices are respectively convolved for each channel and finally stacked and restored.

[0022] Preferably, in the above technical solution, the external force field and the displacement field in step (4) are unified with the naming of the conformal factor obtained in step (1), so that the node factor on the grid node corresponds to the external force and the displacement.

[0023] Preferably, in the above technical solution, the offline learning data set in step (5) is divided into two parts, one part is used to train the global attention U—NET neural network, and the other part is used to verify the global attention U—NET neural network.

[0024] Preferably, in the above technical solution, the training of the global attention U—NET neural network in step (5) includes: continuously inputting the data set into the global attention U—NET neural network according to batch conditions, and adaptively adjusting the learning rate parameter of the global attention U—NET neural network until the loss function value calculated by the cross-entropy loss function converges.

[0025] Preferably, in the above technical solution, the level set topology optimization method used in step (6) is an extended level set method combined with conformal factors, which drives the evolution of the water turbine blade topology according to the fluid-structure coupling analysis results.

[0026] Preferably, in the above technical solution, the multiple optimization iterations of the model are 10-30 optimization iterations.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] (1) The present application combines a double-line learning neural network and a water turbine blade topology optimization method based on conformal mapping. First, a data sample library composed of water turbine blade models is created, and each sample is mapped to a two-dimensional plane using a conformal mapping algorithm to obtain a mapped two-dimensional grid and corresponding conformal factors. Subsequently, fluid-structure coupling analysis is performed on each sample to obtain the external force field and displacement field of the model nodes under specific flow field conditions. By replacing the smooth transition layer of the U—NET neural network with a global attention layer, a global attention U—NET neural network is obtained to improve the accuracy of prediction.

[0029] (2) The method of the present application not only includes the step of preprocessing the water turbine blade model using the conformal mapping algorithm, but also involves offline and online training of the model using the global attention U—NET neural network, ultimately realizing the topology optimization of the water turbine blade. This method can effectively reduce the finite element analysis process relied on by traditional topology optimization, improve the optimization efficiency, and at the same time maintain or improve the performance of the water turbine blade.

[0030] (3) Through the implementation of the present application, the mechanical properties of the water turbine blade can be significantly improved, the structural design thereof can be optimized, and the overall performance and reliability of the water turbine can be further improved, thereby providing technical support for improving the energy conversion efficiency and economic benefits of the hydropower station. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a step diagram of a water turbine blade topology optimization method combining a double-line learning neural network and conformal mapping according to the present application;

[0032] Figure 2 is a diagram showing the change in the blade grid before and after mapping in the method according to the present application.

[0033] Figure 3 is a schematic diagram of the training of an offline network in the method according to the present application;

[0034] Figure 4 is a schematic diagram of the structure of a global attention U-NET neural network in the method according to the present application.

[0035] Figure 5 is a schematic diagram of the detailed structure of a global attention layer in the method according to the present application.

[0036] Figure 6 is a schematic diagram of the iteration process of an online learning neural network for the first 20 times in the method according to the present application.

[0037] Figure 7 is a schematic diagram of the iteration process of an online learning neural network after 20 times in the method according to the present application. DETAILED DESCRIPTION

[0038] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings, but it should be understood that the scope of protection of the present application is not limited by the specific embodiments.

[0039] Unless otherwise explicitly stated, throughout the specification and claims, the term "comprise" or variations such as "comprises" or "comprising" will be understood to imply the inclusion of a stated element or component but not the exclusion of any other element or component.

[0040] As shown in Figures 1 to 7 , a water turbine blade topology optimization method combining a double-line learning neural network and conformal mapping according to the specific embodiments of the present application comprises the following steps:

[0041] S1: Create a data sample library composed of water turbine blade models. As Figure 2 , use a conformal mapping algorithm for each sample to map the sample to a two-dimensional plane, thereby obtaining a mapped two-dimensional grid and conformal factors corresponding to the grid nodes.

[0042] In this embodiment, 50 water turbine blade models with different shapes are collected, and 60 different structural topologies are set for each model. The sample individuals are defined based on the shape and structural topology. The 3000 samples form a sample database. Meanwhile, the conformal mapping algorithm shown in step S1 can be divided into:

[0043] 1) The conformal mapping method used in step S1 is the RICCI flow algorithm for discrete triangular meshes. Meanwhile, the mapped two-dimensional plane is a bounded rectangular plane.

[0044] 2) Find the largest inscribed rectangle in the surface of the water turbine blade. The four nodes of the rectangle are set as fixed points for mapping.

[0045] 3) The obtained conformal factor is converted into a matrix consistent in size with the two-dimensional grid by the barycentric interpolation algorithm. In this embodiment, the two-dimensional grid size is 150x150.

[0046] Step S1 realizes the purpose of reducing the dimension of the water turbine blade model to a two-dimensional plane, and the conformal factor representing the original surface characteristics of the model is collected for subsequent processing.

[0047] S2: Perform fluid-structure interaction analysis on each sample to obtain the external force field and displacement field of the model nodes under specific flow field conditions.

[0048] In the topology optimization of the water turbine blade, the finite element analysis is a fluid-structure interaction analysis. In this embodiment, all samples are placed in a uniform incompressible flow field with a flow rate of 2 m / S. The fluid-structure interaction analysis is performed on each data sample of step S1 to obtain the external force field and displacement field generated by each blade in the given flow field. These two field data will provide a basis for the subsequent optimization step.

[0049] S3: Replace the smooth transition layer of the U—NET neural network with a global attention layer to obtain a global attention U—NET neural network.

[0050] The network architecture used in the present application is improved based on the U—NET neural network. As shown in Figure 3 The global attention layer is a network architecture that allows matrices to perform convolution operations along different dimensions and then restores them. Compared with the smooth transition layer, the global attention layer performs convolution processing in different dimensions, so it has better analytical ability for local information of matrices and is more suitable for topology optimization problems.

[0051] S4: taking the conformal factor and the structure topology of each sample in step S1 as the input of the network, and taking the external force field and the displacement field of the model node generated in step S2 as the output of the network, integrating the input and the output of the network to form a single training sample matrix.

[0052] The finite element analysis in topology optimization is mainly used for calculating the external force field and the displacement field of the model under the current structure topology. Therefore, the neural network used in the present application can predict the corresponding external force field and displacement field according to the conformal factor and the current structure topology of the model. Therefore, in the present embodiment, the conformal factor and the structure topology of each sample in step S1 are taken as the input of the network; the external force field and the displacement field of the model node generated in step S2 are taken as the output of the network.

[0053] S5: integrating all the training sample matrices obtained in step S4 to serve as the offline learning dataset for training the global attention U-NET neural network in step S3.

[0054] The double-line neural network proposed in the present application is a neural network trained through the two processes of online learning and offline learning. The offline learning process is as shown in steps S1 to S5, as shown in the figure. Figure 4 The purpose of offline learning is to enable the neural network to learn the relationship between the two input factors of the conformal factor and the structure topology of the water turbine blade and the two output factors of the external force field and the displacement field through a wide range of data samples. At the same time, in order to facilitate data processing, the input data and the output data in step S4 are both matrices of the same size.

[0055] The optimizer for training the neural network is Adam; the proportion of the training set and the test set in the offline learning dataset is 7:3; and the loss function used in the training is the cross-entropy loss function.

[0056] S6: processing the water turbine blade model to be optimized through the conformal mapping algorithm in step S1 to obtain the mapped grid and the conformal factor, and completing the initialization of the level set topology optimization method and performing 20 optimization iterations on this basis.

[0057] The level set method used in step S6 is the extended level set method combined with conformal mapping, which is suitable for optimization objects such as water turbine blades with complex surfaces.

[0058] S7: integrating the external force field, the displacement field, the structure topology and the conformal factor generated in step S6, and referring to the online learning dataset obtained in step S5.

[0059] In order to ensure the portability of the global attention U-NET neural network, the offline learning and the online learning are consistent in the form of learning, and the only difference lies in the different training data.

[0060] S8: The globally attentive U-NET neural network trained in step S5 is further trained by the online learning dataset.

[0061] The procedure of online learning is shown in steps S6-S8 as Figure 5 The purpose of online learning is to let the neural network learn the relationship between the structure topology of the current blade to be optimized and the external force field and displacement field. The combination of offline learning and online learning can reduce the network training burden for a specific hydraulic turbine blade model while ensuring the accuracy of the grid prediction.

[0062] S9: The globally attentive U-NET neural network trained by online learning in step S8 is used to replace the fluid-structure coupling analysis step of the level set method to generate a level set method combined with double-line learning.

[0063] The overall flowchart of double-line learning based on neural network and conformal mapping is shown in Figure 6 The globally attentive U-NET neural network trained by offline learning and online learning can accurately predict the corresponding external force field and displacement field according to the structure topology of the model to be optimized in step S6.

[0064] S10: The level set method combined with double-line learning obtained in step S9 is used for the optimization iteration after 20 times of the model selected in step S5.

[0065] The foregoing description of specific exemplary embodiments of the application is intended to be illustrative only and is not intended to limit the application to the precise forms described. Many modifications and variations are possible in light of the above teachings without departing from the spirit or essential characteristics of the application. The exemplary embodiments were chosen and described in order to explain the principles of the application and its practical application and to allow others skilled in the art to understand the application for various exemplary embodiments with various modifications being suitable. The scope of the application is intended to be defined by the claims and their equivalents.

Claims

1. A method for topology optimization of hydraulic turbine blades combining doublet learning neural network with conformal mapping, characterized in that, The method comprises the following steps: (1) creating a data sample library composed of water turbine blade models, using conformal mapping algorithm to map each sample to a two-dimensional plane, obtaining the mapped two-dimensional grid and the conformal factor corresponding to the grid nodes; (2) performing fluid-structure coupling analysis on each sample under different structural topologies to obtain the external force field and displacement field of the model nodes under specific flow field conditions; (3) replacing the smooth transition layer of the U—NET neural network with a global attention layer to obtain a global attention U—NET neural network; (4) taking the conformal factor in step (1) and the structural topology as the input of the network, and taking the external force field and displacement field of the model nodes generated in step (2) as the output of the network, integrating the input and output of the network to form a separate training sample matrix; (5) integrating each training sample matrix obtained in step (4) as an offline learning data set to train the global attention U—NET neural network in step (3); (6) processing the water turbine blade model to be optimized by the conformal mapping algorithm to obtain the mapped grid and conformal factor, and initializing the level set topology optimization method based on this, and performing multiple optimization iterations; (7) integrating the external force field, displacement field, structural topology and conformal factor generated in step (6), and referring to the online learning data set obtained in step (5); (8) training the global attention U—NET neural network trained in step (5) through online learning data set for secondary training; (9) replacing the fluid-structure coupling analysis step of the level set method with the global attention U—NET neural network trained in step (8) to generate a level set method combined with double-line learning; (10) using the level set method combined with double-line learning obtained in step (9) for multiple subsequent optimization iterations of the model selected in step (5).

2. The method of claim 1, wherein the method is characterized by, The conformal mapping algorithm comprises the following steps: 1) using RICCI flow algorithm for discrete triangular grid, and the two-dimensional plane mapped to is a bounded rectangular plane; 2) finding the largest inscribed rectangle in the surface of the water turbine blade, and setting the four nodes of the rectangle as fixed points of mapping; 3) converting the obtained conformal factor into a matrix consistent in size with the two-dimensional grid by barycentric interpolation algorithm.

3. The method of claim 1, wherein the method is characterized by, The external force field and displacement field generated in step (2) are represented in the form of a matrix, and the size of the matrix is consistent with the matrix in step 3.

4. The method of claim 1, wherein, The global attention layer in step (3) is a network architecture that convolves the matrix for each channel and finally stacks and restores it.

5. The method of claim 1, wherein, In step (4), the external force field, displacement field and conformal factor obtained in step (1) are uniformly named to realize the correspondence of the node factor on the grid node with the external force and displacement.

6. The method of claim 1, wherein, The offline learning data set in step (5) is divided into two parts, one part is used to train the global attention U—NET neural network, and the other part is used to verify the global attention U—NET neural network.

7. The method of claim 1, wherein, The training of the global attention U-Net neural network in step (5) comprises: continuously inputting the data set into the global attention U-Net neural network according to batch conditions, and adaptively adjusting the learning rate parameter of the global attention U-Net neural network until the loss function value calculated by using the cross-entropy loss function converges.

8. The method of claim 1, wherein, The level set topology optimization method used in step (6) is an extended level set method combined with a conformal factor, which drives the evolution of the water turbine blade topology according to the fluid-structure coupling analysis result.

9. The method of claim 1, wherein, The number of optimization iterations of the model is 10-30 optimization iterations.

Images