Impeller machinery incremental design method and system based on modular network architecture
By decoupling design parameters through a modular network architecture and using independent sub-network modules to train weight parameters, the problems of long time consumption and large data requirements in traditional methods are solved, and efficient and scalable physical field prediction of impeller machinery design is achieved.
Patent Information
- Application Number
- CN202510825644.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-09-26
AI Technical Summary
Traditional numerical simulation methods are time-consuming and difficult to iterate quickly in the design of power machinery equipment. Deep learning models need to be retrained when the design changes, resulting in increased costs and cycles, and a lack of targeted modeling of local differences in design parameters.
A modular network architecture is adopted to decouple the impeller machinery design parameters into independent sub-network modules. The weight parameters are trained through pairwise sample combination and diversified adjustment paths, and the parameter adjustment is implemented step by step to achieve incremental prediction of the physical field.
It realizes the step-by-step progressiveness of physical field prediction and the scalability of the model, reduces data requirements, avoids "dimensionality explosion" and overfitting, and improves design efficiency.
Smart Images
Figure CN120706310A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of power machinery, and in particular relates to an incremental design method and system for impeller machinery based on a modular network architecture. Background Art
[0002] The design and optimization of power machinery (such as gas turbine blades and centrifugal compressor impellers) requires high precision and efficiency in physical field predictions. While traditional numerical simulation methods (such as CFD) can provide accurate results, a single simulation often takes hours, making it difficult to support rapid iteration. While deep learning models have shown potential in accelerating predictions, they still face two major bottlenecks when responding to design changes: the "curse of dimensionality," where the size of the required training samples increases exponentially as the number of design variables increases; and the rigidity of the model. The introduction or adjustment of any new design parameters requires the re-collection of large amounts of data and retraining of the model, significantly increasing costs and timelines.
[0003] The root cause of the above problems lies in the fact that existing technologies generally regard the design parameters and physical fields as a one-step static mapping process. In fact, the more common approach in engineering design is to achieve the ideal design result through gradual fine-tuning based on a baseline design (for example, optimizing the airflow angle first, then adjusting the thickness), focusing on the local differences from one design sample to the next and their impact on the physical field. However, traditional model prediction methods often only focus on the "overall" relationship from design to performance, lacking targeted modeling for these small but critical local differences, resulting in the need to retrain the model covering the entire parameter space every time the parameters are expanded or the combination changes. Summary of the Invention
[0004] The object of the present invention is to provide a method and system for incremental design of turbomachinery based on a modular network architecture to solve the above problems.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] In a first aspect, the present invention provides an incremental design method for a turbomachinery based on a modular network architecture, comprising:
[0007] Define the design variables of the turbomachinery, generate a multidimensional sample combination within the design variable parameter space, perform numerical simulation on each sample, obtain the physical field distribution data corresponding to the design parameters, and form an initial training data set;
[0008] Build a modular sub-network architecture based on the initial training dataset;
[0009] Based on the initial data set, input and output data pairs are constructed by combining two samples, and multiple forward reasoning paths with different parameter adjustment orders are generated for each set of data. The prediction is performed by calling the series of sub-network modules, calculating the loss function between the predicted physical field and the actual result, and backpropagating to update the corresponding sub-network weight parameters. The training is iterated until convergence to complete the weight optimization.
[0010] A baseline design scheme is selected as the prediction starting point, and a sub-network call path is generated based on the target design parameter adjustment requirements. By connecting the sub-network modules in series in a random or fixed order, the incremental impact of parameter adjustment on the physical field is implemented step by step, and finally the complete physical field prediction results of the target design are output.
[0011] Furthermore, forming an initial training data set includes:
[0012] Define the design variables of the turbomachinery, including geometric parameters, state parameters, and target physical field type; generate multidimensional sample combinations in the design parameter space based on the Latin hypercube sampling method, and numerically simulate each sample using computational fluid dynamics or finite element methods to obtain physical field distribution data that strictly corresponds to the design parameters to form an initial training data set.
[0013] Furthermore, the construction of a modular sub-network architecture based on the initial training data set includes:
[0014] The adjustment operation of a single design parameter is mapped to an independent sub-network module, and each sub-network is only associated with the impact of the incremental change of a specific parameter on the physical field; for newly added design parameters, the corresponding sub-network module is dynamically expanded and forms a decoupled architecture with the historical sub-network to ensure that geometric parameters, state parameters and multi-physical quantity boundary conditions are handled by independent modules respectively.
[0015] Furthermore, the weight optimization is completed, including:
[0016] Data pair matching generation: construct N samples in pairs Group input and output data pairs; generate M different forward reasoning paths for each group of data The N samples generated A set of input and output data available for training;
[0017] Forward reasoning prediction: Based on the constructed path, the corresponding network module is called in turn to adjust each step Achieving result prediction
[0018] Update weight parameters: Calculate the error loss between the predicted results and the actual results, and perform backpropagation based on the loss function to update the weight parameters of each module model;
[0019] Iterative training: Repeat the above information to achieve convergence of the weight parameters of each module model.
[0020] Furthermore, the selected baseline design solution as a prediction starting point includes:
[0021] When performing actual design tasks, there is often a benchmark solution. In the method of the present invention, the benchmark design that is not included as a data set can also be used directly as a prediction starting point. If there is no existing benchmark solution for a known physical field, the Euclidean distance or similarity matching algorithm can be used to retrieve the optimal initial solution x base .
[0022] Furthermore, the step of adjusting the demand generation sub-network call path based on the target design parameters includes:
[0023] After determining the starting and end points of the inference, a path is generated for calling different networks for prediction. The calls to the sub-networks of different design steps are performed in a random order or a fixed order. The parameter adjustment is decomposed into the serial operation of independent modules, and the physical field reconstruction is realized through the serial connection of operators.
[0024] In a second aspect, the present invention provides an incremental design system for turbomachinery based on a modular network architecture, comprising:
[0025] The data acquisition module is used to define the design variables of the turbomachinery, generate multidimensional sample combinations within the design variable parameter space, perform numerical simulation on each sample, obtain the physical field distribution data corresponding to the design parameters, and form an initial training data set;
[0026] Sub-network construction module, used to build a modular sub-network architecture based on the initial training dataset;
[0027] The weight optimization module is used to construct input and output data pairs based on the initial data set by combining two samples, and generate multiple forward reasoning paths with different parameter adjustment orders for each set of data. It performs predictions by calling the series of sub-network modules, calculates the loss function between the predicted physical field and the actual result, and updates the corresponding sub-network weight parameters through backpropagation. It iterates the training until convergence to complete the weight optimization.
[0028] The prediction output module is used to select a baseline design scheme as the prediction starting point and generate a sub-network call path based on the target design parameter adjustment requirements; by connecting each sub-network module in series in a random or fixed order, the incremental impact of parameter adjustment on the physical field is implemented step by step, and finally the complete physical field prediction results of the target design are output.
[0029] Furthermore, the weight optimization is completed, including:
[0030] Data pair matching generation: construct N samples in pairs Group input and output data pairs; generate M different forward reasoning paths for each group of data The N samples generated A set of input and output data available for training;
[0031] Forward reasoning prediction: Based on the constructed path, the corresponding network module is called in turn to adjust each step Achieving result prediction
[0032] Update weight parameters: Calculate the error loss between the predicted results and the actual results, and perform backpropagation based on the loss function to update the weight parameters of each module model;
[0033] Iterative training: Repeat the above information to achieve convergence of the weight parameters of each module model.
[0034] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the incremental design method for impeller machinery based on a modular network architecture when executing the computer program.
[0035] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the incremental design method for impeller machinery based on a modular network architecture.
[0036] Compared with the prior art, the present invention has the following technical effects:
[0037] This approach achieves step-by-step progress and model scalability for physical field prediction by establishing independent network units with parameter decoupling and a dynamic constraint training strategy. Compared to traditional end-to-end models, this approach eliminates the need for global network reconstruction when adding new design parameters and significantly reduces data requirements, effectively resolving the conflict between the scarcity of high-quality data and the need for dynamic parameter expansion in engineering scenarios.
[0038] The present invention expands N samples into The proposed method significantly improves data utilization by creating trainable input-output pairs and solves the "dimensionality explosion" problem caused by parameter coupling in traditional methods. Furthermore, by dynamically adjusting the distribution of intermediate nodes, it effectively suppresses overfitting of local parameter combinations.
[0039] The present invention adopts a modular network architecture to map the adjustment process of specific design parameters into independent operator modules, thereby achieving the scalability of the design space. g+1When training is completed, only the corresponding submodule needs to be added and a small amount of incremental data is used. The input and output combination of historical data and incremental data (K×N groups of input and output pairs) is used to fine-tune only the newly added network submodule, without rebuilding the global network.
[0040] The present invention adopts a modular network architecture to process geometric information separately from various dimensional boundary conditions, reasonably deal with the diversity of different physical quantities, and avoid the dimensional distribution confusion caused by unified normalization. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 Flowchart of an embodiment of the present invention.
[0042] Figure 2 Schematic diagram of a modular network structure according to an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The following, based on the present invention and in conjunction with the embodiments, further details a system and method for incremental turbomachinery design based on a modular network architecture. The following is an application of the present invention, but is not limited thereto. Implementers may modify the parameters based on specific circumstances.
[0044] Example 1, an incremental design method for a turbomachinery based on a modular network architecture, comprising:
[0045] 1. Generate a dataset
[0046] Definition of design variables: Determine the design parameters, including design geometric parameters (such as blade angle, thickness), state parameters (such as speed, inlet pressure), x = {x1, x2, ..., x g}, θ={θ1,θ2,…,θ b The target physical field (flow field pressure distribution, temperature field) can be fully determined based on the design parameters. where u represents all combinations of physical fields within the prediction target range.
[0047] Sample collection: Latin hypercube sampling method is used to generate N parameter combination samples, covering typical working conditions within the set design parameter space to ensure data diversity.
[0048] Numerical simulation: Based on CFD (computational fluid dynamics) or finite element method, a high-precision physical field simulation model is constructed, and numerical calculations are performed on each sample to obtain the physical field distribution data U = {u1, u2, ..., u K Here we assume that a total of k different physical fields are predicted.
[0049] 2. Modular network construction
[0050] The adjustment of one or several design parameters is defined as a design operation, such as p(t,A,B) represents the adjustment x of the t-th design parameter. (t) :A→B. The adjustment of a specific design variable is a design operation, and a corresponding sub-network is established for each design operation. The role of the network is to predict the specific impact of the adjustment on the target physical field based on the adjustment process of characteristic parameters.
[0051] 3. Network weight training
[0052] Data pair matching generation: construct N samples in pairs For each set of data, M different forward reasoning paths are generated (e.g. ), so N samples are generated A set of input and output data available for training.
[0053] Forward reasoning prediction: Based on the constructed path, the corresponding network module is called in turn to adjust each step Achieving result prediction
[0054] Update weight parameters: Calculate the error loss between the predicted results and the actual results, and perform backpropagation based on the loss function to update the weight parameters of each module model.
[0055] Iterative training: Repeat the above information to achieve convergence of the weight parameters of each module model.
[0056] 4. Module Serial Inference Prediction
[0057] Benchmark solution selection: When performing actual design tasks, there is often a benchmark solution. In the method of the present invention, the benchmark design that is not included as a data set can also be used directly as a prediction starting point. If there is no existing benchmark solution for the known physical field, the Euclidean distance or similarity matching algorithm can be used to retrieve the optimal initial solution x base .
[0058] Reasoning path setting: After determining the starting and ending points of the reasoning, generate paths for calling different networks for prediction. The calls to the sub-networks of different design steps can be made in a random order or in a fixed order (such as ) is executed, and the output of the previous network becomes the input of the next network.
[0059] Module-by-module iterative prediction: decompose parameter adjustment into serial operations of independent modules, and realize physical field reconstruction through operator serial connection.
[0060] In Example 2, the specific implementation object selected by the present invention is the classic transonic compressor rotor blade cascade Rotor 37, whose geometric model mainly includes three parts: the blade body, the end wall surface and the casing surface.
[0061] The first step is numerical simulation modeling and data preparation.
[0062] First, we need to determine the scope of the study for the Rotor 37 rotor. Here, we only consider the change of the geometric parameter x and keep the state parameter α fixed. Here, the selected equipment physical field f includes static pressure p, static temperature t, absolute speed v, relative speed w, mass flow rate ρv m Five physical quantities. The equipment performance indicators Ψ were selected as flow field efficiency and pressure ratio. The state parameters α of the Rotor37 cascade are set as shown in Table 1. The geometric parameters x are defined as shown in the attached figure, where 28 parameters are used to adjust the blade geometry. A Rotor37 flow channel CFD calculation model was established based on the above state parameters and geometry.
[0063] Table 1 Rotor37 cascade geometry and operating design parameters
[0064] Working parameters Parameter value Design speed [rpm] 17188.7 Number of blades 36 Blade aspect ratio 1.19 Tip clearance [mm] 0.356 Total entrance temperature [K] 288.15 Total inlet pressure [Pa] 101325
[0065] Within the range of the determined geometric parameter x, 3000 parameter combinations are selected using Latin hypercube uniform sampling to establish 3000 samples. 3000 sets of data sets are obtained through parameter calculation of the established CFD model and are recorded as {x, α, g, f}~D.
[0066] The original data set is normalized by mean-standard deviation to obtain a normalized data set.
[0067] {x norm ,α norm ,g norm ,f norm}~D norm
[0068] For the normalized dataset D norm , randomly shuffle the input data and divide it into training set and training set according to the ratio of 4:1 and validation set
[0069] The second step is to build a modular deep operator neural network.
[0070] The specific implementation of the present invention using a modular physical field prediction network is as follows:
[0071] according to Figure 2 The structure shown in the figure can be expressed as the following formula:
[0072]
[0073] in, represents the selection mechanism of the submodule network that represents the adjustment of the design variables. i =σ(W·[x;E(P)]) (σ is the activation function and W is the learnable parameter).
[0074] Specifically in the modeling process, the encoder and decoder are constructed using multi-layer perceptrons.
[0075] The third step is modular deep operator neural network training.
[0076] For the established network structure, the constructed input and output data pairs are used to complete the network training. The network training process is expressed as follows, where represents the mean square error:
[0077] Based on the diagram structure and formula requirements, the revised training process is mathematically expressed as follows:
[0078]
[0079] k=argmax(g(E(x)))
[0080] Step 4: Carry out design tasks based on the Internet
[0081] The physical field is predicted by the trained TNO network. For a given set of geometric parameters x, input them into the trained network to obtain the predicted physical field output Then denormalize it to get the predicted physical field tensor data of the real scale
[0082] Based on the obtained physical field tensor data, the required performance index estimates can be obtained through explicit calculations Based on the obtained performance parameters, the optimization design of Rotor37 can be carried out to obtain design results under different objectives.
[0083]
[0084] In yet another embodiment of the present invention, a turbomachinery incremental design system based on a modular network architecture is provided, which can be used to implement the aforementioned turbomachinery incremental design method based on a modular network architecture. Specifically, the system includes:
[0085] The data acquisition module is used to define the design variables of the turbomachinery, generate multidimensional sample combinations within the design variable parameter space, perform numerical simulation on each sample, obtain the physical field distribution data corresponding to the design parameters, and form an initial training data set;
[0086] Sub-network construction module, used to build a modular sub-network architecture based on the initial training dataset;
[0087] The weight optimization module is used to construct input and output data pairs based on the initial data set by combining two samples, and generate multiple forward reasoning paths with different parameter adjustment orders for each set of data. It performs predictions by calling the series of sub-network modules, calculates the loss function between the predicted physical field and the actual result, and updates the corresponding sub-network weight parameters through backpropagation. It iterates the training until convergence to complete the weight optimization.
[0088] The prediction output module is used to select a baseline design scheme as the prediction starting point and generate a sub-network call path based on the target design parameter adjustment requirements; by connecting each sub-network module in series in a random or fixed order, the incremental impact of parameter adjustment on the physical field is implemented step by step, and finally the complete physical field prediction results of the target design are output.
[0089] The module division in the embodiments of the present invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in various embodiments of the present invention may be integrated into a single processor, exist physically as separate modules, or two or more modules may be integrated into a single module. The integrated modules may be implemented in either hardware or software functional modules.
[0090] In another embodiment of the present invention, a computer device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the incremental design method of impeller machinery based on a modular network architecture.
[0091] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the incremental design method for impeller machinery based on a modular network architecture in the above embodiment.
[0092] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0093] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0094] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.
[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.
[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.
Claims
1. An incremental design method for turbomachinery based on a modular network architecture, characterized in that: include: Define the design variables of the turbomachinery, generate a multidimensional sample combination within the design variable parameter space, perform numerical simulation on each sample, obtain the physical field distribution data corresponding to the design parameters, and form an initial training data set; Build a modular sub-network architecture based on the initial training dataset; Based on the initial data set, input and output data pairs are constructed by combining two samples, and multiple forward reasoning paths with different parameter adjustment orders are generated for each set of data. The prediction is performed by calling the series of sub-network modules, calculating the loss function between the predicted physical field and the actual result, and backpropagating to update the corresponding sub-network weight parameters. The training is iterated until convergence to complete the weight optimization. A baseline design scheme is selected as the prediction starting point, and a sub-network call path is generated based on the target design parameter adjustment requirements. By connecting the sub-network modules in series in a random or fixed order, the incremental impact of parameter adjustment on the physical field is implemented step by step, and finally the complete physical field prediction results of the target design are output.
2. The incremental design method for turbomachinery based on modular network architecture according to claim 1, characterized in that: The forming of the initial training data set includes: Define the design variables of the turbomachinery, including geometric parameters, state parameters, and target physical field type; generate multidimensional sample combinations in the design parameter space based on the Latin hypercube sampling method, and numerically simulate each sample using computational fluid dynamics or finite element methods to obtain physical field distribution data that strictly corresponds to the design parameters to form an initial training data set.
3. The incremental design method for turbomachinery based on modular network architecture according to claim 1, characterized in that: The modular sub-network architecture is constructed based on the initial training data set, including: The adjustment operation of a single design parameter is mapped to an independent sub-network module, and each sub-network is only associated with the impact of the incremental change of a specific parameter on the physical field; for newly added design parameters, the corresponding sub-network module is dynamically expanded and forms a decoupled architecture with the historical sub-network to ensure that geometric parameters, state parameters and multi-physical quantity boundary conditions are handled by independent modules respectively.
4. The incremental design method for turbomachinery based on modular network architecture according to claim 1, characterized in that: The weight optimization is completed, including: Data pair matching generation: construct N samples in pairs Group input and output data pairs; generate M different forward reasoning paths for each group of data The N samples generated A set of input and output data available for training; Forward reasoning prediction: Based on the constructed path, the corresponding network module is called in turn to adjust each step Achieving result prediction Update weight parameters: Calculate the error loss between the predicted results and the actual results, and perform backpropagation based on the loss function to update the weight parameters of each module model; Iterative training: Repeat the above information to achieve convergence of the weight parameters of each module model.
5. The incremental design method for turbomachinery based on modular network architecture according to claim 1, characterized in that: The selected baseline design solution is used as a starting point for prediction, including: When performing actual design tasks, there is often a benchmark solution. In the method of the present invention, the benchmark design that is not included in the solution as a data set can also be used directly as a prediction starting point. If there is no existing benchmark solution with a known physical field, the Euclidean distance or similarity matching algorithm can be used to retrieve the optimal initial solution x. base .
6. The incremental design method for turbomachinery based on modular network architecture according to claim 1, characterized in that: The step of adjusting the demand generation sub-network call path based on the target design parameters includes: After determining the starting and end points of the inference, a path is generated for calling different networks for prediction. The calls to the sub-networks of different design steps are performed in a random order or a fixed order. The parameter adjustment is decomposed into the serial operation of independent modules, and the physical field reconstruction is achieved through the serial connection of operators.
7. The incremental design system for turbomachinery based on modular network architecture is characterized by: include: The data acquisition module is used to define the design variables of the turbomachinery, generate multidimensional sample combinations within the design variable parameter space, perform numerical simulations on each sample, obtain the physical field distribution data corresponding to the design parameters, and form an initial training data set; Sub-network construction module, used to build a modular sub-network architecture based on the initial training dataset; The weight optimization module is used to construct input and output data pairs based on the initial data set by combining two samples, and generate multiple forward reasoning paths with different parameter adjustment orders for each set of data. It performs predictions by calling the series of sub-network modules, calculates the loss function between the predicted physical field and the actual result, and updates the corresponding sub-network weight parameters through backpropagation. It iterates the training until convergence to complete the weight optimization. The prediction output module is used to select a baseline design scheme as the prediction starting point and generate a sub-network call path based on the target design parameter adjustment requirements; by connecting each sub-network module in series in a random or fixed order, the incremental impact of parameter adjustment on the physical field is implemented step by step, and finally the complete physical field prediction results of the target design are output.
8. The turbomachinery incremental design system based on modular network architecture according to claim 7, characterized in that: The weight optimization is completed, including: Data pair matching generation: construct N samples in pairs Group input and output data pairs; generate M different forward reasoning paths for each group of data The N samples generated A set of input and output data available for training; Forward reasoning prediction: Based on the constructed path, the corresponding network module is called in turn to adjust each step Achieving result prediction Update weight parameters: Calculate the error loss between the predicted results and the actual results, and perform backpropagation based on the loss function to update the weight parameters of each module model; Iterative training: Repeat the above information to achieve convergence of the weight parameters of each module model.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the incremental design method of impeller machinery based on modular network architecture as claimed in any one of claims 1 to 6 are implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the incremental design method for turbomachinery based on a modular network architecture as claimed in any one of claims 1 to 6 are implemented.