A method and device for optimizing the size of a valve core of a rotary proportional directional control valve and a readable storage medium thereof

Abstract

This invention discloses a method, apparatus, and readable storage medium for optimizing the structural dimensions of a rotary proportional directional control valve core, relating to the field of hydraulic component optimization design technology. The valve core structural dimensions include the valve core step width and the thinnest wall thickness. A two-stage optimization strategy is employed to optimize the surrogate model, including: using a Bayesian optimization algorithm to quickly locate the global optimal region of the design variables; exploring and utilizing the acquisition function balance to screen out the effective optimization intervals for the valve core step width and the thinnest wall thickness; based on the optimal region determined in the first stage, using the gradient descent method for precise optimization; calculating the gradient of the objective function using the difference method; iteratively optimizing to obtain the optimal values ​​of the design variables; determining whether the flow rate corresponding to the optimization result meets the preset design flow rate requirements and whether the axial torque is within the preset safety threshold range; if not, repeating the iteration; if yes, outputting the optimal valve core step width and the thinnest wall thickness.

Description

Technical Field

[0001] This invention relates to the field of hydraulic component optimization design technology, and in particular to a method, apparatus and readable storage medium for optimizing the structural dimensions of a rotary proportional directional control valve core. Background Technology

[0002] As the core control element of a hydraulic control system, the servo valve's flow regulation accuracy and operational stability directly determine the overall performance of the hydraulic system. The valve core, as a key moving component of the servo valve, has its step width and minimum wall thickness as core structural parameters affecting flow characteristics and axial torque: the valve core step width directly changes the flow area, affecting flow regulation sensitivity; the valve core's minimum wall thickness not only relates to structural strength and reliability but also influences the magnitude of axial torque by altering the force distribution on the valve core.

[0003] The traditional valve core structure and size design has the following prominent problems:

[0004] The design process relies heavily on engineers’ experience, determining the valve core step width and wall thickness by looking up tables or using trial and error. There is a lack of systematic research on the coupling relationship between these two factors and flow rate and axial torque, making it difficult to achieve multi-objective collaborative optimization.

[0005] The optimization methods are too simplistic, often relying on single simulations or simple optimization algorithms. This results in either low global search efficiency or insufficient local optimization accuracy, leading to long design cycles and high R&D costs.

[0006] The lack of standardized optimization processes, the neglect of the coupling effect between fluid dynamics and structural mechanics, and the focus on optimizing only a single performance indicator make it difficult to balance flow requirements and structural stability, resulting in inconsistent product reliability.

[0007] In existing technologies, valve core structure optimization mostly focuses on adjusting a single parameter or simulating a single physics field. A two-stage collaborative optimization scheme targeting the valve core step width and wall thickness has not yet been proposed, failing to efficiently solve the technical challenge of balancing global search and precise local optimization. Therefore, a systematic and efficient optimization design method is urgently needed to achieve precise matching of valve core structure dimensions with flow rate and axial torque through scientific optimization strategies. Summary of the Invention

[0008] The main objective of this invention is to propose a method, apparatus, and readable storage medium for optimizing the structural dimensions of a rotary proportional directional control valve core, aiming to solve the technical problem of poor synergy in existing valve core optimization.

[0009] To achieve the above objectives, this invention proposes a method for optimizing the structural dimensions of a rotary proportional directional control valve core. The valve core structural dimensions include the valve core step width and the thinnest wall thickness of the valve core. The method includes:

[0010] Step 1: Use Latin hypercube sampling to obtain an initial dataset, which covers the range of values ​​for valve core step width and valve core thinnest wall thickness.

[0011] Step 2: Establish a parametric model of the rotary proportional directional control valve, setting the valve core step width and valve core thinnest wall thickness as design variables, and using flow rate and axial torque as the core response targets;

[0012] Step 3: Construct a multiphysics coupled simulation model, which includes a fluid dynamics model and a structural mechanics model, and configure the coupling conditions between the two models;

[0013] Step 4: Run the multiphysics coupling simulation model to obtain response values ​​with flow rate and axial torque as the core.

[0014] Step 5: Based on the initial dataset and the response values, construct a proxy model using an artificial neural network;

[0015] Step 6: Optimize the proxy model using a two-stage optimization strategy, which includes:

[0016] Step 61, first stage: Use Bayesian optimization algorithm to quickly locate the global optimal region of design variables, and use the acquisition function to balance exploration and utilization to screen out the effective optimization range of valve core step width and valve core thinnest wall thickness;

[0017] Step 62, Second Stage: Based on the optimal region determined in the first stage, the gradient descent method is used for precise optimization, the gradient of the objective function is calculated by the difference method, and the optimal value of the design variable is obtained by iterative optimization.

[0018] Step 7: Determine whether the flow rate corresponding to the optimization result meets the preset design flow rate requirements and whether the axial torque is within the preset safety threshold range; if not, return to steps 2 to 6 and repeat the iteration; if it meets the requirements, output the optimal valve core step width and the thinnest valve core wall thickness.

[0019] In one embodiment, the process of obtaining the initial dataset through Latin hypercube sampling in step 1 includes:

[0020] Step 11: Divide the probability distributions of the valve core step width and the valve core thinnest wall thickness into several equal probability intervals, the number of which is the same as the preset total number of samples;

[0021] Step 12: Randomly generate a sample point within each of the equal probability intervals;

[0022] Step 13: Randomly permutate the sample points generated for the valve core step width and the valve core thinnest wall thickness, and combine them to form multiple independent final samples, which constitute the initial dataset.

[0023] In one embodiment, the process of constructing the parameterized model in step 2 includes:

[0024] Step 21: Based on the actual structural dimensions of the rotary proportional directional control valve, construct three-dimensional models of core components such as valve core, valve sleeve, and valve body, and restore the stepped structure of the valve core, the geometric features of the internal flow channel, and the assembly details of the sealing elements.

[0025] Step 22: Remove redundant features through geometric cleaning and optimize the model structure through topological simplification. Extract the fluid domain involved in the flow of the medium and the solid domain that affects the mechanical properties of the structure to form a fully parameterized digital model. The digital model can dynamically update the structural morphology by adjusting the valve core step width and the valve core thinnest wall thickness parameters.

[0026] In one embodiment, the construction of the multiphysics coupling simulation model in step 3 includes:

[0027] Step 31: The fluid dynamics model adopts a polyhedral mesh discretization strategy adapted to the valve core flow channel, configures the k-ε turbulence model and pressure-based solver, and defines the physical properties of the hydraulic oil and the inlet and outlet boundary conditions.

[0028] Step 32: The structural mechanics model adopts a hexahedral mesh discretization strategy, selects a linear elastic material constitutive model, and defines the contact relationship and constraint conditions between the valve core and the valve sleeve.

[0029] Step 33: The coupling condition maps the wall pressure data of the fluid dynamics model to the interface of the structural domain through an interpolation algorithm, so as to realize the accurate transfer of fluid-structure load and simulate the coupling effect of valve core step width and wall thickness on flow rate and axial moment.

[0030] In one embodiment, the process of constructing the surrogate model using an artificial neural network in step 5 includes:

[0031] Step 51: Preprocess the initial dataset and response values ​​by eliminating dimensional differences and removing outlier samples through Z-score standardization;

[0032] Step 52: Divide the preprocessed data into training set, validation set and test set in a ratio of 7:1.5:1.5;

[0033] Step 53: Design a neural network structure with an input layer dimension of 2, 2-3 hidden layers, and an output layer dimension of 2. The ReLU function is used as the activation function for the hidden layers.

[0034] Step 54: Train the model using the Adam optimizer and mean squared error loss function, and prevent overfitting using Dropout regularization.

[0035] Step 55: Use the coefficient of determination (R²) and root mean square error (RMSE) to evaluate the accuracy of the surrogate model. When R² ≥ 0.95, the model is deemed to meet the usage requirements.

[0036] In one embodiment, the implementation of the Bayesian optimization algorithm in step 61 includes:

[0037] Step 611: Use the surrogate model as the probabilistic model and the lower confidence bound (LCB) as the acquisition function;

[0038] Step 612: Optimize the acquisition function by improving the whale algorithm to obtain new sampling points;

[0039] Step 613: Input the newly added sampling points into the multiphysics coupling simulation model to obtain response values, update the dataset and reconstruct the surrogate model;

[0040] Step 614: Repeat the above operation until the prediction error of the surrogate model converges, and output the global optimal region of the design variables.

[0041] In one embodiment, the implementation of the gradient descent method in step 6.2 includes:

[0042] Step 621: Use the optimal region center value output in the first stage as the initial parameter;

[0043] Step 622: Calculate the gradient of the objective function using the backward difference method. The gradient calculation formula is:

[0044]

[0045] Where h is the step size of the variable perturbation;

[0046] The value of h is shown in the formula:

[0047]

[0048] Where w is the scaling factor of the function variable x, and can be assigned a value according to actual needs.

[0049] Step 623: Update the design variables according to the formula:

[0050]

[0051] k represents the number of iterations.

[0052] Where α is the learning rate, initially set to 0.01, and dynamically adjusted according to the changes in the objective function during the iteration process;

[0053] Step 624: Stop optimization when the difference between the design variables of two adjacent iterations is less than 1e-5 or the preset number of iterations is reached.

[0054] In one embodiment, the judgment criteria in step 7 are: the flow rate is greater than or equal to the preset design flow rate, and the axial torque is less than or equal to the preset safety threshold; the preset design flow rate and the preset safety threshold can be adjusted according to the specific application scenario of the rotation proportional directional control valve.

[0055] Secondly, the present invention provides a device for optimizing the structural dimensions of a rotary proportional direction control valve core, comprising:

[0056] The data sampling module is used to obtain an initial dataset using Latin hypercube sampling, and the initial dataset covers the range of values ​​for valve core step width and valve core thinnest wall thickness.

[0057] The parametric modeling module is connected in communication with the data sampling module and is used to establish a parametric model of the rotary proportional directional control valve, setting the valve core step width and the valve core thinnest wall thickness as design variables.

[0058] The multiphysics coupling simulation module is communicatively connected to the parametric modeling module. It is used to construct a multiphysics coupling simulation model that includes a fluid dynamics model and a structural mechanics model, configure coupling conditions and run the simulation to obtain response values ​​with flow rate and axial torque as the core.

[0059] The surrogate model construction module is communicatively connected to the data sampling module and the multiphysics coupling simulation module, and is used to construct and verify the surrogate model based on the initial dataset and response values ​​through an artificial neural network.

[0060] The two-stage optimization module, which is communicatively connected to the proxy model construction module, includes a global region positioning unit and a precise optimization unit. The global region positioning unit uses a Bayesian optimization algorithm to locate the optimal region for the design variables. The precise optimization unit uses the gradient descent method to precisely optimize within the optimal region.

[0061] The iterative judgment and result output module is connected to the parametric modeling module and the two-stage optimization module respectively. It is used to judge whether the optimization result meets the design requirements. If not, it triggers each module to repeat the corresponding operation. If yes, it outputs the optimal valve core step width and the valve core thinnest wall thickness.

[0062] Thirdly, the present invention provides a readable storage medium storing a computer program, the computer program including program code, the program code being used to execute the above-described valve core structure size design method based on two-stage optimization.

[0063] The technical solution of this invention adopts a two-stage optimization method for optimization. In the first stage, Bayesian optimization driven by an improved whale algorithm is used to quickly lock the global optimal variable region, which solves the problem of insufficient global search capability of a single algorithm. In the second stage, gradient descent optimization based on the difference method achieves accurate optimization, which makes up for the deficiency of local accuracy in global search. The two stages work together to greatly improve the optimization efficiency and accuracy. At the same time, the improved whale algorithm introduces a strategy of "accepting inferior individuals with a preset probability", which further enhances the scientific nature and stability of global search.

[0064] Through closed-loop iterative judgment and optimization, the valve core structure size parameters can be continuously adjusted until the design requirements are met, ensuring that the final output valve core structure size is accurately matched with the flow rate and axial torque performance. This significantly improves the flow regulation accuracy, operational stability and structural reliability of the proportional directional control valve, shortens the product development cycle and reduces R&D costs, and provides a systematic and standardized optimization process for the valve core structure size design of the proportional directional control valve. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0066] Figure 1 A flowchart illustrating a method for optimizing the structural dimensions of a rotary proportional directional control valve core, provided by the present invention.

[0067] Figure 2 A schematic diagram of the valve core structure of a rotary proportional directional control valve provided by the present invention;

[0068] Figure 3 A flow curve diagram of a rotary proportional directional control valve provided by the present invention;

[0069] Figure 4 A pressure diagram of a rotary proportional directional control valve provided by the present invention;

[0070] Figure 5 A cross-sectional view of a rotary proportional directional control valve provided by the present invention;

[0071] Figure 6 A cross-sectional deformation cloud diagram of a rotational proportional directional control valve provided by the present invention;

[0072] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0074] It should be noted that if directional indicators (such as up, down, left, right, front, back, etc.) are involved in the embodiments of this invention, these directional indicators are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly. Unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0075] Furthermore, if the embodiments of the present invention involve descriptions using terms such as "first," "second," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Furthermore, the use of "and / or" or "and / or" throughout the text includes three parallel options; for example, "A and / or B" includes option A, option B, or options where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0076] To achieve the above objectives, this invention proposes a method for optimizing the structural dimensions of a rotary proportional directional control valve core. The valve core structural dimensions include the valve core step width and the thinnest wall thickness of the valve core. The method includes:

[0077] Step 1: Use Latin hypercube sampling to obtain the initial dataset, which covers the range of values ​​for valve core step width and valve core thinnest wall thickness.

[0078] Step 2: Establish a parametric model of the rotary proportional directional control valve, setting the valve core step width and valve core thinnest wall thickness as design variables, and using flow rate and axial torque as the core response targets;

[0079] Step 3: Construct a multiphysics coupled simulation model, which includes a fluid dynamics model and a structural mechanics model, and configure the coupling conditions between the two models;

[0080] Step 4: Run the multiphysics coupling simulation model to obtain response values ​​with flow rate and axial torque as the core.

[0081] Step 5: Based on the initial dataset and response values, construct a proxy model using an artificial neural network;

[0082] Step 6: Optimize the proxy model using a two-stage optimization strategy. The two-stage optimization strategy includes:

[0083] Step 61, first stage: Use Bayesian optimization algorithm to quickly locate the global optimal region of design variables, and use the acquisition function to balance exploration and utilization to screen out the effective optimization range of valve core step width and valve core thinnest wall thickness;

[0084] Step 62, Second Stage: Based on the optimal region determined in the first stage, the gradient descent method is used for precise optimization, the gradient of the objective function is calculated by the difference method, and the optimal value of the design variable is obtained by iterative optimization.

[0085] Step 7: Determine whether the flow rate corresponding to the optimization result meets the preset design flow rate requirements and whether the axial torque is within the preset safety threshold range; if not, return to steps 2 to 6 and repeat the iteration; if it meets the requirements, output the optimal valve core step width and the thinnest valve core wall thickness.

[0086] In one embodiment, the process of obtaining the initial dataset through Latin hypercube sampling in step 1 includes:

[0087] Step 11: Divide the probability distributions of the valve core step width and the valve core thinnest wall thickness into several equal probability intervals, with the number of equal probability intervals being the same as the preset total number of samples.

[0088] Step 12: Randomly generate a sample point within each equally probable interval;

[0089] Step 13: Randomly permutate the sample points generated for the valve core step width and the valve core thinnest wall thickness, and combine them to form multiple independent final samples, which constitute the initial dataset.

[0090] In one embodiment, the process of constructing the parameterized model in step 2 includes:

[0091] Step 21: Based on the actual structural dimensions of the rotary proportional directional control valve, construct three-dimensional models of core components such as valve core, valve sleeve, and valve body, and restore the stepped structure of the valve core, the geometric features of the internal flow channel, and the assembly details of the sealing elements.

[0092] Step 22: Remove redundant features through geometric cleanup and optimize the model structure through topology simplification. Extract the fluid domain involved in the flow of the medium and the solid domain that affects the mechanical properties of the structure to form a fully parameterized digital model. The digital model can dynamically update the structural morphology by adjusting the valve core step width and the valve core thinnest wall thickness parameters.

[0093] In one embodiment, the construction of the multiphysics coupling simulation model in step 3 includes:

[0094] Step 31: The fluid dynamics model adopts a polyhedral mesh discretization strategy adapted to the valve core flow channel, configures the k-ε turbulence model and pressure-based solver, and defines the physical properties of the hydraulic oil and the inlet and outlet boundary conditions.

[0095] Step 32: The structural mechanics model adopts a hexahedral mesh discretization strategy, selects a linear elastic material constitutive model, and defines the contact relationship and constraint conditions between the valve core and the valve sleeve.

[0096] Step 33: The coupling condition maps the wall pressure data of the fluid dynamics model to the interface of the structural domain through an interpolation algorithm, so as to realize the accurate transfer of fluid-structure load and simulate the coupling effect of valve core step width and wall thickness on flow rate and axial moment.

[0097] In one embodiment, the process of constructing the surrogate model using an artificial neural network in step 5 includes:

[0098] Step 51: Preprocess the initial dataset and response values ​​by eliminating dimensional differences and removing outlier samples through Z-score standardization;

[0099] Step 52: Divide the preprocessed data into training set, validation set and test set in a ratio of 7:1.5:1.5;

[0100] Step 53: Design a neural network structure with an input layer dimension of 2, 2-3 hidden layers, and an output layer dimension of 2. The ReLU function is used as the activation function for the hidden layers.

[0101] Step 54: Train the model using the Adam optimizer and mean squared error loss function, and prevent overfitting using Dropout regularization.

[0102] Step 55: Use the coefficient of determination (R²) and root mean square error (RMSE) to evaluate the accuracy of the surrogate model. When R² ≥ 0.95, the model is deemed to meet the usage requirements.

[0103] In one embodiment, the implementation of the Bayesian optimization algorithm in step 61 includes:

[0104] Step 611: Use the surrogate model as the probabilistic model and the lower confidence bound (LCB) as the acquisition function;

[0105] Step 612: Optimize the acquisition function by improving the whale algorithm to obtain new sampling points;

[0106] Step 613: Input the newly added sampling points into the multiphysics coupling simulation model to obtain response values, update the dataset and reconstruct the surrogate model;

[0107] Step 614: Repeat the above operation until the prediction error of the surrogate model converges, and output the global optimal region of the design variables.

[0108] In one embodiment, the implementation of the gradient descent method in step 6.2 includes:

[0109] Step 621: Use the optimal region center value output in the first stage as the initial parameter;

[0110] Step 622: Calculate the gradient of the objective function using the backward difference method. The gradient calculation formula is:

[0111]

[0112] Where h is the step size of the variable perturbation;

[0113] The value of h is shown in the formula:

[0114]

[0115] Here, w is the scaling factor of the function variable x, which can be assigned a value according to actual needs.

[0116] Step 623: Update the design variables according to the formula:

[0117]

[0118] k represents the number of iterations.

[0119] Where α is the learning rate, initially set to 0.01, and dynamically adjusted according to the changes in the objective function during the iteration process;

[0120] Step 624: Stop optimization when the difference between the design variables of two adjacent iterations is less than 1e-5 or the preset number of iterations is reached.

[0121] In one embodiment, the judgment criteria in step 7 are: the flow rate is greater than or equal to the preset design flow rate, and the axial torque is less than or equal to the preset safety threshold; the preset design flow rate and the preset safety threshold can be adjusted according to the specific application scenario of the rotation proportional directional control valve.

[0122] In one embodiment, the preset design flow rate is 20 L / min, and the preset safety threshold is 18 mN·m.

[0123] Secondly, the present invention provides a device for optimizing the structural dimensions of a rotary proportional direction control valve core, comprising:

[0124] The data sampling module is used to obtain the initial dataset using Latin hypercube sampling. The initial dataset covers the range of values ​​for valve core step width and valve core thinnest wall thickness.

[0125] The parametric modeling module communicates with the data sampling module and is used to establish a parametric model of the rotary proportional directional control valve, setting the valve core step width and the valve core thinnest wall thickness as design variables.

[0126] The multiphysics coupling simulation module communicates with the parametric modeling module to construct a multiphysics coupling simulation model that includes a fluid dynamics model and a structural mechanics model. It configures the coupling conditions and runs the simulation to obtain response values ​​with flow rate and axial torque as the core parameters.

[0127] The surrogate model construction module communicates with the data sampling module and the multiphysics coupling simulation module, respectively, and is used to construct and verify the surrogate model based on the initial dataset and response values ​​through an artificial neural network.

[0128] The two-stage optimization module communicates with the surrogate model construction module and includes a global region localization unit and a precise optimization unit. The global region localization unit uses the Bayesian optimization algorithm to locate the optimal region for the design variables. The precise optimization unit uses the gradient descent method to precisely optimize within the optimal region.

[0129] The iterative judgment and result output module communicates with the parametric modeling module and the two-stage optimization module respectively. It is used to judge whether the optimization result meets the design requirements. If not, it triggers each module to repeat the corresponding operation. If yes, it outputs the optimal valve core step width and the valve core thinnest wall thickness.

[0130] Thirdly, the present invention provides a readable storage medium storing a computer program, the computer program including program code, the program code being used to execute the above-described valve core structure size design method based on two-stage optimization.

[0131] The technical solution of this invention adopts a two-stage optimization method for optimization. In the first stage, Bayesian optimization driven by an improved whale algorithm is used to quickly lock the global optimal variable region, which solves the problem of insufficient global search capability of a single algorithm. In the second stage, gradient descent optimization based on the difference method achieves accurate optimization, which makes up for the deficiency of local accuracy in global search. The two stages work together to greatly improve the optimization efficiency and accuracy. At the same time, the improved whale algorithm introduces a strategy of "accepting inferior individuals with a preset probability", which further enhances the scientific nature and stability of global search.

[0132] Through closed-loop iterative judgment and optimization, the valve core structure size parameters can be continuously adjusted until the design requirements are met, ensuring that the final output valve core structure size is accurately matched with the flow rate and axial torque performance. This significantly improves the flow regulation accuracy, operational stability and structural reliability of the proportional directional control valve, shortens the product development cycle and reduces R&D costs, and provides a systematic and standardized optimization process for the valve core structure size design of the proportional directional control valve.

[0133] The present invention will be further described in detail below with reference to specific embodiments:

[0134] 1. Initial Data Set Acquisition

[0135] 1. Latin hypercube sampling yields the initial dataset:

[0136] The core of Latin hypercube sampling (LHS) is stratified sampling, which can uniformly cover the range of values ​​of all variables with a small number of samples, greatly improving sampling efficiency.

[0137] Layered interval calculation:

[0138] The purpose of this step is to divide the probability distribution of each input variable into N equally probable intervals (N is the total number of samples to be drawn).

[0139] If variable X follows a uniform distribution on the interval [a, b], the interval is directly divided into equal parts according to the length.

[0140] The probability of each interval is:

[0141]

[0142] No. The range of each interval is:

[0143]

[0144] in, , where a is the minimum value of the variable and b is the maximum value of the variable.

[0145] Generate N random numbers in the range [0,1]. Then, generate random points within the interval:

[0146]

[0147] The N random points generated for each variable are randomly permuted and combined into N final samples.

[0148] Specifically, the valve core step width (x1) is set to range from 0.15 mm to 0.45 mm, and the valve core thinnest wall thickness (x2) is set to range from 0.8 mm to 1.0 mm. Latin hypercube sampling is used to divide the probability distributions of the two design variables into 30 equally probable intervals. One random sample point is generated for each interval, and these are combined through permutation to obtain 30 initial samples, forming the initial dataset.

[0149] 2. Parametric Model Construction

[0150] Based on the actual dimensional parameters of a certain type of rotary proportional directional control valve, 3D models of core components such as the valve core, valve sleeve, and valve body are constructed in 3D modeling software. This process recreates the valve core's stepped structure (width x1, wall thickness x2), the curvature of the internal flow channel bends, the gradual change in cross-sectional dimensions, and the O-ring sealing structure. Redundant chamfer features are removed through geometric cleanup, and the model structure is optimized through topology simplification. The fluid domain of the flow channel and the solid domain of the valve core are extracted to generate a fully parameterized digital model that supports dynamic adjustment of parameters x1 and x2.

[0151] 3. Construction and operation of multiphysics coupling simulation models

[0152] 3.1 Fluid Dynamics Model: A polyhedral mesh was used to discretize the fluid domain, with a global mesh size of 0.1 mm to 2.0 mm, and a local mesh size of 0.1 mm in the throttling region of the flow channel. A k-ε turbulence model was configured, which, through an improved ε equation solution strategy, has better prediction accuracy than the standard k-ε model for separated flows (such as backflow in the diffuser section of the flow channel) and rotating flows under strong pressure gradients, and is more suitable for complex flow regimes within the valve. An extensible wall function was used for near-wall processing, balancing computational efficiency in the high Reynolds number region with analytical accuracy in the low Reynolds number region. The hydraulic oil physical properties were set to a density of 850 kg / m³ and a dynamic viscosity of 0.0391 Pa·s; the inlet was set to a pressure inlet (7 MPa), the outlet to a pressure outlet (0 MPa), and the valve core surface was set to a non-slip wall surface.

[0153] 3.2 Structural Mechanics Model: The solid domain of the valve core is discretized using a hexahedral mesh with a mesh size of 0.2 mm to 0.5 mm, and locally refined to 0.2 mm at the root of the step; the valve core material is 440C steel with an elastic modulus of 200 GPa and a Poisson's ratio of 0.28; the contact relationship between the valve core and the valve sleeve is defined with a friction coefficient of 0.08, and axial translation constraints are set at both ends of the valve core;

[0154] The final result is approximately 3.93 million high-quality mesh elements, with a mesh Jacobian ratio better than 0.7 (far exceeding the acceptable threshold of 0.5), effectively avoiding stress calculation errors caused by mesh distortion.

[0155] 3.3 Coupling Condition Configuration: The fluid domain wall pressure data is mapped to the solid domain interface using a radial basis function interpolation algorithm to achieve fluid-solid load transfer. The simulation model was run, and the flow rate (Q) and axial moment (M) response values ​​for 30 sets of samples were obtained. Some sample data are shown in the table below:

[0156]

[0157] 4. Proxy Model Construction

[0158] The initial dataset and response values ​​were preprocessed as follows: Z-score standardization was used to eliminate dimensional differences, and one group of outlier samples was removed using box plotting. The dataset was then divided into training (20 groups), validation (4 groups), and test (5 groups) sets in a 7:1.5:1.5 ratio. The neural network structure was designed as follows: input layer with 2 neurons (x1, x2), hidden layers with 2 layers (32 neurons in the first layer and 16 neurons in the second layer), and output layer with 2 neurons (Q, M). The ReLU activation function was used for the hidden layers. The Adam optimizer (learning rate 0.001) and mean squared error loss function were selected, and the dataset was trained for 500 epochs. Dropout (probability 0.2) was used to prevent overfitting.

[0159] The verification results show that the surrogate model has a determination coefficient R² = 0.97, a root mean square error RMSE = 0.32 L / min (flow rate) and 0.45 mN·m (axial torque), which meets the usage requirements.

[0160] 5. Two-stage optimization

[0161] 5.1 First Stage: Bayesian Optimization to Locate the Global Optimal Region. The constructed surrogate model is used as the probabilistic model, and the lower confidence bound (LCB) is used as the acquisition function. The formula is as follows:

[0162]

[0163] Where k is set to 2.5, the acquisition function is optimized using an improved whale algorithm, the population size is set to 50, and the number of iterations is set to 30. One new sampling point is added in each iteration, the response value is obtained from the multiphysics coupled simulation model, the dataset is updated, and the surrogate model is reconstructed. After 30 iterations, the prediction error of the surrogate model converges, and the globally optimal region is output: x1∈0.25 mm~0.35 mm, x2∈0.95 mm~1.0 mm.

[0164] It should be noted that, in order to optimize and improve the initial structure, the optimization objective function is constructed by focusing on the quantitative control of key performance indicators, selecting the maximum bolt stress and the maximum deformation of the oil port at the bottom of the valve body as the optimization objectives:

[0165]

[0166] in, Overall goal For the response values ​​of different indicators, .

[0167] Since the maximum bolt stress (unit: MPa) and the maximum port deformation (unit: mm) are performance indicators with different physical dimensions, their numerical magnitudes differ significantly. Therefore, it is necessary to first normalize the two objective functions by introducing a dimensionless transformation:

[0168]

[0169] Using artificial neural networks (ANNs) to build surrogate models is a method that simplifies the approximation of the behavior of complex systems. The core idea is to leverage the nonlinear fitting capabilities of ANNs to replace computationally expensive and time-consuming original models (such as physical simulations and finite element analyses), enabling rapid prediction and analysis. The specific construction approach and steps are as follows:

[0170] Design variables: These are the key parameters that affect the system output, denoted as... The dimension is n.

[0171] Response value: The output of the original model, denoted as The dimension is m.

[0172] The performance of ANNs depends on the quality of the training data. Representative input-output samples need to be obtained from the original model. The sampling method used is Latin hypercube sampling as described in section 1. A balance needs to be struck between accuracy and cost (the computational time of the original model). Typically, the sample size is 10-100 times the input dimension. The final dataset is obtained as follows:

[0173]

[0174] Where N is the sample size.

[0175] Data preprocessing (eliminating the impact of differences in data volume on ANN training), common operations:

[0176] Normalization: Maps data to the interval [0,1] or [-1,1].

[0177] Standardization: Transforming data into a distribution with a mean of 0 and a variance of 1, such as... ;

[0178] : Raw data (a feature value of a single sample).

[0179] The mean of this set of data (corresponding to the feature) reflects the central tendency of the data.

[0180] The standard deviation of this set of data (corresponding to the feature) reflects the degree of dispersion of the data.

[0181] Outlier handling: Outlier samples are removed using methods such as box plots and Z-scores to avoid interfering with training;

[0182] Dataset splitting: Divide D into a training set (70-80%, used for fitting), a validation set (10-15%, used to monitor overfitting), and a test set (10-15%, used for final evaluation).

[0183] Neural network architecture design

[0184] Designing the network topology based on problem complexity involves the core configuration of the input layer, hidden layer, and output layer: Input layer: Number of neurons = dimension n of the input variables, directly receives the preprocessed input. .

[0185] Hidden layer:

[0186] Number of layers: 1-3 layers are commonly used (shallow networks are easy to train, while deep networks may overfit).

[0187] Number of neurons: typically 2-10 times the input dimension (can be adjusted through grid search / Bayesian optimization), but too many should be avoided to prevent overfitting.

[0188] Output layer: Number of neurons = dimension m of output variable. The activation function is selected according to the output range: If the output is a continuous value (such as a regression problem), the linear activation function (no activation) is commonly used; if the output is in [0,1], the sigmoid function can be used.

[0189] Activation functions: ReLU (to alleviate gradient vanishing and is computationally fast) and tanh (for symmetric output) are commonly used in hidden layers, while sigmoid (which is prone to gradient vanishing in deeper layers) should be avoided.

[0190] Model training and optimization:

[0191] Minimize prediction error using the backpropagation algorithm; adjust core parameters as follows:

[0192] Loss function: Mean squared error (MSE) is commonly used in regression problems. ), mean absolute error (MAE), or weighted loss (for unbalanced outputs).

[0193] Optimizer: Adam (adaptive learning rate, fast convergence) is preferred, followed by SGD (learning rate needs to be manually adjusted).

[0194] Regularization: Prevents overfitting.

[0195] L2 regularization (weight decay): limits the size of weights;

[0196] Dropout: Randomly discards some neurons during training (e.g., with a probability of 0.2-0.5).

[0197] Early Stopping: Stop training when the validation set loss does not decrease for several consecutive rounds.

[0198] Learning rate scheduling: The initial learning rate is set to 0.001~0.0001, and it is decayed in the later stages of training (decaying to 0.5 of the original rate every 10 rounds) to avoid oscillations.

[0199] Model validation and evaluation:

[0200] Use a test set to evaluate the accuracy of the proxy model. Common metrics include:

[0201] Coefficient of determination ( The closer the result is to 1, the better the fit. ;

[0202] The true response value of a single sample;

[0203] All samples The average value;

[0204] The predicted response value of the surrogate model (constructed by an artificial neural network) to this sample.

[0205] Root mean square error (RMSE): reflects the magnitude of the error. ;

[0206] Mean Absolute Percentage Error (MAPE): Suitable for assessing relative error. ;

[0207] Cross-validation: K-fold cross-validation (such as 5-fold cross-validation) reduces the randomness of data partitioning and more robustly evaluates generalization ability.

[0208] The Whale Optimization Algorithm (WOA) is primarily used to solve optimization problems and is applicable to both continuous and discrete optimization. Its core idea is to simulate the foraging behavior of humpback whales, and it mainly includes three operations:

[0209] (1) Encircle the prey

[0210] Once a pod of whales knows the location of its prey, it adjusts its position based on its relative position to the prey. This behavior can be represented by the following formula:

[0211]

[0212]

[0213] in, The distance between the whale and its prey is represented by t, where t represents the current iteration number. and These represent the solutions for the present and the next generation, respectively. This indicates the position of the optimal candidate solution prior to the present era. and The coefficient vector is shown in the following formula:

[0214]

[0215]

[0216] in, and Represents two random numbers between 0 and 1. The number that decreases linearly to 0 as it iterates is represented by the following formula:

[0217]

[0218] Where G represents the total number of iterations.

[0219] (2) Spiral attack

[0220] To simulate the spiral movement of humpback whales, specifically their bubble-net attack, a spiral equation was constructed between the whale and its prey to reproduce the whale's spiral trajectory. The formula is shown below:

[0221]

[0222]

[0223] Where b is a constant, usually set to 1. For a random number between [-1, 1], the parameter update is shown in the formula:

[0224]

[0225]

[0226] in, A random number between 0 and 1.

[0227] The choice between spiral attack and surrounding the prey is determined randomly. When the probability p is less than 0.5, surrounding the prey is chosen; otherwise, spiral attack is chosen. The expression is shown in the equation:

[0228]

[0229] (3) Random search

[0230] Random search is used in the whale optimization algorithm to perform a global search to help escape local optima, and its representation is shown in the equation:

[0231]

[0232]

[0233] in, For the randomly selected whale, the formula for random search is very similar to the formula for surrounding prey, the only difference being that the direction is changed from the current optimal candidate solution to a randomly selected whale. The choice between surrounding prey and random search depends on... The value when | When |<1, the prey-surrounding behavior is adopted; when | If |≥1, then a random search behavior is selected.

[0234] The following improvements were made to the whale algorithm:

[0235] In the standard Whale Optimization algorithm, it is prone to prematurely getting trapped in local optima during iterations. Its main update mechanism relies on the position of the best individual in the population, causing other individuals to tend to search around the optimal solution, thus exacerbating population clustering and rapidly reducing population diversity. While the algorithm's "shrinking encirclement mechanism" and "spiral update" strategies are beneficial for quickly approaching the global optimum in the early stages, once the best individual gets trapped in a local optimum, the entire population guided by it struggles to escape the trap. Compared to Particle Swarm Optimization or Artificial Bee Colony Optimization, WOA lacks dynamic weights, velocity perturbations, or mutation operations to maintain the population's exploratory capabilities. Although WOA has an "exploration phase," this phase only controls the switching between exploration and exploitation by controlling parameters A and C, lacking finer-grained adjustments to inter-individual differences and failing to effectively control population diversity. In WOA, individual updates rely only on the current optimal solution, rather than multiple globally optimal solutions. This update method limits the propagation of information among individuals, making the population prone to forming a single search direction. To alleviate this problem, a strategy of accepting inferior individuals with a certain probability is introduced to enhance the algorithm's global search capability. Accepting inferior individuals with a certain probability refers to selectively accepting new solutions with poor fitness during the individual update process. Suppose we have a new candidate solution. Its fitness value is And the current solution fitness value If the new solution has better fitness, that is... If the new solution has poor fitness, then accept the solution directly. If so, the solution will be accepted with a certain probability.

[0236] During the update process, the steps for accepting inferior solutions with a certain probability are as follows: (1) Generate new solutions: Generate new candidate solutions for the current individual. (2) Calculate fitness difference: Calculate the fitness difference between the new solution and the current solution. (3) Judge the quality: Compare the fitness of the new solution and the current solution. If the new solution is better than the current solution, it is accepted directly. (4) Accept inferior solutions with a certain probability: If the new solution is poor, a random number is generated. ,when If the solution is poor, accept it; otherwise, keep the current solution unchanged.

[0237] The acceptance probability can be determined by drawing inspiration from the simulated annealing algorithm. Simulated annealing is a global optimization method for solving complex optimization problems, inspired by the annealing process in physics. Annealing is a metal heat treatment technique that slowly cools a system to a stable state with the lowest energy. Analogously, the simulated annealing algorithm performs a random search in the solution space and introduces a gradually decreasing "temperature" parameter to adjust the search range, thereby effectively avoiding getting trapped in local optima and improving the ability to find the global optimum. The formula is as follows:

[0238]

[0239]

[0240] The probability of accepting a "worse solution";

[0241] The energy difference between the new solution and the current solution;

[0242] Current temperature

[0243] The probability of accepting a "worse objective function value";

[0244] The objective function value of the new solution;

[0245] The objective function value of the current solution;

[0246] The flowchart of the improved whale algorithm is as follows: Figure 4 As shown.

[0247] 5.2 Second Stage: Precise Optimization Using Gradient Descent. Using the global optimal region center values ​​(x1 = 0.3 mm, x2 = 0.975 mm) as initial parameters, the gradient of the objective function is calculated using the backward differencing method, with a perturbation step size h = 0.001 mm. The initial learning rate α = 0.01. The learning rate doubles when the objective function value decreases during iteration and halves when the objective function value increases. After 20 iterations, the difference between two adjacent design variables is less than 1e-5, at which point the optimization stops, yielding the optimal solution: x1 = 0.3 mm, x2 = 1.0 mm.

[0248] It should be noted that gradient descent calculates the gradient (i.e., the first derivative) of the objective function and then updates the parameters in the opposite direction of the gradient, thereby gradually approaching the minimum value.

[0249] The gradient descent method combined with the post-difference method has the following main process: (1) Initialize parameters: Select the initial values ​​of the model parameters. Usually, these parameters are randomly initialized, but in this paper, considering the optimization method of the first stage, the initial values ​​are set to the optimal values ​​after the first stage of optimization. (2) Calculate the objective function value: Calculate the value of the objective function based on the current parameter values. (3) Calculate the gradient: Calculate the gradient value of the current objective function using the difference method, which indicates the direction of parameter update. (4) Update parameters: Update the parameters according to a certain learning rate based on the calculated gradient. (5) Check convergence: Determine whether the parameters have converged. If the change of the objective function is very small, or the set number of iterations is reached, stop training; otherwise, return to step (2) and continue iterating.

[0250] Update the design variables according to the formula:

[0251]

[0252] k represents the number of iterations.

[0253] Where α is the learning rate, initially set to 0.01, and dynamically adjusted according to the changes in the objective function during the iteration process;

[0254] x is the model parameter. It is the gradient of the objective function with respect to the parameter x.

[0255] The learning rate α is a key hyperparameter in gradient descent, determining the step size of parameter updates in each iteration and significantly impacting model convergence and optimization performance. If the learning rate is set too large, the objective function may diverge during iterations; excessively large variable update steps could not only miss the optimal solution but also cause oscillations or even divergence. Conversely, if the learning rate is too small, the convergence speed may slow down significantly, increasing overall training time and potentially leading to getting stuck in local optima or oscillating slowly around the optimal solution, making rapid convergence difficult. Therefore, choosing a reasonable learning rate is crucial for ensuring the stability and efficiency of the optimization process.

[0256] The learning efficiency definition formula used in this article is:

[0257]

[0258] in, This is the reduction factor of the learning rate relative to the variable x, and can be assigned a value according to the specific problem. During the iteration process of the minimization problem, the learning rate changes as follows:

[0259]

[0260] in, It is the value of the objective function before updating the variables. It refers to the objective function value before and after updating the variable. When the function value decreases, the learning rate doubles. When the function value increases, the learning rate halves.

[0261] 6. Optimize result verification

[0262] The optimal solution (x1=0.3mm, x2=1.0mm) was input into a multiphysics coupled simulation model for verification, yielding a flow rate Q=27.2L / min and an axial moment M=16.9mN·m. This result meets the preset design requirements (flow rate ≥20 L / min, axial moment ≤18 mN·m).

[0263] This invention achieves efficient and precise design of valve core step width and wall thickness through Latin hypercube sampling, parametric modeling, multiphysics coupled simulation, surrogate model construction, and a two-stage optimization strategy. Compared with traditional design methods, this invention has the following advantages:

[0264] Optimization efficiency: The two-stage optimization strategy shortens the design cycle;

[0265] Optimized accuracy: The prediction error for flow rate and axial moment is ≤5%, which is 30% more accurate than traditional methods;

[0266] Practicality: The optimized valve core structure dimensions balance flow regulation requirements with structural stability;

[0267] Versatility: This method can be extended to the valve core design of different models of rotary proportional directional control valves, and has broad application prospects.

[0268] The two-stage optimization strategy of this invention not only solves the technical problem of valve core structure size design, but also provides new ideas and methods for multi-objective optimization of other mechanical parts, and has important engineering application value.

[0269] Definitions:

[0270] Z-score standardization (also known as standard score): measures the degree of deviation of a specific data point from the mean of its dataset, and is expressed in units of standard deviation.

[0271] The formula for calculating the Z-score is:

[0272]

[0273] in:

[0274] : The Z-score value of this data point.

[0275] : The original values ​​of the data points that need to be calculated.

[0276] : The mean of the dataset.

[0277] : Standard deviation of the dataset.

[0278] It should be understood that the terms "one embodiment" or "one example" throughout the specification mean that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in one example" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Those skilled in the art should also recognize that the embodiments described in the specification are optional embodiments, and the actions and modules involved are not necessarily essential to the invention.

[0279] In various embodiments of the present invention, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0280] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It is particularly important to note that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0281] The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for optimizing the structural dimensions of a rotary proportional directional control valve core, characterized in that: The valve core structural dimensions include the valve core step width and the valve core thinnest wall thickness; the method includes: Step 1: Use Latin hypercube sampling to obtain an initial dataset, which covers the range of values ​​for valve core step width and valve core thinnest wall thickness. Step 2: Establish a parametric model of the rotary proportional directional control valve, setting the valve core step width and valve core thinnest wall thickness as design variables, and using flow rate and axial torque as the core response targets; Step 3: Construct a multiphysics coupled simulation model, which includes a fluid dynamics model and a structural mechanics model, and configure the coupling conditions between the two models; Step 4: Run the multiphysics coupling simulation model to obtain response values ​​with flow rate and axial torque as the core. Step 5: Based on the initial dataset and the response values, construct a proxy model using an artificial neural network; Step 6: Optimize the proxy model using a two-stage optimization strategy, which includes: Step 61, first stage: Use Bayesian optimization algorithm to quickly locate the global optimal region of design variables, and use the acquisition function to balance exploration and utilization to screen out the effective optimization range of valve core step width and valve core thinnest wall thickness; Step 62, Second Stage: Based on the optimal region determined in the first stage, the gradient descent method is used for precise optimization, the gradient of the objective function is calculated by the difference method, and the optimal value of the design variable is obtained by iterative optimization. Step 7: Determine whether the flow rate corresponding to the optimization result meets the preset design flow rate requirements and whether the axial torque is within the preset safety threshold range; if not, return to steps 2 to 6 and repeat the iteration; if it meets the requirements, output the optimal valve core step width and the thinnest valve core wall thickness.

2. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The process of obtaining the initial dataset by Latin hypercube sampling in step 1 includes: Step 11: Divide the probability distributions of the valve core step width and the valve core thinnest wall thickness into several equal probability intervals, the number of which is the same as the preset total number of samples; Step 12: Randomly generate a sample point within each of the equal probability intervals; Step 13: Randomly permutate the sample points generated for the valve core step width and the valve core thinnest wall thickness, and combine them to form multiple independent final samples, which constitute the initial dataset.

3. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The process of constructing the parameterized model in step 2 includes: Step 21: Based on the actual structural dimensions of the rotary proportional directional control valve, construct three-dimensional models of core components such as valve core, valve sleeve, and valve body, and restore the stepped structure of the valve core, the geometric features of the internal flow channel, and the assembly details of the sealing elements. Step 22: Remove redundant features through geometric cleaning and optimize the model structure through topological simplification. Extract the fluid domain involved in the flow of the medium and the solid domain that affects the mechanical properties of the structure to form a fully parameterized digital model. The digital model can dynamically update the structural morphology by adjusting the valve core step width and the valve core thinnest wall thickness parameters.

4. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The construction of the multiphysics coupling simulation model in step 3 includes: Step 31: The fluid dynamics model adopts a polyhedral mesh discretization strategy adapted to the valve core flow channel, configures the k-ε turbulence model and pressure-based solver, and defines the physical properties of the hydraulic oil and the inlet and outlet boundary conditions. Step 32: The structural mechanics model adopts a hexahedral mesh discretization strategy, selects a linear elastic material constitutive model, and defines the contact relationship and constraint conditions between the valve core and the valve sleeve. Step 33: The coupling condition maps the wall pressure data of the fluid dynamics model to the interface of the structural domain through an interpolation algorithm, so as to realize the accurate transfer of fluid-structure load and simulate the coupling effect of valve core step width and wall thickness on flow rate and axial moment.

5. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The process of constructing the surrogate model using the artificial neural network in step 5 includes: Step 51: Preprocess the initial dataset and response values ​​by eliminating dimensional differences and removing outlier samples through Z-score standardization; Step 52: Divide the preprocessed data into training set, validation set and test set in a ratio of 7:1.5:1.5; Step 53: Design a neural network structure with an input layer dimension of 2, 2-3 hidden layers, and an output layer dimension of 2. The ReLU function is used as the activation function for the hidden layers. Step 54: Train the model using the Adam optimizer and mean squared error loss function, and prevent overfitting using Dropout regularization. Step 55: Use the coefficient of determination (R²) and root mean square error (RMSE) to evaluate the accuracy of the surrogate model. When R² ≥ 0.95, the model is deemed to meet the usage requirements.

6. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The implementation of the Bayesian optimization algorithm in step 61 includes: Step 611: Use the surrogate model as the probabilistic model and the lower confidence bound (LCB) as the acquisition function; Step 612: Optimize the acquisition function by improving the whale algorithm to obtain new sampling points; Step 613: Input the newly added sampling points into the multiphysics coupling simulation model to obtain response values, update the dataset and reconstruct the surrogate model; Step 614: Repeat the above operation until the prediction error of the surrogate model converges, and output the global optimal region of the design variables.

7. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The implementation of the gradient descent method in step 6.2 includes: Step 621: Use the optimal region center value output in the first stage as the initial parameter; Step 622: Calculate the gradient of the objective function using the backward difference method. The gradient calculation formula is: ； Where h is the step size of the variable perturbation; The value of h is shown in the formula: ； Where w is the scaling factor of the function variable x, and can be assigned a value according to actual needs. Step 623: Update the design variables according to the formula: ， k represents the number of iterations. Where α is the learning rate, initially set to 0.01, and dynamically adjusted according to the changes in the objective function during the iteration process; Step 624: Stop optimization when the difference between the design variables of two adjacent iterations is less than 1e-5 or the preset number of iterations is reached.

8. The method for optimizing the structural dimensions of a rotary proportional directional control valve core as described in claim 1, characterized in that: The judgment criteria in step 7 are: the flow rate is greater than or equal to the preset design flow rate, and the axial torque is less than or equal to the preset safety threshold; the preset design flow rate and the preset safety threshold can be adjusted according to the specific application scenario of the rotation proportional directional control valve.

9. A device for optimizing the structural dimensions of a rotary proportional directional control valve core, characterized in that, include: The data sampling module is used to obtain an initial dataset using Latin hypercube sampling, and the initial dataset covers the range of values ​​for valve core step width and valve core thinnest wall thickness. The parametric modeling module, which is connected to the data sampling module, is used to establish a parametric model of the rotary proportional directional control valve, and sets the valve core step width and the valve core thinnest wall thickness as design variables. The multiphysics coupling simulation module is communicatively connected to the parametric modeling module. It is used to construct a multiphysics coupling simulation model that includes a fluid dynamics model and a structural mechanics model, configure coupling conditions and run the simulation to obtain response values ​​with flow rate and axial torque as the core. The surrogate model construction module is communicatively connected to the data sampling module and the multiphysics coupling simulation module, and is used to construct and verify the surrogate model based on the initial dataset and response values ​​through an artificial neural network. The two-stage optimization module, which is communicatively connected to the proxy model construction module, includes a global region positioning unit and a precise optimization unit. The global region positioning unit uses a Bayesian optimization algorithm to locate the optimal region for the design variables. The precise optimization unit uses the gradient descent method to precisely optimize within the optimal region. The iterative judgment and result output module is connected to the parametric modeling module and the two-stage optimization module respectively. It is used to judge whether the optimization result meets the design requirements. If not, it triggers each module to repeat the corresponding operation. If yes, it outputs the optimal valve core step width and the valve core thinnest wall thickness.

10. A readable storage medium, characterized in that: The storage medium stores a computer program, which includes program code for executing the valve core structure size design method based on two-stage optimization as described in any one of claims 1 to 8.

Images