Valve structure automatic design method and system based on generative AI

Abstract

The application discloses an automatic valve structure design method and system based on generative AI, and belongs to the technical field of intelligent design of industrial equipment and artificial intelligence. The steps are as follows: first, multi-dimensional representation of a valve structure and construction of a data set; second, training of a conditional generative AI model; third, automatic generation of a valve structure; fourth, rapid verification and optimization of a generated result; and fifth, reconstruction and output of a three-dimensional parameterized valve model. The automatic valve structure design system comprises five modules, namely, construction of a valve data set, training of a conditional generative AI model, generation of a valve structure, verification and optimization, and reconstruction and output of a three-dimensional parameterized valve model. The application can generate a new high-performance valve structure with a non-traditional topology and a flow channel, can provide multiple Pareto optimal schemes at one time, and realizes full-process, automatic and intelligent structure design of a valve from performance requirements to a producible and manufacturable model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial equipment intelligent design and artificial intelligence, and relates to a valve structure automatic design method and system based on generative AI. BACKGROUND

[0002] As a key component of fluid control systems, valves are widely used in major engineering fields such as petroleum and chemical industry, power, water conservancy, aerospace, and nuclear industry. The performance of a valve directly affects the safety, efficiency, and energy consumption of the entire system. The performance of a valve is almost entirely achieved through structural design. An excellent valve structure design means higher safety, lower energy consumption, longer service life, and lower total cost, and its contribution to industrial equipment and even national economic development is fundamental and strategic.

[0003] Current valve structure design has been extensively studied. Zong Chaoyong et al. proposed a dynamic modeling method for nuclear power lift valves under liquid medium based on CFD (Chinese invention patent CN202410743363.8). However, the method requires a completed, artificially pre-set three-dimensional valve model as input, and still needs to manually return and modify the model for recalculation based on experience. The design is not automated, and the high-fidelity CFD calculation is extremely time-consuming, which severely restricts the breadth and speed of structural design. Zhang Liu et al. proposed a method for designing an airborne jet control valve, a control valve, and a jet actuation system (Chinese invention patent CN202311014129.3). However, the design and optimization space of the invention method is strictly limited within a certain valve working principle and structural framework pre-set to achieve a specific system function, and is only limited to specialized and systematic design in specific application scenarios, unable to explore completely different valve design configurations in principle that may have better performance. In summary of existing research, traditional valve structure design mainly has the following significant defects:

[0004] High dependence on design experience, low efficiency; great limitations in innovation, difficult to break out of traditional structural framework; long design iteration period and high cost due to manual model modification and simulation parameter setting; difficulty in handling multi-objective optimization problems, making it extremely difficult for manual optimization when multiple conflicting objectives such as flow characteristics, structural strength, weight, and cost are encountered in design.

[0005] With the rapid development of artificial intelligence technology, generative AI technology provides a new approach to solving the above problems. Generative AI can recognize and understand the relevance of different structural forms by learning and analyzing a large amount of structural design data, enabling it to generate a large number of innovative design candidates in a short time beyond traditional design frameworks.

[0006] Therefore, in order to overcome the defects of the prior art, it is necessary to propose a valve structure automatic design method and system based on generative AI. SUMMARY

[0007] In view of the problems and deficiencies in the prior art, the present application provides a valve structure automatic design method and system based on generative AI, which constructs a multi-modal training sample database by using a hybrid representation method of voxels and boundary representation, which fuses valve geometric representation, key performance, working condition and material condition, iteratively trains the established conditional generative AI model using a composite loss function, establishes a valve structure automatic generation mechanism, and quickly verifies and optimizes the generation result through multi-physics CAE simulation, and finally outputs a standard format file that can be directly used for production and manufacturing using parameterization reconstruction technology, so as to realize the whole process, automation and intelligent structure design of the valve from performance requirements to a producible and manufacturable model, and significantly improve the efficiency and innovation of valve design.

[0008] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is:

[0009] A valve structure automatic design method based on generative AI, the valve structure automatic design method comprising the following steps:

[0010] First, valve structure multi-dimensional representation and data set construction. A hybrid representation method of voxels and boundary representation is used to construct a training sample data set based on generated voxel tensors, extracted boundary features, obtained key performance vectors and encoded condition vectors. Specifically as follows:

[0011] Step 1.1, generate valve voxelization representation. Based on the historical design database of the valve, select valve three-dimensional CAD models and valve CAE simulation results of successful products, wherein the value range is 5000~8000, and the three-dimensional CAD model of the valve is converted into a high-resolution voxel grid. The voxel grid is composed of voxel units, and each voxel unit is represented by a binary value 0 or 1 indicating whether the position of the voxel unit is occupied by material or empty. Let the valve structure geometric space be , the corresponding voxel tensor represents the probability of the position of each voxel unit being occupied:

[0012] (1)

[0013] wherein, are the resolutions in x, y and z dimensions, respectively, which are used to capture the core structural features of the valve such as the outline of the valve body, the direction of the flow passage and the general shape of the valve disc.

[0014] Step 1.2, boundary feature extraction. Based on the valve 3D CAD model selected in step 1.1, the boundary representation method (B-Rep) is used to further extract boundary representation information, including the topological relationship of faces, loops, edges, and points, as well as geometric parameters including curvature and normal vector, and a graph variational autoencoder is used to generate B-Rep graph structure data , achieving accurate preservation of topological and geometric information.

[0015] Step 1.3, performance data quantification. Based on the valve CAE simulation results selected in step 1.1, the key performance vectors are obtained:

[0016] (2)

[0017] wherein, is the flow coefficient, is the maximum equivalent stress, is the maximum deformation, is the opening and closing torque, is the valve mass, is the sealing specific pressure, is the pressure drop.

[0018] Step 1.4, working condition and material condition coding. Based on the valve 3D CAD model selected in step 1.1, the design input conditions are coded into condition vectors :

[0019] (3)

[0020] wherein, is the nominal diameter, is the nominal pressure, is the working temperature, is the medium density, is the medium viscosity, is the material elastic modulus, is the material yield strength.

[0021] Step 1.5, construction of training sample data set. Based on the voxel tensor generated in step 1.1, the boundary features extracted in step 1.2, the key performance vectors obtained in step 1.3, and the condition vectors coded in step 1.4, a training sample data set is constructed, each training sample is a triple , wherein is the geometric representation composed of voxel tensor and B-Rep graph structure data, P is the key performance vector, and C is the condition vector, specifically:

[0022] (4)

[0023] wherein, is the geometric representation of the th valve, is the key performance vector of the th valve, is the condition vector of the th valve.

[0024] Second step, conditional generative AI model training. Based on the training sample dataset constructed in the first step, the conditional generative AI model architecture is established, and its iterative training is carried out by defining the composite loss function, so as to realize the automatic generation of high-performance, manufacturable valve design scheme according to the given performance index and working condition and material condition. Specifically as follows:

[0025] Step 2.1, design the architecture of conditional generative AI model. Based on the training sample dataset constructed in the first step , the architecture of conditional generative AI model is designed, including generator , discriminator and performance predictor .

[0026] The generator inputs a noise vector sampled from normal distribution, target performance vector and target condition vector , wherein is the identity matrix, respectively from the key performance vector and the condition vector obtained in the first step. The output is a generated valve geometric expression , wherein is the target voxel tensor, is the target B-Rep graph structure data.

[0027] The output of the generator corresponds to the geometric representation obtained in the first step in structure and generation method, ensuring the consistency of the generator output data with the geometric real data.

[0028] The discriminator inputs a target voxel tensor based on the first step database and the corresponding target condition vector through 3D convolution network and full connection layer, to judge whether the valve geometric structure is real and reasonable under the given condition.

[0029] The performance predictor also inputs the target voxel tensor and the corresponding target condition vector , the output is the target performance vector , i.e. the performance exhibited by the arbitrary geometry under the given condition, to establish a mapping from the geometry space to the performance space.

[0030] Step 2.2, define the composite loss function. Based on the conditional generative AI model architecture designed in step 2.1, define a composite loss function for the generator to guide the adjustment and optimization of the generator parameters, including the adversarial loss function, the conditional consistency loss function, the performance-oriented loss function and the geometry reconstruction loss function.

[0031] The adversarial loss function drives the generator to produce valve geometry as close as possible to the real geometry:

[0032] (5)

[0033] where, is the expectation, is the generated data distribution; is the discriminator scores the output of the generator under the corresponding target condition vector . is the output of the generator subject to the expectation of the generated data distribution .

[0034] The conditional consistency loss function ensures that the valve geometry generated by the generator strictly matches the input condition:

[0035] (6)

[0036] where, is the condition encoder, is the mean square error, is the output of the generator input to the condition encoder after the output result is obtained.

[0037] The performance-oriented loss function ensures that the valve geometry generated by the generator meets the expected performance indicators:

[0038] (7)

[0039] The geometry reconstruction loss function ensures the rationality of the B-Rep graph structure data generated by the generator in topology and geometry:

[0040] (8)

[0041] wherein, is a binary cross-entropy loss function; is a mean square error loss function; are the adjacency matrices of the real valve and the generated valve B-Rep graph, respectively; are the node feature matrices of the real valve and the generated valve B-Rep graph, respectively; is a hyper-parameter related to the binary cross-entropy loss function, is a hyper-parameter related to the mean square error loss function.

[0042] The composite loss function of the generator is the weighted sum of the above loss functions:

[0043] (9)

[0044] wherein, is a hyper-parameter related to the adversarial loss function, is a hyper-parameter related to the conditional consistency loss function, is a hyper-parameter related to the performance-oriented loss function, is a hyper-parameter related to the geometry reconstruction loss function.

[0045] Step 2.3, iterative training of the conditional generative AI model. Based on the composite loss function obtained in step 2.2 , the training is systematically repeated by adjusting the parameters of the conditional generative AI model, as follows:

[0046] Step 2.3.1, for the joint training of the generator and the discriminator , the following two main loop paths are implemented:

[0047] First loop path: fixing the generator and the performance predictor , updating the discriminator . A batch of real data is sampled from the training sample data set in the first step, and a batch of generated data is generated by the generator , and the discriminator is trained to better judge the geometric authenticity.

[0048] Second loop path: fixing the discriminator and the performance predictor , updating the generator . Calculate the gradient according to the composite loss function obtained in step 2.2, and update the generator the parameters of the performance-oriented loss function The gradient of the performance-oriented loss function flows through the performance predictor to the generator , thus guiding the generator to generate a higher performance valve geometry.

[0049] Step 2.3.2, for the pre-training of the performance predictor , use the training sample dataset in the first step to conduct independent training until its prediction accuracy reaches an acceptable level.

[0050] Step 2.3.3, after repeated iterative training, the generator realizes the mapping of a random noise vector to a real, manufacturable, and high-performance valve three-dimensional structure under the strict constraints of the target performance vector and the target condition vector

[0051] .

[0052] Step 3.1, condition input. Based on the trained generator in Step 2, input the valve working condition and material information as well as the key performance requirement information of the valve, and encode them into the corresponding target condition vector and target performance vector .

[0053] Step 3.2, latent space search. Based on the target condition vector and target performance vector encoded in Step 3.1, randomly sample different noise vectors from a noise space subject to a multivariate standard normal distribution . For each sampled , concatenate it with the target condition vector and target performance vector encoded in Step 3.1, and input them together into the trained generator to quickly generate initial valve geometries .

[0054] Step 3.3, preliminary screening. Based on the valve geometry generated in step 3.2 and the target condition vector input in step 3.1 , input into the second step trained performance predictor to obtain the predicted value of the corresponding key performance indicators , and filter according to the set performance requirements.

[0055] Step 3.4, diversity generation. Based on the valve geometry screened in step 3.3, a set of Pareto optimal solutions is obtained through Pareto sorting, and the most representative valve structure design candidate scheme is selected according to the performance.

[0056] Fourth step, rapid verification and optimization of the generated results. Based on the valve structure design candidate scheme obtained in the third step, high-fidelity CAE simulation is carried out, and the performance predictor is fine-tuned according to the error between the valve key performance simulation results and the key performance indicator prediction values to realize the continuous optimization of the candidate scheme. Specifically as follows:

[0057] Step 4.1, carry out high-fidelity CAE simulation. Based on the valve structure design candidate scheme obtained in the third step, full-process finite element (FEM) and fluid mechanics (CFD) simulation is carried out to obtain accurate valve key performance simulation results .

[0058] Step 4.2, simulation and generation difference optimization. Based on the valve key performance simulation results obtained in step 4.1 , calculate the error with the key performance indicator prediction values obtained in the third step to fine-tune the performance predictor , and improve the reliability of future prediction.

[0059] The fine-tuning strategy adopted is: when the error is greater than the preset threshold, the candidate scheme that meets the condition is selected to join the fine-tuning data set, and the learning rate of the performance predictor is reduced to of the current learning rate using the Adam optimizer, so as to realize small-scale and targeted parameter update, and improve the prediction accuracy of the performance predictor . When the error is not greater than the preset threshold, the performance predictor does not need to be fine-tuned.

[0060] Step 5, valve three-dimensional parametric model reconstruction and output. Based on the fourth step of the optimized valve structure design candidate scheme, the corresponding valve geometric voxel tensor is converted into a triangular mesh model, and a three-dimensional model of the valve is obtained through B-Rep reconstruction. Finally, the valve three-dimensional parametric model is obtained through parameterized feature recognition and reconstruction, and is converted into a standard format file that can be directly used for output. Specifically as follows:

[0061] Step 5.1, valve geometric voxel tensor conversion. Based on the valve geometric voxel tensor generated by the generator after optimization in the fourth step , the moving cube algorithm is used to convert it into a triangular mesh model, so that it can clearly represent the geometric shape of the valve.

[0062] Step 5.2, B-Rep-based valve three-dimensional model reconstruction. Based on the triangular mesh model obtained in step 5.1, combine it with the valve B-Rep graph structure data output by the generator after optimization in the fourth step, perform surface fitting and topological reconstruction to obtain a smooth and closed three-dimensional model of the valve.

[0063] Step 5.3, parameterized feature recognition and reconstruction. Based on the three-dimensional model of the valve reconstructed in step 5.2, use feature recognition algorithms for parameterized reconstruction to restore the parameterized features of the model, and obtain the three-dimensional parametric model of the valve.

[0064] Step 5.4, valve three-dimensional parametric model output. Based on the three-dimensional parametric model of the valve reconstructed in step 5.3, convert it into a standard format file that can be directly used for production and manufacturing, and realize seamless integration with CAD / CAE.

[0065] A valve structure automatic design system based on generative AI, which realizes the above-mentioned valve structure automatic design method based on generative AI. The valve structure automatic design system includes a valve data set construction module, a conditional generative AI model training module, a valve structure generation module, a verification and optimization module, and a valve three-dimensional parametric model reconstruction and output module. Specifically:

[0066] The valve data set construction module is composed of a data interface and a geometric representation engine, which is used to realize data extraction of the valve historical design database, perform voxelization representation generation, boundary feature extraction, performance data quantization, and working condition and material condition coding, and is the data basis of the entire system.

[0067] The conditional generative AI model training module accesses the valve data set construction module and is composed of a high-performance computing server, which is used to store pre-trained conditional generative AI models, provide an integrated environment for conditional generative AI model training, reasoning, management and deployment, and is the training cluster of the entire system.

[0068] The valve structure generation module accesses the conditional generative AI model training module and the valve dataset construction module, is composed of a design generation engine, and is used for providing a valve structure automatic design and generation service, receiving a design requirement, and presenting a generation result, and is a generation portal of the whole system.

[0069] The verification and optimization module accesses the valve structure generation module, the conditional generative AI model training module and the valve dataset construction module, is composed of a CAE simulation software system, is used for accurately verifying and optimizing the design generated by AI, and is an optimization center of the whole system.

[0070] The valve three-dimensional parametric model reconstruction and output module accesses the verification and optimization module, the valve structure generation module, the conditional generative AI model training module and the valve dataset construction module, is composed of a geometric reconstruction engine and a format converter, is used for realizing parametric reconstruction and format conversion of the valve three-dimensional parametric model, and is an output window of the whole system.

[0071] The beneficial effects of the present application are:

[0072] (1) The present application constructs a multi-modal training sample database fusing valve geometric representation, key performance, working condition and material condition by using a hybrid representation method of voxels and boundary representation, iteratively trains the established conditional generative AI model by using a composite loss function, establishes a valve structure automatic generation mechanism, can generate a new high-performance valve structure with a non-traditional topology and a flow channel, provides multiple Pareto optimal schemes at one time, realizes multi-objective decision-making, and significantly improves the efficiency and innovation of valve design.

[0073] (2) The present application verifies and optimizes the generation result by multi-physics field CAE simulation, finally outputs a standard format file which can be directly used for production and manufacturing by using a parametric reconstruction technology, so as to realize a full-process, automatic and intelligent structure design of a valve from performance requirements to a producible and manufacturable model. At the same time, the final output is a standard format file, which is completely compatible with the existing industrial software chain, can be seamlessly connected with subsequent engineering analysis, process design and manufacturing, and has strong landing performance. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 It is a valve structure automatic design method flowchart based on generative AI.

[0075] Figure 2 It is a valve structure automatic design system structure schematic diagram based on generative AI. DETAILED DESCRIPTION

[0076] The application is further described below in combination with a specific implementation case, which is to design a stop valve for a medium-temperature steam system, with the goal of maximizing its flow coefficient under the premise of meeting strength requirements .

[0077] A valve structure automatic design method based on generative AI, comprising the following steps:

[0078] Step 1, valve structure multi-dimensional representation and data set construction. A hybrid representation method of voxel and boundary representation is adopted, and a training sample data set is constructed based on the generated voxel tensor, extracted boundary features, obtained key performance vectors and encoded condition vectors. Specifically as follows:

[0079] Step 1.1, generate the voxelized representation of the valve. Based on the historical design database of the valve, select three-dimensional CAD models and valve CAE simulation results of successful products from the valve, wherein the value range is 5000-8000, and the three-dimensional CAD model of the valve is converted into a high-resolution voxel grid. The voxel grid is composed of voxel units, and each voxel unit is represented by a binary value 0 or 1 indicating whether the position of the voxel unit is occupied by material or empty. Let the valve structure geometric space be , the corresponding voxel tensor represents the probability of the position of each voxel unit being occupied:

[0080] (1)

[0081] wherein are the resolutions in x, y and z dimensions, respectively, which are used to capture the core structural features of the valve such as the outline of the valve body, the direction of the flow passage and the general shape of the valve disc.

[0082] In this embodiment, , .

[0083] Step 1.2, boundary feature extraction. Based on the valve three-dimensional CAD model selected in step 1.1, the boundary representation method (B-Rep) is further used to extract boundary representation information, including the topological relationship of faces, loops, edges and points, as well as the geometric parameters including curvature and normal vector, and a graph variational autoencoder is used to generate B-Rep graph structure data , realizing the preservation of accurate topological and geometric information.

[0084] Step 1.3, performance data quantification. Based on the valve CAE simulation results selected in step 1.1, obtain the key performance vectors :

[0085] (2)

[0086] wherein, is the flow coefficient, is the maximum equivalent stress, is the maximum deformation, is the opening and closing torque, is the valve mass, is the sealing specific pressure, is the pressure drop.

[0087] Step 1.4, working condition and material condition coding. Based on the valve 3D CAD model selected in step 1.1, the design input conditions are coded into a condition vector :

[0088] (3)

[0089] wherein, is the nominal diameter, is the nominal pressure, is the working temperature, is the medium density, is the medium viscosity, is the material elastic modulus, is the material yield strength.

[0090] Step 1.5, construction of training sample data set. Based on the voxel tensor generated in step 1.1, the boundary feature extracted in step 1.2, the key performance vector obtained in step 1.3 and the condition vector coded in step 1.4, a training sample data set is constructed, each training sample is a triple , wherein is the geometric representation composed of voxel tensor and B-Rep graph structure data, P is the key performance vector, and C is the condition vector, in detail:

[0091] (4)

[0092] wherein, is the geometric representation of the th valve, is the key performance vector of the th valve, is the condition vector of the th valve.

[0093] Step 2, conditional generative AI model training. Based on the training sample data set constructed in step 1, the conditional generative AI model architecture is established, and its iterative training is carried out by defining the composite loss function, so as to realize the automatic generation of high-performance and manufacturable valve design scheme according to the given performance index and working condition and material condition. The specific steps are as follows:

[0094] Step 2.1, design the conditional generative AI model architecture. Based on the training sample dataset constructed in the first step , design the conditional generative AI model architecture, including the generator , discriminator and performance predictor .

[0095] The generator inputs a noise vector sampled from a normal distribution , target performance vector and target condition vector , where is an identity matrix, are respectively the key performance vector and condition vector obtained in the first step. The output is a generated valve geometry representation , where is the target voxel tensor, is the target B-Rep graph structure data.

[0096] The output of the generator is fully corresponding to the geometric representation structure obtained in the first step and the generation method, ensuring the consistency of the generator output data with the geometric real data.

[0097] The discriminator inputs a target voxel tensor obtained based on the first step database and the corresponding target condition vector through a 3D convolution network and a fully connected layer, to judge whether the valve geometry structure is real and reasonable under the given condition.

[0098] The performance predictor also inputs the target voxel tensor and the corresponding target condition vector through a 3D convolution network and a fully connected layer, and outputs the target performance vector , i.e. the performance of any geometry structure under the given condition, establishing a mapping from the geometry space to the performance space.

[0099] Step 2.2, define the compound loss function. Based on the conditional generative AI model architecture designed in step 2.1, define a compound loss function for the generator to guide the adjustment and optimization of the generator parameters, including the adversarial loss function, the conditional consistency loss function, the performance-oriented loss function and the geometry reconstruction loss function.

[0100] The adversarial loss function The valve geometry generated by the generator approximates the real geometry as closely as possible:

[0101] (5)

[0102] where, is the expectation, is the generated data distribution; is the discriminator evaluates the output of the generator under the corresponding target condition vector and scores it; is the output of the generator subject to the expectation of the generated data distribution .

[0103] The condition consistency loss function ensures that the valve geometry generated by the generator strictly matches the input condition:

[0104] (6)

[0105] where, is the condition encoder, is the mean square error, is the output of the generator after inputting it into the condition encoder .

[0106] The performance-oriented loss function ensures that the valve geometry generated by the generator meets the expected performance indicators:

[0107] (7)

[0108] The geometry reconstruction loss function ensures that the B-Rep graph structure data generated by the generator is reasonable in topology and geometry:

[0109] (8)

[0110] where, is the binary cross-entropy loss function; is the mean square error loss function; are the adjacency matrices of the real valve and the generated valve B-Rep graph, respectively; are the node feature matrices of the real valve and the generated valve B-Rep graph, respectively; is the hyperparameter related to the binary cross-entropy loss function, is the hyperparameter related to the mean square error loss function.

[0111] Composite loss function of generator is a weighted sum of the above loss functions:

[0112] (9)

[0113] wherein, is a hyper-parameter related to the adversarial loss function, is a hyper-parameter related to the conditional consistency loss function, is a hyper-parameter related to the performance-oriented loss function, is a hyper-parameter related to the geometric reconstruction loss function.

[0114] In this embodiment, at the initial time , , the optimal value can be obtained through repeated parameter tuning.

[0115] Step 2.3, iterative training of the conditional generative AI model. Based on the composite loss function obtained in step 2.2 , the training is realized through systematic and repeated parameter tuning of the conditional generative AI model system, as follows:

[0116] Step 2.3.1, for the joint training of the generator and the discriminator , the following two main loop paths are realized:

[0117] First loop path: fix the generator and the performance predictor , update the discriminator . Sample a batch of real data from the training sample data set in the first step, at the same time use the generator to generate a batch of generated data, train the discriminator to better judge the geometric authenticity.

[0118] Second loop path: fix the discriminator and the performance predictor , update the generator . Calculate the gradient according to the composite loss function obtained in step 2.2, and update the parameters of the generator through back propagation. In this process, the gradient of the performance-oriented loss function flows through the performance predictor to the generator , thereby guiding the generator to generate a valve geometry expression with higher performance.

[0119] Step 2.3.2, for the pre-training of the performance predictor , directly use the training sample data set in the first step Independent training is performed until its prediction accuracy reaches an acceptable level.

[0120] Step 2.3.3, after repeated iterative training, the generator achieves the implementation of mapping a random noise vector to a real, manufacturable and high-performance valve three-dimensional structure under the strict constraints of the target performance vector and the target condition vector .

[0121] In this embodiment, the Adam optimizer is used for training, and the training continues until the discriminator loss and the generator loss converge, and the performance predictor has a prediction error on the test set.

[0122] Third step, automatic generation of valve structure. Based on the trained conditional generative AI model in the second step, input the valve working condition and material information and valve key performance requirement information for coding, generate the initial valve geometry structure through latent space search, and preliminarily screen according to the set performance requirements. Finally, according to any specified design requirements, automatically and quickly generate multiple feasible valve structure design schemes, and select the most representative valve structure design candidate scheme according to the performance advantages and disadvantages. The specific process is as follows:

[0123] Step 3.1, condition input. Based on the trained generator in the second step, input the valve working condition and material information and valve key performance requirement information, and encode them into corresponding target condition vector and target performance vector .

[0124] In this embodiment, each item in the target condition vector is as follows:

[0125] , = 16, , the medium is water, , , the valve body and valve cover material is cast carbon steel, , .

[0126] Each item in the target performance vector is as follows:

[0127] · = 130 MPa, , , ,

[0128] , , the optimization objective is to maximize , whose initial value can be set to 140.

[0129] Step 3.2, latent space search. Based on the target condition vector and the target performance vector encoded in step 3.1, randomly sample different noise vectors from a noise space subject to a multivariate standard normal distribution . For each sampled , concatenate it with the target condition vector and the target performance vector encoded in step 3.1, and input them together into the trained generator to quickly generate initial valve geometries .

[0130] Step 3.3, preliminary screening. Based on the valve geometries generated in step 3.2 and the target condition vector input in step 3.1, input them into the second trained performance predictor to obtain the predicted values of the corresponding key performance indicators , and filter them according to the set performance requirements.

[0131] Step 3.4, diversity generation. Based on the valve geometries screened in step 3.3, obtain a set of Pareto optimal solutions through Pareto sorting, and select the most representative valve structure design candidate scheme among them according to the performance advantages and disadvantages.

[0132] In this embodiment, the better valve structure design candidate scheme is candidate scheme No. 3,

[0133] Fourth step, rapid verification and optimization of the generated results. Based on the valve structure design candidate schemes obtained in the third step, carry out high-fidelity CAE simulation, and fine-tune the performance predictor according to the error between the valve key performance simulation results and the predicted values of the key performance indicators, to realize continuous optimization of the candidate schemes. Specifically as follows:

[0134] Step 4.1, carry out high-fidelity CAE simulation. Based on the valve structure design candidate schemes obtained in the third step, carry out full-process finite element (FEM) and fluid mechanics (CFD) simulation to obtain accurate valve key performance simulation results .

[0135] Step 4.2, simulation and generated difference optimization. Based on the valve key performance simulation results obtained in step 4.1 , calculate the error of its key performance index prediction value obtained in the third step , for fine-tuning the performance predictor , improve the reliability of future prediction.

[0136] The fine-tuning strategy adopted is: when the error is greater than the preset threshold, which is in the range of 5%~10%, select the candidate scheme that meets the condition to join the fine-tuning data set, and use the Adam optimizer to reduce the learning rate of the performance predictor to of the current learning rate, so as to realize small-scale and targeted parameter update, and improve the accuracy of the performance predictor to predict performance.

[0137] In this embodiment, Ansys is used for simulation calculation, and the simulation results of the third candidate scheme , and there is an error. The data pair is added to the valve data set, which is used for subsequent adjustment and optimization of the predictor .

[0138] In the fine-tuning strategy, the preset threshold range is set to 10%, and the learning rate of the performance predictor is reduced to of the current learning rate using the Adam optimizer.

[0139] After multiple rounds of training and optimization, the optimal valve structure design candidate scheme is obtained, and the flow coefficient .

[0140] Step 5, valve three-dimensional parameterized model reconstruction and output. Based on the valve structure design candidate scheme optimized in the fourth step, the corresponding valve geometric voxel tensor is converted into a triangular mesh model, and a valve three-dimensional model is obtained through B-Rep reconstruction, and finally a valve three-dimensional parameterized model is obtained through parameterized feature recognition and reconstruction, and converted into a standard format file that can be directly used for output. The specific steps are as follows:

[0141] Step 5.1, valve geometric voxel tensor conversion. Based on the valve geometric voxel tensor output by the generator after optimization in the fourth step, the moving cube algorithm is used to convert it into a triangular mesh model, so that it can clearly represent the geometric shape of the valve.

[0142] Step 5.2, B-Rep-based valve three-dimensional model reconstruction. Based on the triangular mesh model obtained in step 5.1, it is combined with the valve B-Rep graph structure data The surface fitting and topological reconstruction are performed in combination to obtain a smooth and closed three-dimensional valve model.

[0143] Step 5.3, parameterized feature recognition and reconstruction. Based on the three-dimensional valve model reconstructed in step 5.2, a parameterized reconstruction is performed using a feature recognition algorithm to restore the parameterized features of the model, thereby obtaining a three-dimensional parameterized valve model.

[0144] Step 5.4, three-dimensional parameterized valve model output. Based on the three-dimensional parameterized valve model reconstructed in step 5.3, it is converted into a standard format file that can be directly used for production and manufacturing, realizing seamless integration with CAD / CAE.

[0145] A valve structure automatic design system based on generative AI, which realizes the above-mentioned valve structure automatic design method based on generative AI, the valve structure automatic design system includes a valve data set construction module, a conditional generative AI model training module, a valve structure generation module, a verification and optimization module, and a valve three-dimensional parameterized model reconstruction and output module. Specifically:

[0146] The valve data set construction module is composed of a data interface and a geometric representation engine, which is used to realize data extraction of the valve historical design database, perform voxelization representation generation, boundary feature extraction, performance data quantization, and working condition and material condition coding, and is the data basis of the whole system.

[0147] The conditional generative AI model training module accesses the valve data set construction module and is composed of a high-performance computing server, which is used to store the pre-trained conditional generative AI model, provide an integrated environment for conditional generative AI model training, reasoning, management and deployment, and is the training cluster of the whole system.

[0148] The valve structure generation module accesses the conditional generative AI model training module and the valve data set construction module and is composed of a design generation engine, which is used to provide valve structure automatic design and generation services, receive design requirements and present generation results, and is the generation portal of the whole system.

[0149] The verification and optimization module accesses the valve structure generation module, the conditional generative AI model training module and the valve data set construction module and is composed of a CAE simulation software system, which is used to accurately verify and optimize the AI generated design, and is the optimization center of the whole system.

[0150] The valve three-dimensional parameterization model reconstruction and output module accesses the verification and optimization module, the valve structure generation module, the conditional generative AI model training module and the valve data set construction module, is composed of a geometric reconstruction engine and a format converter, and is used for realizing parameterization reconstruction and format conversion of the valve three-dimensional parameterization model, and is an output window of the whole system.

[0151] The above-described embodiments only express the implementation manners of the present application, and cannot be understood as the limitation to the scope of the present application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which all belong to the protection scope of the present application.

Claims

1. An automatic valve structure design method based on generative AI, characterized in that, The automatic design method for valve structure Includes the following steps: The first step is to construct a multidimensional representation and dataset of the valve structure. A hybrid representation method of voxel and boundary representation is adopted. The training sample dataset is constructed based on the generated voxel tensor, the extracted boundary features, the obtained key performance vectors and condition vectors. The condition vectors are obtained by encoding valve operating conditions and material information. Based on the historical design database of valves, select from it A 3D CAD model of a valve from a successful product was selected. Based on this model, B-Rep was used to extract boundary representation information, including the topological relationships of faces, loops, edges, and points, as well as geometric parameters including curvature and normal vectors. A graph variational autoencoder was then used to generate B-Rep graph structure data. The B-Rep graph structure data For the extracted boundary features; The second step is conditional generative AI model training. Based on the training sample dataset built in the first step, a conditional generative AI model architecture is established, and iterative training is performed on it by defining a composite loss function, thereby realizing the automatic generation of valve structure design schemes according to given performance indicators, working conditions and material conditions. The third step is to automatically generate the valve structure; Based on the conditional generative AI model trained in the second step, the key performance vector and condition vector are input. After generating the initial valve geometry through latent space search, the initial screening is performed according to the set performance requirements. Finally, the system can automatically and quickly generate multiple valve structure design schemes according to any specified design requirements, and select the most representative valve structure design candidate scheme based on the performance. The fourth step involves rapid verification and optimization of the generated results. Based on the valve structure design candidate schemes obtained in the third step, high-fidelity CAE simulations are conducted. The performance predictor is then fine-tuned according to the error between the simulation results of the valve's key performance and the predicted values ​​of the key performance indicators. This enables continuous optimization of candidate solutions; The fifth step is the reconstruction and output of the valve's three-dimensional parametric model. Based on the valve structure design candidate scheme optimized in the fourth step, the corresponding valve geometric voxel tensor is transformed into a triangular mesh model. The valve's three-dimensional model is then obtained through B-Rep reconstruction. Finally, the valve's three-dimensional parametric model is obtained through parametric feature recognition and reconstruction.

2. The automatic valve structure design method based on generative AI according to claim 1, characterized in that, The first step is specifically as follows: Step 1.1: Generate valve voxelization characterization; based on the historical valve design database, select... Three-dimensional CAD models and CAE simulation results of valves for successful products, including The value range is 5000~8000. The 3D CAD model of the valve is converted into a high-resolution voxel mesh. The voxel mesh is composed of voxel elements, and each voxel element uses a binary value of 1 or 0 to indicate whether the voxel element's location is occupied by material or empty. Let the geometric space of the valve structure be... Corresponding voxel tensor This represents the probability that the location of each voxel unit is occupied: （1）； in, These are the resolutions in the x, y, and z dimensions, respectively. Step 1.2, Boundary feature extraction; Step 1.3, Quantification of performance data; Based on the CAE simulation results of the valve selected in Step 1.1, obtain its key performance vector. : （2）； in, For flow coefficient, For the maximum equivalent stress, For the maximum deformation, This refers to the opening and closing torque. For valve quality, For sealing pressure, For pressure drop; Step 1.4, Valve operating condition and material information coding; Based on the valve 3D CAD model selected in Step 1.1, the design input conditions are coded into condition vectors. : （3）； in, For nominal diameter, For nominal pressure, Operating temperature For the density of the medium, For the viscosity of the medium, The elastic modulus of the material. The yield strength of the material; Step 1.5: Construct the training sample dataset; based on the voxel tensor generated in Step 1.1, the boundary features extracted in Step 1.2, the key performance vector obtained in Step 1.3, and the conditional vector encoded in Step 1.4, construct the training sample dataset. Each training sample is a triple. ,in The geometric representation is composed of voxel tensors and B-Rep graph structure data, where P is the key performance vector and C is the condition vector. Specifically: （4）； in, For the first Geometric representation of a valve, For the first Key performance vectors for each valve For the first The condition vector of each valve.

3. The automatic valve structure design method based on generative AI according to claim 2, characterized in that, The second step is specifically as follows: Step 2.1, Design the conditional generative AI model architecture; based on the training sample dataset constructed in Step 1. Design a conditional generative AI model architecture, including a generator. Discriminator and performance predictor ; Step 2.2, Define the composite loss function; Based on the conditional generative AI model architecture designed in Step 2.1, define a composite loss function for the generator to guide the adjustment and optimization of generator parameters, including adversarial loss function, conditional consistency loss function, performance-oriented loss function and geometric reconstruction loss function; Step 2.3, iterative training of the conditional generative AI model; based on the composite loss function obtained in Step 2.

2. Training is achieved by systematically and repeatedly tuning the parameters of a conditional generative AI model.

4. The automatic valve structure design method based on generative AI according to claim 3, characterized in that, In the second step mentioned above: In step 2.1: The generator Input a noise vector sampled from a normal distribution. Target performance vector and target condition vector ,in It is the identity matrix. These are respectively derived from the key performance vectors obtained in the first step. With condition vector Output a generated valve geometry representation. ,in For the target voxel tensor, For target B-Rep graph structure data; The discriminator Composed of a 3D convolutional network and fully connected layers, it takes a target voxel tensor as input. and the corresponding target condition vector To determine whether the valve geometry is realistic and reasonable under given conditions; The performance predictor Also composed of 3D convolutional networks and fully connected layers, it takes the target voxel tensor as input. and the corresponding target condition vector The output is the target performance vector. That is, to predict the performance of any geometric structure under given conditions and establish a mapping from geometric space to performance space; In step 2.2: The adversarial loss function for: （5）； in, For the expectation, To generate a data distribution; For discriminator Output of the generator In the corresponding target condition vector The rating below; The output of the generator Follow the distribution of generated data Expectations; The conditional consistency loss function for: （6）； in, For condition encoders, Mean square error, To convert the generator's output Input to condition encoder The output result obtained afterwards; The performance-oriented loss function for: （7）； The geometric reconstruction loss function for: （8）； in, The binary cross-entropy loss function; Let the mean squared error loss function be used. These are the adjacency matrices for the actual valve and the generated valve B-Rep graph structure data, respectively. These are the node feature matrices of the B-Rep graph structure data of the real valve and the generated valve, respectively. For hyperparameters related to the binary cross-entropy loss function, These are hyperparameters related to the mean squared error loss function; The composite loss function of the generator The weighted sum of the above loss functions: （9）； in, For hyperparameters related to the adversarial loss function, Hyperparameters related to the conditional consistency loss function The hyperparameters related to the performance-oriented loss function are ξ, which are hyperparameters related to the geometric reconstruction loss function.

5. The automatic valve structure design method based on generative AI according to claim 4, characterized in that, Step 2.3 is as follows: Step 2.3.1, for the generator and discriminator The joint training is achieved through the following two major loop paths: First loop path: Fixed generator and performance predictor Update the discriminator ; From the training sample dataset of the first step A batch of real data is sampled, and a generator is used simultaneously. Generate a batch of generated data to train the discriminator. To better determine geometric authenticity; Second loop path: Fixed discriminator and performance predictor Update generator Based on the composite loss function obtained in step 2.2 Calculate the gradient and backpropagate to update the generator. Parameters, performance-oriented loss function The gradient will flow through the performance predictor To generator Guide generator Generate higher-performance valve geometry representations; Step 2.3.2, for the performance predictor Pre-training, using the training sample dataset from step one. Perform independent training until its prediction accuracy meets the requirements; Step 2.3.3: After repeated iterative training, the generator... Implement the transformation of a random noise vector In the target performance vector and target condition vector Under strict constraints, a realistic, manufacturable, and high-performance three-dimensional valve structure is mapped.

6. The automatic valve structure design method based on generative AI according to claim 5, characterized in that, The third step is specifically as follows: Step 3.1, Condition Input: Encode the valve operating conditions, material information, and key performance requirements into corresponding target condition vectors. and target performance vector Input into the generator trained in the second step. middle; Step 3.2, Latent Space Search; from a multivariate standard normal distribution Random sampling in noisy space Different noise vectors For each sampled This is compared with the target condition vector encoded in step 3.

1. and target performance vector The data is concatenated and input into the trained generator. In the middle, quickly generate An initial valve geometry ; Step 3.3, preliminary screening; the valve geometry generated in step 3.2... and the target condition vector input in step 3.1 The input is fed into the performance predictor trained in the second step. In this process, the predicted values ​​of the corresponding key performance indicators are obtained. And filter according to the set performance requirements; Step 3.4, Diversity Generation: Based on the valve geometry selected in Step 3.3, a set of Pareto optimal solutions are obtained through Pareto sorting, and the most representative valve structure design candidate schemes are selected by combining performance advantages and disadvantages.

7. The automatic valve structure design method based on generative AI according to claim 6, characterized in that, The fourth step is specifically as follows: Step 4.1: Conduct high-fidelity CAE simulation; based on the valve structure design candidate schemes obtained in step 3, perform full-process finite element (FEM) and fluid dynamics (CFD) simulations to obtain simulation results of key valve performance. ; Step 4.2, Simulation and Generation Difference Optimization; Based on the simulation results of key valve performance obtained in step 4.1 Calculate its predicted value compared with the key performance indicator obtained in step three. Error, fine-tuning the performance predictor This improves the reliability of future predictions.

8. The automatic valve structure design method based on generative AI according to claim 7, characterized in that, In step 4.2, the fine-tuning strategy is as follows: when the error exceeds a preset threshold (ranging from 5% to 10%), candidate solutions that meet this condition are added to the fine-tuning dataset, and the Adam optimizer is used to optimize the performance predictor. The learning rate is reduced to the current learning rate. This enables small, targeted parameter updates, improving the performance predictor. Accuracy of performance prediction; When the error is not greater than the preset threshold, there is no need to fine-tune the performance predictor. .

9. The automatic valve structure design method based on generative AI according to claim 8, characterized in that, The fifth step is specifically as follows: Step 5.1, Valve geometry tensor transformation; transform the valve geometry tensor output by the generator after optimization in step four. This is transformed into a triangular mesh model to represent the geometric shape of the valve; Step 5.2, Valve 3D Model Reconstruction Based on B-Rep; The triangular mesh model obtained in Step 5.1 is compared with the valve B-Rep graph structure data output by the generator after optimization in Step 4. By combining these methods, surface fitting and topology reconstruction are performed to obtain a smooth and closed three-dimensional model of the valve. Step 5.3, Parametric Feature Recognition and Reconstruction; Based on the valve 3D model reconstructed in step 5.2, a feature recognition algorithm is used for parametric reconstruction to restore the parametric features of the model and obtain the valve 3D parametric model. Step 5.4, Output the three-dimensional parametric model of the valve; Convert the reconstructed three-dimensional parametric model of the valve in Step 5.3 into a standard format file that can be directly used for production and manufacturing, and achieve seamless integration with CAD / CAE.

10. An automatic valve structure design system based on generative AI, characterized in that, The automatic valve structure design system described herein implements the automatic valve structure design method based on generative AI as described in any one of claims 1-9. The automatic valve structure design system includes a valve dataset construction module, a conditional generative AI model training module, a valve structure generation module, a verification and optimization module, and a valve 3D parametric model reconstruction and output module. Specifically: The valve dataset construction module consists of a data interface and a geometric representation engine, which is used to extract data from the valve historical design database, perform voxelization representation generation, boundary feature extraction, performance data quantification, and valve operating condition and material information encoding. The conditional generative AI model training module is connected to the valve dataset construction module to store pre-trained conditional generative AI models and provide an integrated environment for conditional generative AI model training, inference, management and deployment. The valve structure generation module is connected to the conditional generative AI model training module and the valve dataset construction module, and is used to provide automatic valve structure design and generation services, receive design requirements and present generation results. The verification and optimization module is connected to the valve structure generation module, the conditional generative AI model training module, and the valve dataset construction module, and is used to accurately verify and optimize the AI-generated design. The valve 3D parametric model reconstruction and output module is connected to the verification and optimization module, the valve structure generation module, the conditional generative AI model training module, and the valve dataset construction module, and is used to realize the parametric reconstruction and format conversion of the valve 3D parametric model.

Images