Optimization method and device for head shape and water entry strategy of water-air cross-medium robot

Abstract

The present invention provides a method and device for optimizing the head shape and water entry strategy of a water-air cross-media robot, including determining a design domain; performing Latin hypercube sampling within the design domain to obtain a design parameter vector; drawing a NURBS curve to form a water-air cross-media robot model; resetting parameters; generating a simulation domain and meshing it; performing a finite element solution on the current mesh to extract solution data; constructing low-quality and high-quality data sets; obtaining the optimal hyperparameters of a proxy model through Bayesian optimization; training the proxy model based on the optimal hyperparameters; constructing an optimization target and employing a multi-objective optimization algorithm to obtain a solution that satisfies the biomimetic morphological and component constraints under the optimization target; determining whether the proxy model has converged and outputting the optimal design parameters. By constructing a fully automated design paradigm for parametric modeling, numerical simulation, and parameter optimization, the present invention avoids manual real-time intervention in the intermediate process, significantly improving process efficiency.

Description

Technical Field

[0001] The present invention relates to the field of water-air cross-medium robot design, and in particular to a method and device for optimizing the head shape and water entry strategy of a water-air cross-medium robot. Background Art

[0002] The water-air cross-media robot is a new type of intelligent device capable of freely switching between air and water. The enormous impact loads it faces during water entry can easily damage the robot, thus presenting a key bottleneck hindering the practical application of this technology. Existing research indicates that head shape and entry strategy are key factors determining the impact loads. Therefore, optimization of the head shape and entry strategy of the water-air cross-media robot is necessary to minimize these impact loads.

[0003] Existing technologies mostly design head shape based on empirical formulas or single working condition simulations. Because empirical formulas and single working condition simulations rely on simplified assumptions under specific conditions, and actual water entry involves the coupling of multiple variables, resulting in insufficient model generalization ability, the optimization results are not effective in actual applications and are difficult to adapt to complex and changeable water entry conditions.

[0004] Existing water entry strategies typically utilize CFD simulation technology. This involves building a dataset using design parameters and load data predicted by CFD simulation. Models such as radial basis function neural networks (RBFNs) and dense neural networks (DNNs) are then trained to predict the water entry impact loads of water-air-crossing media robots with different head shapes. However, existing technologies require real-time manual intervention during data sampling and optimization, resulting in inefficiencies. Furthermore, the model's accuracy is limited, making efficient searches in high-dimensional design spaces impossible. Summary of the Invention

[0005] In response to the defects of the existing technology, the present invention provides a method and device for optimizing the head shape and water entry strategy of a water-air cross-medium robot.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In one aspect, the present invention provides a method for optimizing the head shape and water entry strategy of a water-air cross-medium robot, comprising the following steps:

[0008] Determine the design domain based on bionic morphological constraints, component constraints, and motion constraints;

[0009] Perform Latin hypercube sampling in the design domain to obtain the design parameter vector, and determine whether the design parameter vector meets the bionic morphological constraints and component constraints. If so, add the design parameter vector to the sample set; if not, resample. If the number of iterative sampling reaches a threshold, redefine the design domain.

[0010] Based on the design parameter vectors in the current sample set, a NURBS curve is drawn to describe the head shape; a three-dimensional head model is generated based on the NURBS curve, and the components are assembled to form a water-air cross-medium robot model;

[0011] Reset the current water-air cross-media robot model, its position, posture and motion parameters;

[0012] Automatically check and repair surface mesh problems in imported models, generate a simulation domain, and perform meshing. Preset mesh accuracy is used for the water pool and air. Meshes are refined based on the preset mesh accuracy for the water surface, path, water-air cross-medium robot, and its surroundings.

[0013] Perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data;

[0014] Based on the current low-precision solution data and the design parameter vectors in the current sample set, a low-quality data set is constructed to preliminarily screen the design parameter vectors in the sample set;

[0015] Perform finite element solution on the filtered design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the filtered design parameter vector and high-precision solution data, a high-quality dataset is constructed for training the surrogate model.

[0016] Based on high-quality datasets, we use Bayesian optimization to obtain the optimal hyperparameters of the surrogate model.

[0017] Train the surrogate model based on the optimal hyperparameters;

[0018] Construct an optimization goal, use a multi-objective optimization algorithm to obtain a solution that satisfies the bionic morphological constraints and component constraints under the optimization goal, and select solutions with large variance as annotation samples;

[0019] Determine whether the proxy model has converged. If so, output the optimal design parameters. If not, continue to label samples and expand high-quality data sets, and retrain the proxy model until the proxy model converges.

[0020] Furthermore, the design domain is an eight-dimensional space, including water entry strategy parameters and nose shape parameters. The water entry strategy parameters include water entry speed and water entry angle. The nose shape parameters include beak length, beak thickness, horizontal shape of the front end of the head, vertical shape of the front end of the head, horizontal shape of the rear end of the head, and vertical shape of the rear end of the head.

[0021] Furthermore, the design domain that satisfies the bionic morphological constraints and component constraints is calculated according to the following formula:

[0022] Assembly constraints:

[0023] ;

[0024] in, is the feasible design domain; is the design domain; is the component constraint function; The nose shape parameters z Under the condition of NURBS curve X The function value at ;

[0025] Bionic morphological constraints:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] in, is the bionic morphological constraint function; l is the head length.

[0031] Furthermore, it also includes the determination of the number of high-precision grids:

[0032] Starting from the preset grid accuracy, the number of grids is gradually increased. The grid accuracy is judged according to the impact load change curve in the solution data. As the number of grids increases, if the curve changes, the number of grids is continued to increase; if the curve change amplitude is less than the threshold, the number of grids before the increase is taken as the number of grids for the high-precision grid.

[0033] Furthermore, the overlapping grid technique is adopted in the gridding process.

[0034] Furthermore, the proxy model is a deep kernel learning model.

[0035] Furthermore, the optimal hyperparameters of the surrogate model are obtained according to the following steps:

[0036] Construct the Bayesian hyperparameter optimization objective function:

[0037] ;

[0038] ;

[0039] in, is the proxy model; is the search space; n is the number of hyperparameters; is a combination of hyperparameters; is the optimal hyperparameter; is the loss of the proxy model to be trained under the hyperparameter combination on the validation set;

[0040] Random in the search space Extract k Set initial hyperparameter combination ,calculate , forming the initial data set ;

[0041] The acquisition function is used to iteratively calculate the hyperparameter combination that satisfies the Bayesian hyperparameter optimization objective function to obtain the optimal hyperparameters of the surrogate model:

[0042] ;

[0043] ;

[0044] ;

[0045] in, is the mean; K (·) is the covariance kernel function; is the currently known optimal objective function value; For the dataset; , T is the number of cycles.

[0046] Furthermore, in the multi-objective optimization algorithm, the optimization objective is:

[0047] ;

[0048] in, is the mean axial impact load output by the proxy model; is the mean normal impact load output by the proxy model; is the water entry speed;

[0049] The solution that satisfies the bionic morphology constraint and component constraint is obtained according to the following formula:

[0050] ;

[0051] in, To find solutions that satisfy biomimetic morphological and component constraints; The solution obtained by the multi-objective optimization algorithm.

[0052] Furthermore, the following steps are used to determine whether the proxy model has converged:

[0053] Calculate the percentage absolute error between the solution data and the output of the surrogate model. If the percentage absolute error is less than 1%, the surrogate model is considered to have converged.

[0054] The percentage absolute error is calculated according to the following formula:

[0055] ;

[0056] ;

[0057] in, n is the number of labeled samples; is the percentage absolute error of the axial impact load; is the percentage absolute error of the normal impact load; To solve the data i The peak axial impact load of the marked samples; Output of the proxy model i The peak axial impact load of the marked samples; To solve the data i Peak normal impact load of the marked samples; Output of the proxy model i Peak normal impact load of the marked samples.

[0058] In another aspect, the present invention provides a device for optimizing the head shape and water entry strategy of a water-air cross-medium robot, comprising:

[0059] The first module is used to determine the design domain based on bionic morphological constraints, component constraints, and motion constraints;

[0060] The second module is used to perform Latin hypercube sampling in the design domain to obtain the design parameter vector and determine whether the design parameter vector meets the bionic morphological constraints and component constraints. If so, the design parameter vector is added to the sample set. If not, resampling is performed. If the number of iterative sampling reaches a threshold, the design domain is re-determined.

[0061] The third module is used to draw NURBS curves based on the design parameter vectors in the current sample set to describe the head shape; generate a three-dimensional head model based on the NURBS curves, and complete component assembly to form a water-air cross-medium robot model;

[0062] The fourth module is used to reset the model of the current water-air cross-medium robot, its position, posture and motion parameters;

[0063] The fifth module is used to automatically check and repair surface mesh problems in the imported model, generate the simulation domain, and perform meshing. The water pool and air use the preset mesh accuracy; the water surface, path, water-air cross-medium robot, and its surroundings are meshed based on the preset mesh accuracy.

[0064] The sixth module is used to perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data;

[0065] The seventh module is used to construct a low-quality data set based on the current low-precision solution data and the design parameter vectors in the current sample set; it is used to preliminarily screen the design parameter vectors in the sample set;

[0066] The eighth module is used to perform finite element solution on the screened design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the screened design parameter vector and the high-precision solution data, a high-quality dataset is constructed for training the surrogate model.

[0067] The ninth module is used to obtain the optimal hyperparameters of the surrogate model through Bayesian optimization based on high-quality datasets;

[0068] The tenth module is used to train the agent model based on the optimal hyperparameters;

[0069] The eleventh module is used to construct the optimization target, use the multi-objective optimization algorithm to obtain the solution that satisfies the bionic morphological constraints and component constraints under the optimization target, and select the solution with large variance as the annotation sample;

[0070] The twelfth module is used to determine whether the proxy model has converged. If it has converged, the optimal design parameters are output. If it has not converged, the samples are continued to be labeled and the high-quality data set is expanded, and the proxy model is retrained until the proxy model converges.

[0071] Compared with the prior art, the beneficial technical effects of the present invention are:

[0072] The method and device for optimizing the head shape and water entry strategy of the water-air cross-medium robot provided by the present invention determine the design domain by integrating bionic morphological constraints, component constraints, and motion constraints, and use NURBS curves to describe the head shape. This makes the head shape designed by the present invention diverse and can be optimized based on the bionic shape, thereby achieving a good load reduction effect.

[0073] The present invention constructs a low-quality data set to conduct a preliminary screening of the design parameter vectors in the sample set, determine the approximate impact of different design parameters on the impact load, and exclude most non-critical areas through screening, thereby reducing the number of high-precision simulations and reducing simulation time. At the same time, by determining whether the proxy model has converged, the high-quality data set is expanded when it has not converged, thereby introducing an active learning strategy to conduct targeted exploration of key areas, achieving efficient search of high-dimensional design space and improving the model prediction accuracy in key areas. On the other hand, the present invention constructs a full-process automated design paradigm of parametric modeling-numerical simulation-parameter optimization, thereby avoiding real-time manual intervention in the intermediate process and greatly improving process efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0075] Figure 1 A schematic flow chart of a method for optimizing the head shape and water entry strategy of a water-air cross-medium robot provided in one embodiment;

[0076] Figure 2 A diagram showing the relationship between the control point positions and the head shape provided in one embodiment;

[0077] Figure 3 A schematic diagram of the division of the simulation calculation domain provided by an embodiment;

[0078] Figure 4 A schematic diagram of a gridded simulation calculation domain provided by an embodiment. DETAILED DESCRIPTION

[0079] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0080] Reference Figure 1 One embodiment provides a method for optimizing the head shape and water entry strategy of a water-air cross-medium robot, including the following steps:

[0081] Determine the design domain based on bionic morphological constraints, component constraints, and motion constraints;

[0082] Perform Latin hypercube sampling in the design domain to obtain the design parameter vector, and determine whether the design parameter vector meets the bionic morphological constraints and component constraints. If so, add the design parameter vector to the sample set; if not, resample. If the number of iterative sampling reaches a threshold, redefine the design domain.

[0083] Based on the design parameter vectors in the current sample set, a NURBS curve is drawn to describe the head shape; a three-dimensional head model is generated based on the NURBS curve, and the components are assembled to form a water-air cross-medium robot model;

[0084] Reset the current water-air cross-media robot model, its position, posture and motion parameters;

[0085] Automatically check and repair surface mesh problems in imported models, generate a simulation domain, and perform meshing. Preset mesh accuracy is used for the water pool and air. Meshes are refined based on the preset mesh accuracy for the water surface, path, water-air cross-medium robot, and its surroundings.

[0086] Perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data;

[0087] Based on the current low-precision solution data and the design parameter vectors in the current sample set, a low-quality data set is constructed to preliminarily screen the design parameter vectors in the sample set;

[0088] Perform finite element solution on the filtered design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the filtered design parameter vector and high-precision solution data, a high-quality dataset is constructed for training the surrogate model.

[0089] Based on high-quality datasets, we use Bayesian optimization to obtain the optimal hyperparameters of the surrogate model.

[0090] Train the surrogate model based on the optimal hyperparameters;

[0091] Construct an optimization goal, use a multi-objective optimization algorithm to obtain a solution that satisfies the bionic morphological constraints and component constraints under the optimization goal, and select solutions with large variance as annotation samples;

[0092] Determine whether the proxy model has converged. If so, output the optimal design parameters. If not, continue to label samples and expand high-quality data sets, and retrain the proxy model until the proxy model converges.

[0093] In one embodiment, the design domain is an eight-dimensional space, including water entry strategy parameters and nose shape parameters. The water entry strategy parameters include water entry speed and water entry angle. The nose shape parameters include beak length, beak thickness, horizontal shape of the front end of the head, vertical shape of the front end of the head, horizontal shape of the rear end of the head, and vertical shape of the rear end of the head. Specifically, referring to Figure 2 The head shape of the kingfisher is segmented using non-uniform rational B-spline curves (NURBS curves). Four control points are used to control the nose shape. Two of these control points control the beak length and thickness, respectively. Both control points have one degree of freedom of movement: the first control point can move horizontally to change the beak length, and the second control point can move vertically to change the beak thickness. The other two control points control the head shape. Both control points have two degrees of freedom of movement: horizontal and vertical, controlling the shape of the front and rear ends of the head, respectively. These constitute the nose shape parameters. Since the water impact load is related to the entry strategy, the entry strategy parameters of water velocity and angle are also considered.

[0094] Motion constraints are the water entry strategy constraints determined by the actual motion capability and the lower limit of the motion index of the water-air cross-medium robot. Motion constraints are directly reflected in the upper and lower limits of the water entry strategy parameters.

[0095] The design domain that satisfies the bionic morphological constraints and component constraints is calculated according to the following formula:

[0096] Assembly constraints:

[0097] ;

[0098] in, is a feasible design domain, which is a design domain that satisfies the component constraint function; is the design domain; is the component constraint function; The nose shape parameters z Under the condition of NURBS curve X The function value at ;

[0099] Bionic morphological constraints:

[0100] ;

[0101] ;

[0102] ;

[0103] ;

[0104] in, is the bionic morphological constraint function;l is the head length.

[0105] In one embodiment, the design domain is shown in Table 1:

[0106] Table 1 Design domain parameter range

[0107]

[0108] Based on the design parameter vector in the current sample set, a NURBS curve is drawn, and the NURBS curve is discretized and sampled to obtain a spline curve coordinate point set. A spline curve is generated according to the spline curve coordinate point set, and a three-dimensional head model is further generated. The generated three-dimensional head model is then assembled with the prepared fuselage model to complete the component assembly and form a water-air cross-medium robot model.

[0109] Reset the model, position, posture, and motion parameters of the current water-air cross-medium robot to adapt to the water entry angle and speed.

[0110] Reference Figure 3 In one embodiment, the size of the simulation design domain is 9m×8m×9m, including a water pool, a water surface, a water-air cross-medium robot, a water-air cross-medium robot block, and a path, wherein the water-air cross-medium robot block completely wraps the water-air cross-medium robot; the path is set along the movement direction of the water-air cross-medium robot.

[0111] The simulation domain is meshed, where the water pool and air use the preset mesh accuracy; the water surface, path, water-air cross-medium robot and its surroundings are meshed based on the preset mesh accuracy. Figure 4 As shown in the figure, it can be seen that the density of the water surface, path, water-air cross-medium robot and its surroundings is significantly increased after refinement, which can improve the solution accuracy.

[0112] Automatically check and repair surface mesh issues on imported models before generating and meshing the design domain, ensuring mesh feasibility.

[0113] In a preferred embodiment, overlapping grid technology is used for gridding. The overlapping grid technology can effectively process complex structures with relative motion without the need for grid deformation or regeneration of the grid, thereby enabling adaptive grid refinement to refine or coarsen the grids on both sides of the overlapping grid interface, thereby improving computational efficiency while ensuring computational accuracy.

[0114] In a preferred embodiment, the gridding process further includes determining the number of high-precision grids:

[0115] Starting from the preset grid accuracy, the number of grids is gradually increased. The grid accuracy is judged according to the impact load change curve in the solution data. As the number of grids increases, if the curve changes, the number of grids is continued to increase; if the curve change amplitude is less than the threshold, the number of grids before the increase is taken as the number of grids for the high-precision grid.

[0116] By determining the number of high-precision grids as described above, the feasibility of grid division can be ensured.

[0117] Perform a finite element solution on the current mesh and extract the solution data. The solution data includes low-precision solution data and high-precision solution data. The peak impact load obtained from the low-precision mesh simulation is used as the true value in the low-quality data set, while the peak impact load obtained from the high-precision mesh simulation is used as the true value in the high-precision data set. Before solving, you need to clear the previous solution data and then solve again.

[0118] Based on the current low-precision solution data and the design parameter vectors in the current sample set, a low-quality data set is constructed to preliminarily screen the design parameter vectors in the sample set;

[0119] In one embodiment, 500 sets of low-precision solution data obtained by simulation on a grid with a preset grid precision (i.e., a low-precision grid) are used to construct a low-precision dataset. The low-precision dataset is then used to filter samples from the sample set, identifying 200 candidate samples with low impact loads. These 200 selected candidate samples with low impact loads are then simulated on a high-precision grid to obtain high-precision solution data, constructing a high-quality dataset. This high-quality dataset is then divided into a training set, a validation set, and a test set in a 6:2:2 ratio for training the proxy model.

[0120] The design parameter vectors are screened through low-quality data sets, and the influence of different design parameter vectors on the impact load is utilized to exclude most non-critical areas, reduce the number of high-precision simulations, improve processing efficiency, and save energy.

[0121] In one embodiment, the proxy model is a deep kernel learning model (DKL model). The DKL model uses Gaussian process regression, constructs a kernel function based on DNN (deep neural network) features, and outputs predicted expectation and variance; the output expectation (i.e., mean) is used as an estimate of the impact load, and the variance is used as a measure of output uncertainty.

[0122] The optimal hyperparameters of the surrogate model are obtained through Bayesian optimization. The optimal hyperparameters of the surrogate model are obtained according to the following steps:

[0123] Construct the Bayesian hyperparameter optimization objective function:

[0124] ;

[0125] ;

[0126] in, is the proxy model; is the search space; n is the number of hyperparameters; is a combination of hyperparameters; is the optimal hyperparameter; is the loss of the proxy model to be trained under the hyperparameter combination on the validation set;

[0127] Random in the search space Extract k Set initial hyperparameter combination ,calculate , forming the initial data set ;

[0128] The acquisition function is used to iteratively calculate the hyperparameter combination that satisfies the Bayesian hyperparameter optimization objective function to obtain the optimal hyperparameters of the surrogate model:

[0129] ;

[0130] ;

[0131] ;

[0132] in, is the mean; K (·) is the covariance kernel function; is the currently known optimal objective function value; For the dataset; , T is the number of cycles.

[0133] In one embodiment, the NSGA-III multi-objective optimization algorithm is used to obtain a solution that satisfies the bionic morphological constraints and component constraints under the optimization objectives.

[0134] In the multi-objective optimization algorithm, the optimization objectives are:

[0135] ;

[0136] in, is the mean axial impact load output by the proxy model; is the mean normal impact load output by the proxy model; is the water entry speed; by adding the water entry speed into the optimization objective, the requirement of rapid water entry of the water-air cross-medium robot can be met.

[0137] The solution that satisfies the bionic morphology constraint and component constraint is obtained according to the following formula:

[0138] ;

[0139] in, To find solutions that satisfy biomimetic morphological and component constraints; is the solution obtained by the multi-objective optimization algorithm. According to the variance Sort by size in descending order and filter out the top n The sample corresponding to the solution with the largest sum of variances is taken as the labeled sample.

[0140] Determine whether the proxy model has converged, and thus whether the accuracy of the model in the key area is high enough. If it has converged, output the optimal design parameters. If it has not converged, expand the high-quality dataset and retrain the proxy model with the expanded high-quality dataset until the proxy model converges.

[0141] In one embodiment, whether the proxy model has converged is determined according to the following steps:

[0142] Calculate the percentage absolute error between the solution data and the output of the proxy model. If the percentage absolute error is less than 1%, the proxy model is considered to have converged. If the proxy model has not converged, n The labeled samples are added to the high-quality dataset to expand the high-quality dataset.

[0143] The percentage absolute error is calculated according to the following formula:

[0144] ;

[0145] ;

[0146] in, n is the number of labeled samples; is the percentage absolute error of the axial impact load; is the percentage absolute error of the normal impact load; To solve the data i The peak axial impact load of the marked samples; Output of the proxy model i The peak axial impact load of the marked samples; To solve the data i Peak normal impact load of the marked samples; Output of the proxy model i Peak normal impact load of the marked samples.

[0147] In one embodiment, a device for optimizing the head shape and water entry strategy of a water-air cross-medium robot is provided, comprising:

[0148] The first module is used to determine the design domain based on bionic morphological constraints, component constraints, and motion constraints;

[0149] The second module is used to perform Latin hypercube sampling in the design domain to obtain the design parameter vector and determine whether the design parameter vector meets the bionic morphological constraints and component constraints. If so, the design parameter vector is added to the sample set. If not, resampling is performed. If the number of iterative sampling reaches a threshold, the design domain is re-determined.

[0150] The third module is used to draw NURBS curves based on the design parameter vectors in the current sample set to describe the head shape; generate a three-dimensional head model based on the NURBS curves, and complete component assembly to form a water-air cross-medium robot model;

[0151] The fourth module is used to reset the model of the current water-air cross-medium robot, its position, posture and motion parameters;

[0152] The fifth module is used to automatically check and repair surface mesh problems in the imported model, generate the simulation domain, and perform meshing. The water pool and air use the preset mesh accuracy; the water surface, path, water-air cross-medium robot, and its surroundings are meshed based on the preset mesh accuracy.

[0153] The sixth module is used to perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data;

[0154] The seventh module is used to construct a low-quality data set based on the current low-precision solution data and the design parameter vectors in the current sample set; it is used to preliminarily screen the design parameter vectors in the sample set;

[0155] The eighth module is used to perform finite element solution on the screened design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the screened design parameter vector and the high-precision solution data, a high-quality dataset is constructed for training the surrogate model.

[0156] The ninth module is used to obtain the optimal hyperparameters of the surrogate model through Bayesian optimization based on high-quality datasets;

[0157] The tenth module is used to train the agent model based on the optimal hyperparameters;

[0158] The eleventh module is used to construct the optimization target, use the multi-objective optimization algorithm to obtain the solution that satisfies the bionic morphological constraints and component constraints under the optimization target, and select the solution with large variance as the annotation sample;

[0159] The twelfth module is used to determine whether the proxy model has converged. If it has converged, the optimal design parameters are output. If it has not converged, the samples are continued to be labeled and the high-quality data set is expanded, and the proxy model is retrained until the proxy model converges.

[0160] Matters not covered by the present invention are known technologies.

[0161] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0162] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements are all within the scope of protection of the present application.

[0163] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the head shape and water entry strategy of a water-air cross-medium robot, characterized in that: The following steps are involved: Determine the design domain based on bionic morphological constraints, component constraints, and motion constraints; Latin hypercube sampling is performed within the design domain to obtain the design parameter vector, and it is determined whether the design parameter vector satisfies the bionic morphological constraints and component constraints. If so, the design parameter vector is added to the sample set. If not, resampling is performed. If the number of iterative sampling reaches a threshold, the design domain is re-determined. The design domain that satisfies the bionic morphological constraints and component constraints is calculated according to the following formula: Assembly constraints: in, is the feasible design domain; is the design domain; is the component constraint function; The nose shape parameters z Under the condition of NURBS curve X The function value at ; Bionic morphological constraints: in, is the bionic morphological constraint function; l is the head length; Based on the design parameter vectors in the current sample set, a NURBS curve is drawn to describe the head shape; a three-dimensional head model is generated based on the NURBS curve, and the components are assembled to form a water-air cross-medium robot model; Reset the current water-air cross-media robot model, its position, posture and motion parameters; Automatically check and repair surface mesh problems of the imported model, generate the simulation domain and mesh it. The water pool and air use the preset mesh accuracy; the water surface, path, water-air cross-medium robot and its surroundings are meshed based on the preset mesh accuracy. Perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data; Based on the current low-precision solution data and the design parameter vectors in the current sample set, a low-quality data set is constructed to preliminarily screen the design parameter vectors in the sample set; Perform finite element solution on the filtered design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the filtered design parameter vector and high-precision solution data, a high-quality dataset is constructed for training the surrogate model. Based on high-quality datasets, we use Bayesian optimization to obtain the optimal hyperparameters of the surrogate model. Train the surrogate model based on the optimal hyperparameters; Construct an optimization goal, use a multi-objective optimization algorithm to obtain a solution that satisfies the bionic morphological constraints and component constraints under the optimization goal, and select solutions with large variance as annotation samples; Determine whether the proxy model has converged. If so, output the optimal design parameters. If not, continue to label samples and expand high-quality data sets, and retrain the proxy model until the proxy model converges.

2. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: The design domain is an eight-dimensional space, including water entry strategy parameters and nose shape parameters. The water entry strategy parameters include water entry speed and water entry angle. The nose shape parameters include beak length, beak thickness, horizontal shape of the front end of the head, vertical shape of the front end of the head, horizontal shape of the rear end of the head, and vertical shape of the rear end of the head.

3. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: It also includes the determination of the number of high-precision grids: Starting from the preset grid accuracy, the number of grids is gradually increased. The grid accuracy is judged according to the impact load change curve in the solution data. As the number of grids increases, if the curve changes, the number of grids is continued to increase; if the curve change amplitude is less than the threshold, the number of grids before the increase is taken as the number of grids for the high-precision grid.

4. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: The overlapping grid technique is used in the gridding process.

5. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: The proxy model is a deep kernel learning model.

6. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: The optimal hyperparameters of the surrogate model are obtained according to the following steps: Construct the Bayesian hyperparameter optimization objective function: in, is the proxy model; is the search space; n is the number of hyperparameters; is a combination of hyperparameters; is the optimal hyperparameter; is the loss of the proxy model to be trained under the hyperparameter combination on the validation set; Random in the search space Extract k Set initial hyperparameter combination ,calculate , forming the initial data set ; The acquisition function is used to iteratively calculate the hyperparameter combination that satisfies the Bayesian hyperparameter optimization objective function to obtain the optimal hyperparameters of the surrogate model: in, is the mean; K (·) is the covariance kernel function; is the currently known optimal objective function value; For the dataset; , T is the number of cycles.

7. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: In the multi-objective optimization algorithm, the optimization objectives are: in, is the mean axial impact load output by the proxy model; is the mean normal impact load output by the proxy model; is the water entry speed; The solution that satisfies the bionic morphology constraint and component constraint is obtained according to the following formula: in, To find solutions that satisfy biomimetic morphological and component constraints; The solution obtained by the multi-objective optimization algorithm.

8. The method for optimizing the head shape and water entry strategy of the water-air cross-medium robot according to claim 1, characterized in that: Follow these steps to determine if the surrogate model has converged: Calculate the percentage absolute error between the solution data and the output of the surrogate model. If the percentage absolute error is less than 1%, the surrogate model is considered to have converged. The percentage absolute error is calculated according to the following formula: in, n is the number of labeled samples; is the percentage absolute error of the axial impact load; is the percentage absolute error of the normal impact load; To solve the data i The peak axial impact load of the marked samples; Output of the proxy model i The peak axial impact load of the marked samples; To solve the data i Peak normal impact load of the marked samples; Output of the proxy model i Peak normal impact load of the marked samples.

9. A device for optimizing the head shape and water entry strategy of a water-air cross-medium robot, characterized in that: include: The first module is used to determine the design domain based on bionic morphological constraints, component constraints, and motion constraints; The second module is used to perform Latin hypercube sampling in the design domain to obtain the design parameter vector and determine whether the design parameter vector satisfies the bionic morphological constraints and component constraints. If so, the design parameter vector is added to the sample set. If not, resampling is performed. If the number of iterative sampling reaches a threshold, the design domain is re-determined. The design domain that satisfies the bionic morphological constraints and component constraints is calculated according to the following formula: Assembly constraints: in, is the feasible design domain; is the design domain; is the component constraint function; The nose shape parameters z Under the condition of NURBS curve X The function value at ; Bionic morphological constraints: in, is the bionic morphological constraint function; l is the head length; The third module is used to draw NURBS curves based on the design parameter vectors in the current sample set to describe the head shape; generate a three-dimensional head model based on the NURBS curves, and complete component assembly to form a water-air cross-medium robot model; The fourth module is used to reset the model of the current water-air cross-medium robot, its position, posture and motion parameters; The fifth module is used to automatically check and repair surface mesh problems in the imported model, generate the simulation domain, and perform meshing. The water pool and air use the preset mesh accuracy; the water surface, path, water-air cross-medium robot, and its surroundings are meshed based on the preset mesh accuracy. The sixth module is used to perform finite element solution on the current grid and extract solution data, which includes low-precision solution data and high-precision solution data; The seventh module is used to construct a low-quality data set based on the current low-precision solution data and the design parameter vectors in the current sample set; it is used to preliminarily screen the design parameter vectors in the sample set; The eighth module is used to perform finite element solution on the screened design parameter vector using a high-precision grid to obtain high-precision solution data. Based on the screened design parameter vector and the high-precision solution data, a high-quality dataset is constructed for training the surrogate model. The ninth module is used to obtain the optimal hyperparameters of the surrogate model through Bayesian optimization based on high-quality datasets; The tenth module is used to train the agent model based on the optimal hyperparameters; The eleventh module is used to construct the optimization target, use the multi-objective optimization algorithm to obtain the solution that satisfies the bionic morphological constraints and component constraints under the optimization target, and select the solution with large variance as the annotation sample; The twelfth module is used to determine whether the proxy model has converged. If it has converged, the optimal design parameters are output. If it has not converged, the samples are continued to be labeled and the high-quality data set is expanded, and the proxy model is retrained until the proxy model converges.

Images