Intelligent design method of component
The physical information neural network (PINN) model trained by artificial intelligence replaces traditional CFD computing and automatically optimizes design parameters, solving the problem of component design relies on personal capabilities and long design cycles in the existing technology, and achieving a fast and efficient component design process .
Patent Information
- Application Number
- CN202411419056.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2024-12-31
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing component design methods rely heavily on the designer's personal abilities, have high labor intensity, long design cycle, high labor costs, and long CFD calculation time, making it impossible to quickly form an ideal design solution.
Using artificial intelligence methods to train a physical information neural network (PINN) model that replaces CFD computing, trains a large amount of simulation and experimental data to form an alternative model, and uses optimization algorithms to automatically iterate the design parameters to generate the best design solution, and replace the traditional CFD calculation process.
The time and labor cost required for component design is greatly shortened and the design efficiency is improved. The PINN model can quickly generate prediction results and have stronger generalization capabilities, which are suitable for similar but slightly different problems.
Smart Images
Figure CN119227545A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of construction design technology and intelligent design technology, and in particular to an intelligent design method for a component. Background Art
[0002] In production practice, many components are used to form a specified flow field to achieve the purpose of improving production efficiency and reducing equipment damage. For example, the duct around the ship propeller is used to improve the wake formed by the propeller and improve the propulsion efficiency of the propeller; the rectifying plate in the water pipe mouth is used to improve the flow velocity and distribution of the outlet fluid to prevent other equipment from being damaged by the impact of high-speed fluid.
[0003] At present, the design of such components often relies on the personal experience of designers, and each time a design is completed, it is necessary to obtain the corresponding flow field information through computational fluid dynamics (CFD) simulation calculation, and then compare the calculated flow field information with the expected flow field information, improve the design of the component according to personal ability, and perform CFD calculation again. After multiple cycles, the final design solution is formed. This design method is heavily dependent on the personal ability of the designer, with high labor intensity and high labor cost. At the same time, each CFD calculation often takes a long time, and it is impossible to form a more ideal design solution in a short time. Summary of the invention
[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an intelligent design method for components to solve the technical problems that the current component design method is heavily dependent on the personal ability of the designer, has high labor intensity, long design cycle and high labor cost.
[0005] In order to solve the above technical problems, the specific technical solutions adopted by the present invention are as follows:
[0006] An intelligent design method for a component that generates a specified flow field comprises the steps of:
[0007] S1. Use artificial intelligence methods to train a large amount of simulation and test data to form an alternative model to replace CFD calculations;
[0008] S2. Obtain design variable parameters of the component;
[0009] S3. Assign initial values to the variables of the parametric design, use the alternative model to quickly generate the target flow field, compare the results of the flow field with the results of the target flow field, calculate the overall root mean square error (RMSE) of the two flow fields, and use the optimization algorithm to automatically iteratively optimize the design parameter variables with the goal of minimizing the root mean square error (RMSE). Finally, the best combination of parameter variables is obtained and the design scheme of the component is generated using the best combination of parameter variables.
[0010] Furthermore, the alternative model for replacing CFD calculation in step S1 includes training using a physical information neural network (PINN) method to ultimately obtain an alternative model for replacing CFD calculation with a specified flow field under constraint conditions.
[0011] Furthermore, the training method using the physical information neural network (PINN) method in step S1 includes:
[0012] S11, sampling within the design range of the component design variable parameters to obtain a design variable parameter data set uniformly distributed in the design variable parameter space;
[0013] S12, calculating the performance parameters of the component according to the flow field data in the flow field data set, and using the flow field data and the performance parameters as label data for training the PINN model;
[0014] S13, randomly sampling a preset number of grid points in the flow field area, and taking the flow field data set and the preset number of grid points as the data set;
[0015] S14. Randomly extract 90% of the data in the data set as the training set, and the remaining 10% of the data as the test set. Use the training set and the test set to train the PINN model to obtain a trained PINN model.
[0016] Furthermore, the method of training the PINN model using the training set and the test set includes:
[0017] A1. The control equations of fluid flow, namely the continuity equation and the NS equation, are used as the prior physical information of PINN, which includes:
[0018]
[0019] In the formula, is the Hamiltonian operator; u=(u,v,w) is the fluid velocity, ρ, v are the fluid density and kinematic viscosity coefficient respectively;
[0020] A2. Set the network structure of the constructed PINN - PINN consists of a fully connected neural network and a residual network. The fully connected neural network consists of 8 layers of neural networks, including 1 input layer, 6 hidden layers and 1 output layer. The input layer and the output layer each contain 5 neurons, each hidden layer contains 30 neurons, and a fully connected network is used between layers;
[0021] Furthermore, the network weights w and biases b of each layer of PINN are initialized using the He initialization method or the orthogonal initialization method, the activation function of each layer of neurons uses the hyperbolic tangent function, and the optimization function is a multi-objective optimization function;
[0022] A3. During the training process, the loss function of PINN is optimized by stochastic gradient descent combined with particle swarm optimization algorithm, and the weight w and bias b are updated;
[0023] A4. The predicted value obtained when the loss function value of PINN is reduced to the set optimization target. At this time, the PINN network parameters are the flow field reconstruction alternative model of PINN that replaces the CFD model.
[0024] Furthermore, optimizing the design variable parameters with the goal of minimizing the overall mean square error RMSE also includes taking the minimum root mean square error RMSE as the goal, automatically iteratively optimizing the design variable parameters using optimization algorithms such as particle swarm and simulated annealing, and finally obtaining the best design variable parameter combination.
[0025] Based on the above, the present invention further provides a design method for a ducted propeller, which applies the intelligent design method for the component described above, and the component is a ducted propeller.
[0026] Preferably, the design variable parameters of the component include: the length L of the catheter edge, the thickness d1 of the catheter edge, the length r of the bracket, the thickness d2 of the bracket, the angle C between each bracket and the midline = (c 1 ,c 2 ,…c n ), and depth h.
[0027] Preferably, after the design scheme of the component is generated by using the best combination of design variable parameters, the scheme can also be passed to humans as a result of preliminary optimization for in-depth optimization.
[0028] The beneficial effect of the present invention is that: the physical information neural network (PINN) is used to replace computational fluid dynamics (CFD) to design components of a specified flow field. Since PINN can quickly generate solutions after training, its prediction speed is much faster than that of the traditional CFD model. Once the PINN model training is completed, for the same or similar results, the replacement model can directly generate prediction results without repeating numerical solutions. In addition, PINN has a stronger generalization ability by learning the intrinsic connection between physical laws and data - as long as the physical laws remain unchanged, PINN can be applied to similar but slightly different problems without repeatedly establishing CFD models and grids. That is, this method does not require the original component design to draw a calculation grid, then use numerical simulation software to calculate and finally calculate the flow field performance; on the contrary, based on the PINN model and algorithm, the flow field data and performance parameters of the component can be directly obtained according to the design variable parameters and the component design can be carried out based on this, thereby greatly shortening the time and labor cost required for the design of the specified flow field component. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 these drawings without paying creative work.
[0030] Figure 1 A flowchart of a component-based intelligent design method of the present invention;
[0031] Figure 2 A flow chart of a method for training a substitute model using a physical information neural network (PINN) method of the present invention;
[0032] Figure 3 A flowchart of the present invention for performing alternative model training on a PINN model using a training set and a test set;
[0033] Figure 4 It is a schematic diagram of a symmetrical propeller duct component of the present invention. DETAILED DESCRIPTION
[0034] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present invention, but are not intended to limit the scope of the present invention. Similarly, the following examples are only partial embodiments of the present invention rather than all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.
[0035] like Figure 1 As shown, an intelligent design method for a component that generates a specified flow field comprises the following steps:
[0036] S1. Use artificial intelligence methods to train a large amount of simulation and test data to form an alternative model to replace CFD calculations;
[0037] S2. Obtain design variable parameters of the component;
[0038] S3. Assign initial values to the variables of the parametric design, use the surrogate model to quickly generate the target flow field, compare the results of the flow field with the results of the target flow field, calculate the overall root mean square error (RMSE) of the two flow fields, and use the optimization algorithm to automatically iteratively optimize the design parameter variables with the goal of minimizing the root mean square error (RMSE), and finally obtain the best combination of parameter variables and use the best combination of parameter variables to generate a design solution for the component;
[0039] Furthermore, when the component generating the specified flow field is a symmetrical propeller duct (or ducted propeller), in step S1, the propeller wake field under the propeller duct constraint can be trained using the physical information neural network (PINN) method, and finally an alternative model for simulating the propeller wake field under the duct constraint is obtained.
[0040] Figure 2 The flowchart of a method for training a substitute model using a physical information neural network (PINN) method in step S1 is shown, comprising:
[0041] S11, sampling within the design range of the component design variable parameters to obtain a design variable parameter data set uniformly distributed in the design variable parameter space;
[0042] S12, calculating the performance parameters of the component according to the flow field data in the flow field data set, and using the flow field data and the performance parameters as label data for training the PINN model;
[0043] S13, randomly sampling a preset number of grid points in the flow field area, and taking the flow field data set and the preset number of grid points as the data set;
[0044] S14. Randomly extract 90% of the data in the data set as the training set, and the remaining 10% of the data as the test set. Use the training set and the test set to train the PINN model to obtain a trained PINN model.
[0045] When using physical information neural network for simulation, there is no need for results from CFD simulation or experimental data as training data sets. The required training data are the spatial coordinates of the flow field.
[0046] Figure 3 The flowchart of step S14 using the training set and the test set to train the PINN model includes:
[0047] A1. The control equations of fluid flow, namely the continuity equation and the NS equation, are used as the prior physical information of PINN, which includes:
[0048]
[0049] In the formula, is the Hamiltonian operator; u = (u, v, w) is the fluid velocity, ρ and v are the fluid density and kinematic viscosity coefficient respectively; the model takes the flow field position coordinates and time (i.e., the spatiotemporal information of the flow field) as the input of the neural network.
[0050] A2. Set the network structure of the constructed PINN - PINN consists of a fully connected neural network and a residual network. The fully connected neural network consists of 8 layers of neural networks, including 1 input layer, 6 hidden layers and 1 output layer. The input layer and the output layer each contain 5 neurons, each hidden layer contains 30 neurons, and a fully connected network is used between layers;
[0051] Furthermore, the network weights w and biases b of each layer of PINN are initialized using the He initialization method or the orthogonal initialization method, the activation function of each layer of neurons uses the hyperbolic tangent function, and the optimization function is a multi-objective optimization function;
[0052] A3. During the training process, the loss function of PINN is optimized by stochastic gradient descent combined with particle swarm optimization algorithm, and the weight w and bias b are updated;
[0053] A4. The predicted value obtained when the loss function value of PINN is reduced to the set optimization target. At this time, the PINN network parameters are the flow field reconstruction alternative model of PINN that replaces the CFD model.
[0054] Figure 4 The figure shows a schematic diagram of a symmetrical propeller duct. Further, when the component generating the specified flow field is a symmetrical propeller duct, the design variable parameters in step S2 are the length L of the duct edge, the thickness d1 of the duct edge, the length r of the bracket, the thickness d2 of the bracket, and the angle C between each bracket and the center line = (c 1 ,c 2 ,…c n ), and depth (perpendicular to the paper) h.
[0055] Furthermore, when the component generating the specified flow field is a symmetrical propeller duct, in step S3, the parameters (L, d1, r, d2, C, h) of the symmetrical propeller duct are assigned initial values, and the PINN replacement model trained by artificial intelligence is used to quickly generate the wake field in the propeller duct design, and the results of the flow field (many data points) are compared with the results of the target flow field (individual data points at the same position), and the overall mean square error RMSE of the two flow fields is calculated, and each physical quantity is calculated by the neural network, that is, the output flow field structure information of the neural network is substituted into the partial differential equation, and the equation error is obtained by automatic differentiation calculation, and then the error between the output flow field information of the time network and the known information under the time and space coordinates is solved, and the error loss function is minimized by the back propagation and gradient descent algorithm of the neural network until the neural network converges. Among them,
[0056]
[0057] Among them, X and Y are the target flow field and the flow field predicted and output by the PINN model, respectively, and m is the number of data points at different positions. The result of the target flow field can be obtained based on theoretical analysis and previous design experience. Subsequently, with the minimum root mean square error RMSE as the goal, the design variable parameters are automatically iteratively optimized using optimization algorithms such as particle swarm and simulated annealing, and finally the best variable combination is obtained.
[0058] Furthermore, after the design scheme of the component is generated by using the best combination of design variable parameters, the scheme can be handed over to the designer as a result of preliminary optimization for in-depth optimization.
[0059] The above descriptions are only some embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for intelligent design of components, characterized in that: It includes the following steps: S1. Use artificial intelligence methods to build models for training simulation and test data to form alternative models to replace CFD calculations; S2. Obtain design variable parameters of the component; S3. Assign initial values to the variables of the parametric design, use the alternative model to quickly generate the target flow field, compare the results of the flow field with the results of the target flow field, calculate the overall root mean square error (RMSE) of the two flow fields, and use the optimization algorithm to automatically iteratively optimize the design parameter variables with the goal of minimizing the root mean square error (RMSE). Finally, the best combination of parameter variables is obtained and the design scheme of the component is generated using the best combination of parameter variables.
2. The intelligent design method of a component according to claim 1, characterized in that: Alternative models that form an alternative to CFD calculations include: The physical information neural network PINN method is used for training, and finally a substitute model for CFD calculation with a specified flow field under constraint conditions is obtained.
3. The intelligent design method of a component according to claim 2, characterized in that: The training method using the physical information neural network PINN method includes: S11, sampling within the design range of the component design variable parameters to obtain a design variable parameter data set uniformly distributed in the design variable parameter space; S12, calculating the performance parameters of the component according to the flow field data in the flow field data set, and using the flow field data and the performance parameters as label data for training the PINN model; S13, randomly sampling a preset number of grid points in the flow field area, and taking the flow field data set and the preset number of grid points as the data set; S14. Randomly extract 90% of the data in the data set as the training set, and the remaining 10% of the data as the test set. Use the training set and the test set to train the PINN model to obtain a trained PINN model.
4. The intelligent design method of a component according to claim 3, characterized in that: The method of using a training set and a test set to train the PINN model includes: A1. The governing equations of fluid flow, namely the continuity equation and the NS equation, are used as the prior physical information of PINN. A2. Set the network structure of the constructed PINN; A3. During the training process, the loss function of PINN is optimized by stochastic gradient descent combined with particle swarm optimization algorithm, and the weight w and bias b are updated; A4. The predicted value obtained when the loss function value of PINN is reduced to the set optimization target. At this time, the PINN network parameters are the flow field reconstruction alternative model of PINN that replaces the CFD model.
5. The intelligent design method of a component according to claim 4, characterized in that: The network structure of the constructed PINN is composed of a fully connected neural network and a residual network, wherein the fully connected neural network is composed of 8 layers of neural networks, namely 1 input layer, 6 hidden layers and 1 output layer; wherein the input layer and the output layer each contain 5 neurons, each hidden layer contains 30 neurons, and a fully connected network is used between layers.
6. The intelligent design method of a component according to claim 4, characterized in that: The network structure of the PINN constructed by the setting, the network weights w and bias b of each layer of the PINN are initialized using the He initialization method or the orthogonal initialization method, the activation function of each layer of neurons uses the hyperbolic tangent function, and the optimization function is a multi-objective optimization function.
7. The intelligent design method of a component according to claim 1, characterized in that: The automatic iterative optimization of the design parameter variables by using the optimization algorithm is to automatically iteratively optimize the design variable parameters by using a particle swarm and / or simulated annealing optimization algorithm, and finally obtain the best design variable parameter combination.
8. A method for designing a ducted propeller, characterized in that: The invention applies the intelligent design method of the component according to any one of claims 1 to 7, wherein the component is a ducted propeller.
9. The method for designing a ducted propeller according to claim 8, characterized in that: The design variable parameters of the component include: the length L of the catheter edge, the thickness d1 of the catheter edge, the length r of the bracket, the thickness d2 of the bracket, the angle C between each bracket and the center line = (c1, c2, ... c n ), and depth h.
10. The method for designing a ducted propeller according to claim 9, characterized in that: After the design scheme of the component is generated by using the best combination of design variable parameters, the scheme can be passed to humans for in-depth optimization as the result of preliminary optimization.