Training method of three-dimensional flow field prediction system of underwater vehicle and application of training method

Abstract

The invention belongs to the technical field related to deep learning, and discloses a training method and application of a three-dimensional flow field prediction system of an underwater vehicle, and the training method comprises the steps: calculating the mass center of a neighbor point set for each surface grid point of a vehicle model, constructing a covariance matrix between the neighbor point set and the centroid of the neighbor point set, and performing eigenvalue decomposition to obtain a normal vector of the point; calculating a normal vector included angle between the surface grid point and each point in the neighbor point set, taking the obtained maximum included angle as the geometric feature measurement of the point, and converting the geometric feature measurement into a weight through a Sigmoid function, thereby obtaining the sampling probability of each point; randomly extracting surface grid points from the original surface grid of the aircraft based on the obtained sampling probability to obtain a point cloud of a corresponding model; and training a point cloud neural network by using the sampled point cloud data to obtain the three-dimensional flow field prediction system of the underwater vehicle. Based on the method, the prediction precision and the flow field detail recovery capability can be improved while the calculation efficiency is maintained.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field related to deep learning, and more particularly, to a training method of a three-dimensional flow field prediction system of an underwater vehicle and application thereof. BACKGROUND

[0002] In the optimization design process of modern underwater vehicles, obtaining accurate three-dimensional flow field distribution is of great significance for drag reduction, noise reduction and shape optimization. Currently, the main ways to obtain high-fidelity flow field data include experimental fluid dynamics (EFD) and computational fluid dynamics (CFD). Although the experimental fluid dynamics method can provide accurate measurement results, it has a long test period, high cost and limited measurable range, making it difficult to meet the needs of early design stages and multi-scheme comparative analysis. The computational fluid dynamics method can accurately obtain flow field information by numerically solving the Navier-Stokes equation and has strong flexibility, but in large-scale parameter space exploration and multi-scheme optimization, it is limited by complex grid partitioning, long computation time and other problems, and has low computational efficiency.

[0003] Based on the above problems, in recent years, data-driven surrogate modeling methods have gradually become a research hotspot. For example, support vector machines, Kriging methods, polynomial response surfaces and other models can replace CFD to some extent, reducing computational cost. However, these traditional surrogate modeling methods often have insufficient accuracy and difficulty in capturing local flow characteristics when dealing with high-dimensional, nonlinear and complex geometric conditions. At the same time, convolutional neural networks (CNN) have also been introduced into the field of flow field prediction, which has improved modeling capabilities to some extent. However, due to its dependence on regular grid input, it must perform interpolation on irregular grid data, which can cause information loss in sensitive areas such as boundary layers, limiting its applicability in complex geometric scenarios.

[0004] Point cloud neural networks can directly process unstructured grid data, providing a new direction for efficient prediction of complex flow fields. However, existing point cloud methods generally use uniform or random sampling strategies, resulting in insufficient sampling density in key areas with high curvature or pressure gradient discontinuity, leading to large local prediction errors and limited ability to capture flow field details.

[0005] Therefore, how to improve the accuracy of three-dimensional flow field prediction while maintaining computational efficiency is a technical problem that needs to be solved at present. SUMMARY

[0006] In view of the above defects or improvement needs of the prior art, the present application provides a training method of a three-dimensional flow field prediction system of an underwater vehicle and application thereof, which aims to improve the accuracy of three-dimensional flow field prediction while maintaining computational efficiency.

[0007] To achieve the above objectives, the following technical solution is proposed.

[0008] According to a first aspect of the present invention, a training method for a three-dimensional flow field prediction system for an underwater vehicle is provided, comprising: Obtain the original dataset, each set of original data containing different three-dimensional geometric models of the vehicle and the flow field physical quantities of each grid point in the surface grid of the vehicle model obtained by computational fluid dynamics simulation method; For each surface grid point of the aircraft model , For the i-th surface mesh point, extract the K nearest mesh points from the surface mesh to form this point. nearest neighbor set K is a preset number of points; calculate the centroid of the nearest neighbor set. Construct a set of nearest neighbors Its center of mass The covariance matrix between the two is analyzed, and eigenvalue decomposition is performed. The eigenvector corresponding to the smallest eigenvalue in the decomposition is taken as the point. normal vector ; Calculation points With each point Angle between the normal vectors , For point set Take the j-th point in the equation and obtain the maximum included angle. As a point Geometric feature measurement ,right Perform Sigmoid function calculation to obtain the point weight , Calculate the weights to find the minimum geometric feature measure of all surface grid points. The ratio of the sum of the weights of all surface mesh points is used to obtain the point. sampling probability ; Based on the obtained sampling probability, surface mesh points are randomly extracted from the original surface mesh of the three-dimensional geometric model of the aircraft to obtain the point cloud of the corresponding model. Using the coordinates of each point in the point cloud as input and the flow field physical quantities of each point as output labels, a point cloud neural network is trained to obtain a three-dimensional flow field prediction system for underwater vehicles.

[0009] According to a second aspect of the present invention, a method for optimizing a three-dimensional geometric model of an underwater vehicle is provided, comprising: Based on the set evaluation objective function, the surface point cloud of the three-dimensional geometric model is taken as the optimization variable, and an optimization algorithm is used to search for optimization in the preset variable space range, so as to obtain the optimal three-dimensional geometric model. The evaluation objective function is used to calculate the total resistance and average adverse pressure gradient of the corresponding model based on the flow field physical quantity of each point of the three-dimensional geometric model, and the flow field physical quantity of each point of the three-dimensional geometric model is obtained by inputting the surface point cloud coordinate combination of the corresponding three-dimensional geometric model into the underwater vehicle three-dimensional flow field prediction system.

[0010] According to a third aspect of the present application, an electronic device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method according to any one of the above aspects when executing the computer program.

[0011] According to a fourth aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method according to any one of the above aspects.

[0012] According to a fifth aspect of the present application, a computer program product is provided, which comprises a computer program or instructions, and the computer program or instructions are executed by a processor to implement the steps of the method according to any one of the above aspects.

[0013] Overall, compared with the prior art, the above technical solutions conceived by the present application mainly have the following beneficial effects: 1. The training method of the underwater vehicle three-dimensional flow field prediction system disclosed in the present application, after obtaining the original data, for each surface grid point, the normal vector thereof is estimated based on the near neighbor point set, and then the geometric feature metric of each surface grid point is obtained by calculating the included angle of the normal vectors, the geometric feature metric of the surface grid point reflects the curvature of the local surface shape, in the area with large curvature, the normal vectors of adjacent points deviate more obviously, so the maximum normal vector included angle value is larger, and the change of the curvature distribution is usually related to the change of the pressure gradient in the flow field, the high curvature area corresponds to the area with sudden change of the pressure gradient, then the Sigmoid function is introduced, the Sigmoid function can perform nonlinear mapping on the geometric feature metric, compared with linear mapping, the high curvature and low curvature points can be more obviously distinguished, so that the points with larger feature values obtain higher values after mapping, thereby enhancing the weight of the key area in sampling, so the geometric feature metric can be converted into a weight, and finally the sampling probability of each surface grid point can be calculated based on the weight of each surface grid point, based on the above algorithm, the sampling probability of each surface grid point can be dynamically adjusted according to the curvature distribution of the model, adaptive sampling based on the geometric curvature feature is realized, so that the sampling in the high curvature area is increased, and the local prediction error is significantly reduced. The sampling mechanism is combined with the point cloud neural network structure, so that the model can directly process unstructured grid point cloud data, while maintaining the computing efficiency, the prediction accuracy and the flow field detail recovery ability are improved.

[0014] 2. The three-dimensional geometric model optimization method of the underwater vehicle disclosed in the present application, since the trained underwater vehicle three-dimensional flow field prediction system is used to directly predict the flow field physical quantity corresponding to different coordinate combinations, the CFD simulation with huge calculation amount is avoided, and the optimization efficiency is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a step flow chart of the training method of the underwater vehicle three-dimensional flow field prediction system in an embodiment of the present application; Figure 2 is a structural schematic diagram of the SUBOFF vehicle main hull in an embodiment; Figure 3 is a size schematic diagram of the CFD design domain in an embodiment of the present application; Figure 4 is a schematic diagram of the CFD unstructured grid division based on the design domain determined in an embodiment of the present application; Figure 3 Figure 5 is an architecture schematic diagram of the cloud neural network in an embodiment of the present application; Figure 6 is a point cloud schematic diagram obtained by uniform sampling in an embodiment; Figure 7 ​is a schematic diagram of a point cloud obtained by adaptive sampling in an embodiment of the present application; Figure 8 is a pressure coefficient prediction effect diagram in an embodiment of the present application; Figure 9 is a shear stress coefficient prediction effect diagram in an embodiment of the present application; Figure 10 is a pressure gradient prediction effect diagram in an embodiment of the present application; Figure 11 is a comparison diagram of geometric models before and after optimization in an embodiment of the present application; Figure 12 is a comparison diagram of pressure coefficient curves before and after optimization in an embodiment of the present application; Figure 13 is a comparison diagram of shear stress curves before and after optimization in an embodiment of the present application; Figure 14 is a comparison diagram of pressure gradient curves before and after optimization in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0017] In a first aspect, the present application claims a training method of an underwater vehicle three-dimensional flow field prediction system.

[0018] As Figure 1 shown is a step flow chart of the training method of the underwater vehicle three-dimensional flow field prediction system in an embodiment of the present application. The step order in the chart is only one possible way provided, but is not limited thereto. The method will be described in detail below in combination with Figure 1 .

[0019] S1, obtain an original data set, each group of original data containing different three-dimensional geometric models of a vehicle and flow field physical quantities of each grid point in a vehicle model surface grid obtained by a computational fluid dynamics simulation method.

[0020] Specifically, before obtaining the original data, a reference model of the vehicle is determined according to the design requirements, and the parameters of the vehicle can be adjusted within a certain range based on the reference model, and the vehicle model with the best flow field physical quantity is obtained through an optimization algorithm. In the present application, in order to quickly obtain the flow field physical quantity corresponding to different vehicle models, a three-dimensional flow field prediction system of the underwater vehicle is trained, and the trained prediction system can directly generate the corresponding flow field physical quantity according to the input vehicle model point element, thereby accelerating the optimization process.

[0021] Specifically, based on the reference model, the vehicle is adjusted to obtain different three-dimensional geometric models of the vehicle, and the computational fluid dynamics simulation method (CFD) is used to simulate the flow field of each vehicle model, and the flow field physical quantity of each grid point in the surface grid of each vehicle model can be obtained. Each three-dimensional geometric model and the flow field physical quantity obtained by simulation thereof constitute a set of original data, and multiple sets of original data constitute an original data set.

[0022] In an embodiment, the surface profile of the vehicle can be generated by a Bezier curve to obtain a three-dimensional geometric model of the corresponding vehicle. As shown in Figure 2 The structure of the SUBOFF vehicle main hull is shown in the schematic diagram, which includes an inlet section, a straight section, an outlet section and a tail section. An 8-order Bezier curve can be used to parameterize the SUBOFF vehicle main hull (for models with more complex shape characteristics or local detail requirements, a higher-order Bezier curve can be used), and the geometric shape is adjusted by controlling the profile control points of the inlet section and the outlet section. The sampling range of the vertical coordinates of the control points of the inlet section is [0.9, 1.2] times of the baseline value, which is used to adjust the transition of the forebody streamline; the sampling range of the vertical coordinates of the control points of the outlet section is [0.7, 1.3], which is used to control the contraction gradient of the tail. Latin hypercube sampling (LHS) is used to generate 1000 sets of parameter combinations in an 8-dimensional design space (1000 sets are only examples, and the actual number can be expanded according to requirements), and the 1000 sets of parameter combinations correspond to 1000 different three-dimensional geometric model shapes, which ensures uniform coverage of the design space, and automatically generates three-dimensional geometric models in batches. It should be noted that the Bezier curve is only an example of a parameterization method, and the present application is not limited thereto.

[0023] For each aircraft model, computational fluid dynamics (CFD) simulation can be used to simulate the flow field and obtain the flow field physical quantities at each grid point on the surface mesh of each aircraft model. For example, Fluent Meshing (this method is not limited to Fluent Meshing; other commercial or open-source mesh generation software with scripting capabilities can also be used). This method automatically meshes the aircraft boundary layer and its surrounding region (design domain), generating an unstructured mesh containing approximately 3 million elements. Fifteen boundary layer meshes are set on the hull surface, with a growth rate of 1.2 (the number of mesh elements and the boundary layer mesh settings can be adjusted according to computational requirements and geometric complexity) to meet the requirement of the turbulence model y+≈1. Figure 3 The diagram shows the dimensions of the CFD design domain in one embodiment of the present invention. The computational domain is set to meet the standard that the distance from the front end of the hull to the inlet is 8D, the distance from the rear end to the outlet is 10D, and the distance from the lateral boundary to the centerline is 6D, where D is the maximum diameter of the hull. A cylindrical enrichment zone with a diameter of 2D is set around the hull to capture near-field flow. Figure 4 The figure shown is based on an embodiment of the present invention. Figure 3 A schematic diagram of the unstructured CFD mesh generation for the defined design domain. The CFD solution employs the SIMPLEC pressure-based coupled algorithm and the SST k-ω turbulence model, with boundary conditions set as velocity inlet, pressure outlet, and no-slip wall. Water was used as the working medium (ρ = 998.2 kg / m³, μ = 0.001003 Pa·s), with an inflow velocity of 6.1728 m / s (operating condition). The Reynolds number was calculated based on the hull length. Steady-state calculations were performed for 2000 iterations. After convergence, the flow field physical quantities on the hull surface were output, including key parameters such as pressure coefficient, shear stress coefficient, and pressure gradient distribution data.

[0024] S2, for each surface grid point of the aircraft model , For the i-th surface mesh point, extract the K nearest mesh points from the surface mesh to form this point. nearest neighbor set Calculate the centroid of the nearest neighbor set. Construct a set of nearest neighbors Its center of mass The covariance matrix between the two is analyzed, and eigenvalue decomposition is performed. The eigenvector corresponding to the smallest eigenvalue in the decomposition is taken as the point. normal vector .

[0025] Specifically, this step involves calculating the normal vector of each surface grid point on the model surface, thereby providing data for subsequent steps to dynamically allocate sampling probabilities by measuring surface curvature characteristics.

[0026] For each surface grid point , the K nearest grid points in the surface grid are extracted to form the nearest neighbor set of the point . , represents the three-dimensional coordinate vector of the jth neighbor point in the corresponding nearest neighbor set . After determining the nearest neighbor set , the three-dimensional coordinate vector of the centroid of the set is calculated as . .

[0027] Subsequently, based on the nearest neighbor set and its centroid, the covariance matrix is constructed, and its expression is .

[0028] The above covariance matrix is a 3*3 dimensional matrix.

[0029] Finally, the eigenvalue decomposition of the matrix is performed, and the eigenvalue decomposition, also known as spectral decomposition, is to decompose the matrix into a product of a set of eigenvalues and eigenvectors. The eigenvector corresponding to the smallest eigenvalue is selected as the normal vector of the surface grid point .

[0030] For each grid point in the surface grid, the above processing is performed to obtain the normal vector of each surface grid point.

[0031] S3, the normal vector angle between the point and each point , is calculated, where j is the jth point in the point set . The maximum angle obtained is taken as the geometric feature measure of the point . The Sigmoid function is calculated for to obtain the weight of the point , where min is the minimum value of the geometric feature measure of all surface grid points. The weight is calculated as the ratio of the weight of the point to the sum of the weights of all surface grid points, and the sampling probability of the point is obtained.

[0032] Specifically, this step aims to dynamically allocate the sampling probability by measuring the surface curvature feature.

[0033] For each surface grid point and its neighboring point set After determining the normal vector of each point, the included angle between the normal vector of each point in the point set and the normal vector of point can be calculated The calculation formula is: .

[0034] In the formula, is the normal vector of point , and is the normal vector of point .

[0035] Subsequently, the included angles , , are compared, and the maximum included angle is selected as the geometric feature metric of point , which can be expressed as: .

[0036] After obtaining the geometric feature metric of each surface grid point , the minimum geometric feature metric is determined by comparison, and the curvature geometric feature quantity is converted into a weight by a Sigmoid function, and the calculation formula is: ; In the formula, k is a preset scaling factor, which can be directly taken as 1, for example.

[0037] After obtaining the weight of each surface grid point , the ratio of the weight to the sum of the weights of all surface grid points is calculated to convert the weight into a sampling probability , and the calculation formula is: ; In the formula, is the weight of the nth surface grid point, and N is the number of surface grid points.

[0038] S4, based on the obtained sampling probability, randomly sampling surface grid points from the original surface grid of the three-dimensional geometric model of the aircraft to obtain a point cloud corresponding to the model.

[0039] ​​​Specifically, for each vehicle model, surface grid points can be randomly extracted from the original surface grid of the vehicle model based on the sampling probability distribution of each surface grid point, to obtain the point cloud of the corresponding model. For example, 5000 points are randomly extracted from the original grid based on the sampling probability distribution to form a point cloud. Since the sampling probability is based on the curvature change of the model, sampling based on the sampling probability can achieve dense sampling in high curvature areas and sparse sampling in flat areas. Each sampling point contains three-dimensional coordinates ( , , ) and corresponding physical quantities ( , , ).

[0040] For each set of original data in the original data, the above method is used for processing to obtain the model point cloud and the flow field physical quantity corresponding to the point cloud, thereby forming the training data.

[0041] In the present application, by processing the original data, a geometric adaptive sampling algorithm based on the change of the normal vector is used to generate a training data set suitable for a point cloud neural network. The local geometric normal vector change rate is used as the core index, the sampling probability is dynamically allocated by measuring the surface curvature characteristics, thereby achieving dense sampling in areas with large curvature and sparse sampling in smooth areas. The CFD simulation results are converted from a structured grid form to a three-dimensional point cloud form that can preserve geometric characteristics. This step reduces the prediction error of the subsequent proxy model while ensuring the fidelity of the geometric information.

[0042] S5, taking the coordinates of each point in the point cloud as input and taking the flow field physical quantity of each point as output label, training the point cloud neural network to obtain an underwater vehicle three-dimensional flow field prediction system.

[0043] Specifically, each set of training data contains the point cloud coordinates obtained by sampling and the flow field physical quantity corresponding to the point cloud. The point cloud coordinates are used as input samples, and the corresponding flow field physical quantity is used as a sample label. The point cloud neural network is trained based on the training data to obtain an underwater vehicle three-dimensional flow field prediction system.

[0044] In an embodiment, both the input and label data can be normalized. Specifically, the coordinates of each point in the point cloud are normalized by the minimum-maximum normalization method and input into the point cloud neural network, and the flow field physical quantity of each point is normalized by the minimum-maximum normalization method and used as the output label.

[0045] The formula for minimum-maximum normalization of any parameter X is: ; In the formula, is the minimum value of parameter X in the global minimum value, Let X be the global maximum value. This is the result after performing min-max normalization on the current parameter x.

[0046] Normalization, which maps data to the [0,1] interval, can improve the stability of network training.

[0047] Point cloud neural networks can adopt conventional architectures, such as Figure 5 The diagram shows the architecture of a cloud neural network in one embodiment of the present invention. The network includes: (1) a 3×3 input transformation module to enhance rotation invariance through a spatial transformation matrix; (2) a shared weight multilayer perceptron sequence for local feature extraction; (3) a 64×64 feature transformation module to stabilize high-dimensional feature representation; (4) a max-pooling symmetric function to aggregate global features and ensure invariance to the input order of the point cloud; and (5) a multilayer perceptron to implement physical quantity regression prediction. The input is the three-dimensional coordinates of the point cloud or the normalized three-dimensional coordinates of the point cloud, and the output is the flow field parameters of each point or the normalized flow field parameters. The Adam optimizer is used, and the mean square error is used as the loss function for training. When the training loss converges to a preset level, it can be used as a three-dimensional flow field prediction system for underwater vehicles. For example, the Adam optimizer, batch size 64, and mean square error loss function are used to train for 3000 rounds. The dataset is divided into training, validation, and test sets in a ratio of 8:1:1. The training results show that the validation loss converges stably to 3×10 -5 Scale, on the test set , and Predicted R 2 They reached 0.9934, 0.9972 and 0.9482 respectively.

[0048] It should be noted that the hyperparameter settings of the neural network (including optimizer type, batch size, learning rate, training epochs, and loss function form) are not limited to the above configurations. They can be adjusted according to the specific task scale, data characteristics, and computing resources to achieve the best training results.

[0049] Overall, while existing point cloud neural networks can directly process unstructured data, they typically rely on uniform or random sampling, resulting in sparse point cloud distribution in key regions such as high curvature and abrupt pressure gradient changes. This leads to significant local prediction errors and severe loss of flow field details. This invention employs an adaptive sampling strategy based on geometric curvature features to dynamically adjust the point cloud sampling density, achieving data encryption in key flow regions and thus significantly improving the accuracy and stability of flow field prediction.

[0050] In a second aspect, the present application also relates to a method for optimizing a three-dimensional geometric model of an underwater vehicle, comprising: based on a set evaluation objective function, taking the coordinate combination of the surface point cloud of the three-dimensional geometric model as the optimization variable, and using an optimization algorithm to optimize in the preset variable space range to obtain the optimal three-dimensional geometric model; wherein the evaluation objective function is used to calculate the total resistance and average adverse pressure gradient of the corresponding model based on the flow field physical quantity of each point of the three-dimensional geometric model, the flow field physical quantity of each point of the three-dimensional geometric model is obtained by inputting the coordinate combination of the surface point cloud of the corresponding three-dimensional geometric model into the underwater vehicle three-dimensional flow field prediction system, and the underwater vehicle three-dimensional flow field prediction system is trained by using the training method as described above.

[0051] Specifically, the trained underwater vehicle three-dimensional flow field prediction system can be embedded into the optimization framework as a surrogate model. In the optimization solving process, the surrogate model is used to directly predict the flow field physical quantity corresponding to different coordinate combinations, thereby avoiding repeated calculation of a large amount of CFD simulation, greatly improving the optimization efficiency, and substituting the predicted flow field physical quantity into the evaluation objective function to calculate the objective function value.

[0052] The optimization algorithm can use conventional algorithms such as non-dominated sorting genetic algorithm (NSGA-II), particle swarm optimization (PSO), differential evolution algorithm (DE), and Bayesian optimization (BO) to realize global search and rapid convergence of the multi-objective design space.

[0053] Taking the genetic algorithm as an example, the population size is set to 100, the crossover probability is 0.9, the mutation probability is 0.1, and the evolution is 100 generations. In the optimization process, the surrogate model is directly called to quickly evaluate the objective function, and a single prediction only takes 0.03 seconds. The Pareto optimal solution obtained is verified by CFD, which shows that compared with the baseline design, the optimal design scheme can reduce the total resistance by 2.8 N, and the maximum average adverse pressure gradient is reduced by 17.7%, and the prediction error of the resistance and pressure gradient surrogate model is less than 2% and 4%, respectively. Geometric analysis shows that the curvature of the bow is more smooth after optimization, and the contraction rate of the stern is adjusted to improve the pressure recovery characteristics.

[0054] The present application combines the trained prediction system with a multi-objective optimization algorithm to realize rapid iterative design and optimization of the underwater vehicle in terms of drag reduction and pressure gradient control.

[0055] In a third aspect, the present application also relates to an electronic device.

[0056] The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0057] The electronic device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The processor can be a central processing unit (CPU), and can 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, and the like. The memory can be used to store computer programs and / or modules, and the processor can realize various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and calling data stored in the memory.

[0058] In a fourth aspect, the present application also relates to a computer readable storage medium.

[0059] The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the above method.

[0060] Specifically, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0061] In a fifth aspect, the present application also relates to a computer program product or a computer program.

[0062] The computer program product or the computer program includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps of the above method.

[0063] In the following, the effect of the technical scheme of the present application will be described through specific experimental data.

[0064] As shown in FIG. 1, a point cloud is obtained by sampling a three-dimensional model according to the present application. Figure 6 FIG. 2 shows a point cloud obtained by uniform sampling in an embodiment, and FIG. 3 shows a point cloud obtained by non-uniform sampling in an embodiment. Figure 7The figure shows a point cloud obtained by adaptive sampling in one embodiment of the present invention. In the figure, the point cloud of uniform sampling is evenly distributed across the entire hull surface, while the point cloud of adaptive sampling is significantly denser in areas with greater curvature, such as the bow and stern. This shows that the adaptive sampling method can increase the sampling density in areas with greater curvature and more fully preserve key surface features.

[0065] like Figure 8 The figure shown is a prediction result of the pressure coefficient in one embodiment of the present invention, where (a) corresponds to the pressure coefficient label value, (b) corresponds to the pressure coefficient prediction value, and (c) corresponds to the deviation between the pressure coefficient prediction value and the label value. It can be seen that the prediction result is highly consistent with the label value, with only a small error in local areas, indicating that the model of the present invention can accurately predict the pressure coefficient distribution.

[0066] like Figure 9 The figure shown is a prediction result of the shear stress coefficient in one embodiment of the present invention, where (a) corresponds to the label value of the shear stress coefficient, (b) corresponds to the predicted value of the shear stress coefficient, and (c) corresponds to the deviation between the predicted value and the label value of the shear stress coefficient. It can be seen that the prediction result maintains a high degree of consistency with the label value, the main distribution characteristics of the bow and midsection are accurately reconstructed, and the overall error is small, indicating that the model of the present invention can accurately predict the shear stress coefficient.

[0067] like Figure 10 The figure shown is a pressure gradient prediction effect diagram in one embodiment of the present invention, where (a) corresponds to the pressure gradient label value, (b) corresponds to the pressure gradient predicted value, and (c) corresponds to the deviation between the pressure gradient predicted value and the label value. It can be seen that the prediction result can reproduce the overall distribution trend of the pressure gradient well. The main change areas at the bow and stern are basically consistent with the label values, and the overall error is small, indicating that the model of the present invention also has high prediction accuracy for the pressure gradient.

[0068] like Figure 11 The diagram shows a comparison of the geometric models before and after optimization in one embodiment of the present invention. (a) corresponds to the inlet profile curve, and (b) corresponds to the outlet profile curve. The original model is the baseline model before optimization, while optimized model 1 and optimized model 2 are the two optimized results output by the optimization algorithm. It can be seen that the optimized models exhibit a smoother curvature transition in the inlet section compared to the original model. In the outlet section, both optimized models adjust the geometry by slowing down or accelerating the contraction rate, respectively. These changes make the shape more conducive to flow development, providing a structural basis for reducing drag and improving the pressure gradient.

[0069] like Figure 12The figure shows a comparison of pressure coefficient curves before and after optimization in one embodiment of the present invention. (a) compares the pressure coefficient curves of the original model and optimized model 1, while (b) compares the pressure coefficient curves of the original model and optimized model 2. In the figure, the optimized model exhibits a smoother pressure drop and a higher minimum pressure in the inlet section compared to the original model. The negative pressure trough value in the outlet section is significantly reduced, and the pressure recovery of optimized model 2 is more stable. This demonstrates that the optimized model, by improving the pressure transition in the inlet section and reducing the negative pressure intensity in the outlet section, helps to reduce pressure differential resistance and improve overall flow characteristics.

[0070] like Figure 13 The figure shows a comparison of shear stress curves before and after optimization in one embodiment of the present invention. (a) corresponds to the shear stress of the original model, (b) corresponds to the shear stress of optimized model 1, and (c) corresponds to the shear stress of optimized model 1. In the figure, the average shear stress of both optimized models in the inlet section is lower than that of the original model, indicating that the smooth transition in the inlet section effectively reduces frictional resistance. In the outlet section, the overall shear stress of optimized model 1 is lower than that of optimized model 2. This shows that both optimization schemes achieve overall drag reduction by reducing frictional resistance, while optimized model 1 has even lower shear resistance in the tail region, resulting in a more significant reduction in total drag.

[0071] like Figure 14 The figure shows a comparison of pressure gradient curves before and after optimization in one embodiment of the present invention. (a) compares the pressure gradient curves of the original model and optimized model 1, while (b) compares the pressure gradient curves of the original model and optimized model 2. In the figure, the pressure gradient change in the inlet section of the optimized model is more gradual than that of the original model. The peak values ​​of the adverse pressure gradient in the outlet section are significantly lower than those of the original model, and the reduction in the pressure gradient of optimized model 2 is more significant. This demonstrates that the optimized model, by effectively reducing the adverse pressure gradient, helps to suppress flow separation and improve overall flow characteristics.

[0072] Overall, the present invention has the following advantages: It directly processes unstructured grid point clouds, avoiding the information loss introduced by traditional interpolation; The adaptive sampling strategy effectively improves the prediction accuracy for high curvature regions and key flow regions; The surrogate model has extremely high computational efficiency, with a single-sample prediction time of approximately 0.03 seconds, which is only 0.0017% of that of CFD. When combined with a multi-objective optimization algorithm, an efficient balance can be achieved between drag reduction and pressure gradient control; It demonstrated significant potential for drag reduction and noise reduction in SUBOFF vehicle validation, and has promising prospects for engineering applications.

[0073] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" in this invention are intended to illustrate the invention and are not intended to limit the invention.

[0074] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method of training an underwater vehicle three-dimensional flow field prediction system, the method comprising: The method comprises the following steps: obtaining an original data set, each set of original data containing a different three-dimensional geometric model of a vehicle and flow field physical quantities of each grid point in a surface grid of the vehicle model obtained by a computational fluid dynamics simulation method; For each surface grid point of the aircraft model , For the i-th surface mesh point, extract the K nearest mesh points from the surface mesh to form this point. nearest neighbor set K is a preset number of points; calculate the centroid of the nearest neighbor set. Construct a set of nearest neighbors Its center of mass The covariance matrix between the two is analyzed, and eigenvalue decomposition is performed. The eigenvector corresponding to the smallest eigenvalue in the decomposition is taken as the point. normal vector ; Computing points The normal vector between each point The angle between the normal vectors , The jth point in the point set Take the maximum angle obtained As the geometric feature metric of point , Sigmoid function calculation is performed on To get the weight of point , , The minimum value of the geometric feature metric of all surface grid points is calculated The ratio of the weight to the sum of the weights of all surface grid points, the sampling probability of point ; ; based on the obtained sampling probability, randomly sampling surface grid points from the original surface grid of the three-dimensional geometric model of the vehicle to obtain a point cloud corresponding to the model; training the point cloud neural network with the coordinates of each point in the point cloud as input and the flow field physical quantities of each point as output labels to obtain a three-dimensional flow field prediction system for underwater vehicles.

2. The training method of claim 1, wherein, The surface contour of the vehicle is generated by a Bezier curve to obtain a three-dimensional geometric model corresponding to the vehicle.

3. The training method of claim 1, wherein The flow field physical quantities include pressure coefficients, shear stress coefficients and pressure gradient distributions.

4. The training method of claim 1, wherein Weight The calculation formula is: ； In the formula, k is a preset scaling factor.

5. The training method of claim 1, wherein, Points of sampling probability The calculation formula is: ； In the formula, is the weight of the nth surface grid point, and N is the number of surface grid points.

6. The training method of claim 1, wherein, The coordinates of each point in the point cloud are input into the point cloud neural network after minimum-maximum normalization processing, and the flow field physical quantities of each point are used as output labels after minimum-maximum normalization processing. The formula for minimum-maximum normalization processing of any parameter X is: ； wherein is the global minimum of the parameter X, is the global maximum of the parameter X, is the result of the min-max normalization of the current parameter x.

7. A method of optimizing a three-dimensional geometric model of an underwater vehicle, the method comprising: The method comprises the following steps: based on the set evaluation objective function, the coordinate combination of the surface point cloud of the three-dimensional geometric model is used as the optimization variable, and an optimization algorithm is used to optimize in the preset variable space range to obtain an optimal three-dimensional geometric model; wherein the evaluation objective function is used to calculate the total drag and average adverse pressure gradient of the corresponding model based on the flow field physical quantities of each point of the three-dimensional geometric model, the flow field physical quantities of each point of the three-dimensional geometric model are obtained by inputting the coordinate combination of the surface point cloud of the corresponding three-dimensional geometric model into the three-dimensional flow field prediction system for underwater vehicles, and the three-dimensional flow field prediction system for underwater vehicles is trained by using the training method of any one of claims 1 to 6. 8.An electronic device comprising a memory and a processor, the memory storing a computer program, wherein, The processor executes the computer program to realize the steps of the method of any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7.

10. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions are executed by the processor to realize the steps of the method of any one of claims 1 to 7.

Images