Neural network automatic training system and method

Abstract

The invention discloses an automatic training system and method for a neural network, and relates to the field of network model training, and the system comprises a parameter configuration module which is used for loading a configuration file, displaying the configuration file in a tree structure in a visual interaction interface, interactively editing the configuration file, and storing the edited configuration file; the data preparation module is used for extracting actual data based on the stored configuration file, interactively obtaining a data set and further integrating model training preparation data; the model training module is used for searching an optimal architecture and hyper-parameter combination of a neural network model by using a genetic algorithm; and the reasoning prediction module is used for loading the optimal architecture and hyper-parameter combination of the neural network model, inputting the received user data into the neural network model and further returning a prediction result. According to the invention, the training efficiency of the neural network model is effectively improved.

Description

Technical Field

[0001] This application relates to the field of network model training, and in particular to an automated neural network training system and method. Background Technology

[0002] In current neural network model training processes, manual configuration of hyperparameters, adjustment of network structure, and repeated trial and error are typically relied upon. This is not only inefficient but also demands a high level of expertise from the user. Especially in complex application scenarios such as engineering simulations, where data dimensionality is high and nonlinearity is strong, traditional training methods struggle to quickly build high-precision, high-generalization models, thus hindering the practical application of neural networks in real-world engineering problems.

[0003] Furthermore, existing training systems generally lack a unified visual configuration and automated optimization mechanism, resulting in chaotic configuration file management, untraceable training processes, and long model iteration cycles. Although some tools have introduced automated parameter tuning techniques, they still have significant shortcomings in multi-objective optimization, integration with CAE simulation data, and training-prediction closed-loop feedback, making it difficult to meet the demands of intelligent engineering design for efficient and accurate modeling. Summary of the Invention

[0004] The purpose of this invention is to provide an automated neural network training system and method, which solves at least one of the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] According to one aspect of the present invention, an automated neural network training system is provided, comprising:

[0007] The parameter configuration module is used to load configuration files and display them in a tree structure within a visual interactive interface, allowing users to interactively edit and save the edited configuration files.

[0008] The data preparation module is used to extract actual data based on the saved configuration file, interactively obtain the dataset, and then integrate the data for model training preparation.

[0009] The model training module is used to search for the optimal architecture and hyperparameter combination of a neural network model using a genetic algorithm.

[0010] The inference and prediction module is used to load the optimal architecture and hyperparameter combination of the neural network model, input the received user data into the neural network model, and then return the prediction results.

[0011] Optionally, in the parameter configuration module, the loaded configuration file is a YAML format configuration file, and the configuration file is parsed to display the parsed configuration file in a tree structure on the visual interactive interface.

[0012] Optionally, the data preparation module includes:

[0013] The path parsing unit is used to extract the actual data of the path parameters based on the values ​​of the path parameters, and to use the actual data of the path parameters and the determined parameters as a preloaded training file;

[0014] The data preprocessing unit is used to interactively obtain the dataset and divide the dataset into a set of independent variable data and a set of dependent variable data, and then divide the dataset into a training set and a validation set.

[0015] Optionally, the model training module includes:

[0016] The population initialization unit is used to initialize the population based on the data prepared for model training, drive the iterative optimization of the population, and use the iteratively optimized population as the optimal architecture and hyperparameter combination of the neural network model.

[0017] The multi-objective optimization unit is used to calculate the individual fitness of the neural network model and optimize the process of searching for the optimal architecture of the neural network model based on the calculation results.

[0018] The population initialization unit is used to complete the population initialization and drive the population iterative optimization.

[0019] The elite retention unit is used to identify elite individuals using an elite retention strategy and select the best individual using a tournament selection strategy. The elite individuals and the best individual are then used as the iteration results to enter the next generation of the population.

[0020] Optionally, in the data preprocessing unit, the set of independent variable data is set as X={x1,x2......xn}, and the set of dependent variable data is set as Y={y1,y2...,ym}, where x1, x2, and xn are the 1st, 2nd, and nth independent variable data, respectively, and y1, y2, and ym are the 1st, 2nd, and mth dependent variable data, respectively.

[0021] The dataset is an engineering design dataset generated by CAE simulation technology. In the data preprocessing unit, the primary data is used as the independent variable data, and the secondary data obtained by applying the primary data is used as the dependent variable data. The primary data is the basic data to ensure the construction of the project, and the secondary data is the data obtained by CAE experimental verification based on the primary data.

[0022] Optionally, in the multi-objective optimization unit, a multi-objective fitness function is set by combining the fitness function with weights, and the training set is trained with the multi-objective fitness function. Then, the multi-objective fitness function index between the test sets is calculated, and the calculation result is fed back to the genetic algorithm as the individual fitness, thereby optimizing the process of searching for the optimal architecture of the neural network model.

[0023] Optionally, the process of setting the multi-objective fitness function is as follows: minimizing the mean squared error (MSE) and the coefficient of determination (R²) 2 The results of maximizing the objective and minimizing the mean absolute error (MAPE) are used as weights to construct a multi-objective fitness function.

[0024] Optionally, the population initialization unit is used to treat each configuration file parameter as a set of candidate parameters, and to define the population search space with the genetic algorithm-related parameters in the saved configuration file, and then complete the population initialization with the population search space and candidate parameters.

[0025] Optionally, the process of driving the population iterative optimization is as follows: for integer parameters among the candidate parameters, the mutation amplitude is adjusted according to the gradient based on the search space range and the user-specified mutation scaling factor; for floating-point parameters, the mutation amplitude is dynamically adjusted by combining random perturbation with the user-specified mutation scaling factor.

[0026] On the other hand, this application also provides an automated training method for neural networks, including: step S1, obtaining a dataset and dividing the dataset into a set of independent variable data and a set of dependent variable data;

[0027] Step S2: Load the configuration file using a visual interactive interface and parse the configuration file;

[0028] Step S3: Train the neural network model through the visual interface, the partitioned dataset and the parsed configuration file, and obtain the training log, then display the training log on the visual interface.

[0029] Step S4: Based on the trained neural network model and the parsed configuration file, the received user data is fed into the neural network model, and then the prediction result is returned.

[0030] Step S5: Modify the corresponding data in the test set with the prediction results, and return to the training process of the neural network model for iteration.

[0031] Compared with existing technologies, the advantages of this invention are as follows: This solution significantly improves the efficiency of model development and application by constructing an automated neural network training system that integrates parameter configuration, model training, and inference prediction. The visual interactive interface supports tree-like display and dynamic editing of YAML configuration files, lowering the barrier to entry. Combined with genetic algorithms and multi-objective optimization strategies, it can automatically search for the optimal network structure and hyperparameter combination, balancing model accuracy and generalization ability. The system is particularly well-suited for engineering data generated by CAE simulations, achieving efficient mapping between primary and secondary data, and accelerating convergence through elite retention and dynamic mutation mechanisms. Furthermore, real-time feedback of training logs and a closed-loop iterative mechanism for prediction results further enhance the model's interpretability and continuous optimization capabilities, providing an efficient, stable, and automated modeling solution for intelligent engineering design. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the structure of an automated neural network training system provided in this application.

[0033] Figure 2 This is a structural diagram of the data preparation module provided in this application.

[0034] Figure 3 This is a schematic diagram of the structure of the model training module provided in this application.

[0035] Figure 4 This is a flowchart of an automated neural network training method provided in this application. Detailed Implementation

[0036] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0037] It should be noted that although the terms first, second, third, etc., may be used in the embodiments of this application for description, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of this application, first can also be referred to as second, and similarly, second can also be referred to as first.

[0038] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0039] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Specifically, the neural network automated training system and method provided in this application are applicable to application scenarios where CAE simulation technology is used to construct neural network models from data for engineering problems, so as to realize the integration of neural networks and CAE. The neural network automated training system and method described in this application aim to realize one-click application of automated training and inference prediction of neural network models by applying a neural network automated training system driven by genetic algorithms and configuration files.

[0041] To apply the above-mentioned application scenarios, this application provides an automated neural network training system, the structural diagram of which can be found in the attached diagram. Figure 1 As shown, it includes:

[0042] The parameter configuration module is used to load configuration files and display them in a tree structure within a visual interactive interface, allowing users to interactively edit and save the configuration files.

[0043] Specifically, in the parameter configuration module, the loaded configuration file is a YAML format configuration file, and the configuration file is parsed to display the parsed configuration file in a tree structure on the visual interactive interface; the tree structure is a display method that supports hierarchical expansion and collapse.

[0044] Specifically, the YAML-formatted configuration file described in this application contains various configuration file parameters. The data format of these parameters is a key-value pair data format of ["keyword", "value", "remark"]. Each key-value pair represents a configuration file parameter, where the keyword is the name of the configuration file parameter, the value is the numerical value of the configuration file parameter, and the remarks are explanations of the configuration file parameter to help users understand it. It is worth noting that the numerical values ​​of the configuration file parameters described in this application are numerical values ​​in the computer field, and they can be non-numeric values, such as numerical ranges, file paths, string collections, etc.

[0045] For example, the configuration file parameters in this application include: absolute path of the project space, data-related parameters, genetic algorithm optimization-related parameters, hyperparameters of the neural network to be optimized, and output-related parameters; the data-related parameters include: relative path of training data, path of test data, column numbers of independent and dependent variables in the dataset, and proportion of validation set; the genetic algorithm optimization-related parameters include: population size, number of iterations, mutation ratio and probability, and crossover probability; the neural network hyperparameters to be optimized include: number of hidden layers, number of neurons per layer, dropout rate, optimizer, learning rate, and batch size; the output-related parameters include: model save path, statistical information output path of the genetic algorithm optimization process, log save path, and normalizer save path.

[0046] Specifically, in the parameter configuration module, the interactive editing process of the configuration file is as follows: double-click the configuration file parameter in the visual interactive interface and enter the value of the configuration file parameter to complete the editing of the configuration file.

[0047] Specifically, in the parameter configuration module, the process of saving the edited configuration file is as follows:

[0048] The configuration file parameters in the edited configuration file are classified into path parameters and defined parameters; the path parameters are configuration file parameters whose values ​​are stored in a specific path, and the defined parameters are configuration file parameters whose values ​​are not stored in a specific path.

[0049] Save the edited configuration file in YAML format.

[0050] Specifically, the parameter configuration module significantly lowers the barrier to entry for training neural network models by loading YAML-formatted configuration files and displaying them in a tree structure within a visual interactive interface. Users can intuitively view, edit, and save various training parameters, including data paths, hyperparameter ranges, and genetic algorithm settings, without writing code. This module supports categorized management of configuration items (such as path parameters and fixed parameters) and allows direct modification of parameter values ​​via double-clicking, greatly improving configuration efficiency and accuracy. Simultaneously, the structured configuration files facilitate version control, experimental reproduction, and team collaboration, effectively solving the problems of chaotic and untraceable configurations in traditional training processes, laying a clear and reliable foundation for subsequent automated training workflows.

[0051] Please continue reading. Figure 1 As shown, the automated neural network training system also includes:

[0052] The data preparation module is used to extract actual data based on the saved configuration file, interactively obtain the dataset, and then integrate the data for model training preparation.

[0053] Build a neural network model and train it.

[0054] Please see Figure 2 The diagram shown is a structural schematic of the data preparation module provided in this application, including:

[0055] The path parsing unit is used to extract the actual data of the path parameters based on the values ​​of the path parameters, and to use the actual data of the path parameters and the determined parameters as a preloaded training file.

[0056] Specifically, in the path parsing unit, the process of "extracting the actual data of the path parameter based on the value of the path parameter" involves searching the computer file system using the value of the path parameter as the file path, and using the search result as the actual data of the path parameter.

[0057] Specifically, the path resolution unit automatically retrieves and extracts the corresponding actual data files from the file system based on the path parameters specified in the configuration file, achieving automation and intelligence in data loading. This unit integrates path parameters and deterministic parameters into a preloaded training file, effectively decoupling data location from training logic and improving the system's flexibility and portability. Users only need to specify the path in the visual interface, and the system can automatically complete data location and reading, avoiding errors that may be caused by manually entering the path. This automation mechanism not only saves manual operation time but also ensures the consistency and reliability of data sources, providing a stable and efficient data input foundation for subsequent data preprocessing and model training.

[0058] Please continue reading. Figure 2 As shown, the data preparation module further includes:

[0059] The data preprocessing unit is used to interactively acquire the dataset and divide the dataset into a set of independent variable data and a set of dependent variable data. Then, the dataset is divided into a training set and a validation set. The divided dataset, the actual data of the path parameters, and the determined parameters are used as the data to prepare for model training.

[0060] Specifically, in the data preprocessing unit, the set of independent variable data is set as X={x1,x2......xn}, and the set of dependent variable data is set as Y={y1,y2...,ym}, where x1, x2, and xn are the 1st, 2nd, and nth independent variable data, respectively, and y1, y2, and ym are the 1st, 2nd, and mth dependent variable data, respectively.

[0061] The dataset is an engineering design dataset generated by CAE simulation technology. In the data preprocessing unit, the primary data is used as the independent variable data, and the secondary data obtained by applying the primary data is used as the dependent variable data. The primary data in this application is the basic data to ensure engineering construction, and the secondary data is the data obtained by CAE experimental verification based on the primary data.

[0062] For example, the dataset of independent variables in this application can be {structure, material, load}, and the dataset of dependent variables can be {stiffness in the first direction, height in the second direction}.

[0063] Specifically, the data preprocessing unit can interactively acquire datasets and automatically divide them into sets of independent and dependent variables, further segmenting them into training and validation sets to construct well-structured model input data. This unit is particularly well-suited for engineering design data generated by CAE simulations, using primary data (such as structure, materials, and loads) as input and secondary data (such as stiffness and deformation) as output, precisely meeting the needs of engineering modeling. Through a standardized data partitioning process, this unit ensures the scientific rigor of the training process and the objectivity of model evaluation, effectively handling complex engineering data with high dimensionality and strong nonlinearity. Simultaneously, its interactive design allows users to flexibly adjust the partitioning ratio, balancing automation and controllability, providing solid data support for high-quality neural network training.

[0064] Please continue reading. Figure 1 As shown, the automated neural network training system also includes:

[0065] The model training module is used to search for the optimal architecture and hyperparameter combination of a neural network model using a genetic algorithm.

[0066] Please see Figure 3 As shown, it is a schematic diagram of the structure of the model training module described in this application, including:

[0067] The population initialization unit is used to initialize the population based on the data prepared for model training, drive the iterative optimization of the population, and use the iteratively optimized population as the optimal architecture and hyperparameter combination of the neural network model.

[0068] Specifically, the population initialization unit is used to treat each configuration file parameter as a set of candidate parameters, and to define the population search space with the genetic algorithm-related parameters in the saved configuration file, and then complete the population initialization with the population search space and candidate parameters.

[0069] The process of driving population iterative optimization is as follows: For integer parameters among the candidate parameters, the mutation amplitude is adjusted according to the gradient based on the search space range and the user-specified mutation scaling factor; for floating-point parameters, the mutation amplitude is dynamically adjusted by combining random perturbation with the user-specified mutation scaling factor.

[0070] Specifically, the genetic algorithm-related parameters are the parameters related to the genetic algorithm stored in the configuration file; the search space range and the user-specified mutation ratio factor are the configuration file parameters that are interactively edited by the user.

[0071] Specifically, in the population initialization unit described in this application, the process of "adjusting the mutation amplitude according to the gradient" is not specifically defined. In an exemplary embodiment, the product of the length of the search space range and the mutation ratio factor can be used as the mutation amplitude. Meanwhile, the random perturbation described in this application is a Gaussian perturbation value generated using a Gaussian distribution. The process of "dynamically adjusting the mutation amplitude by combining the random perturbation with the user-specified mutation ratio factor" is as follows: the product of the length of the search space range and the mutation ratio factor is used as the basic perturbation scale, and then the mutation amplitude is dynamically adjusted by combining the basic perturbation scale with the random perturbation. The implementation process is the existing technology of dynamic adjustment in existing genetic algorithms, which will not be elaborated in this application.

[0072] Specifically, the population initialization unit automatically generates an initial population based on the genetic algorithm parameters and hyperparameter search space defined in the configuration file, laying the foundation for the optimization of the neural network architecture and hyperparameters. This unit treats each configured parameter as a candidate solution and, in conjunction with a user-specified mutation rate factor, adjusts the mutation magnitude using gradients for integer parameters and dynamically adjusts the mutation intensity using Gaussian random perturbations for floating-point parameters. This differentiated mutation strategy balances the breadth and accuracy of the search, avoiding getting trapped in local optima. Through reasonable initialization and adaptive mutation mechanisms, this unit significantly improves the convergence speed and search efficiency of the genetic algorithm, reduces invalid training iterations, and provides a high-quality initial population foundation for subsequent efficient and stable optimization.

[0073] Please continue reading. Figure 3 As shown, the model training module further includes:

[0074] The multi-objective optimization unit is used to calculate the individual fitness of the neural network model and optimize the process of searching for the optimal architecture of the neural network model based on the calculation results.

[0075] Specifically, in the multi-objective optimization unit, the fitness function is set by weight combination, and the training set is trained with the multi-objective fitness function. Then, the multi-objective fitness function index between the test sets is calculated, and the calculation result is fed back to the genetic algorithm as the individual fitness, thereby optimizing the process of searching for the optimal architecture of the neural network model.

[0076] The process of setting the multi-objective fitness function is as follows: minimizing the mean squared error (MSE) and the coefficient of determination (R²) 2The results of maximizing the objective and minimizing the mean absolute error (MAPE) are used as weights to construct a multi-objective fitness function.

[0077] Specifically, the multi-objective optimization unit comprehensively evaluates the overall performance of individual neural networks by constructing a multi-objective fitness function that integrates minimizing mean squared error (MSE), maximizing the coefficient of determination (R²), and minimizing mean absolute percentage error (MAPE). This unit not only optimizes the model on the training set but also provides fitness metrics through the test set, effectively preventing overfitting and improving the model's generalization ability. Users can flexibly adjust the weights of each objective according to engineering needs to achieve customized optimization goals. This multi-dimensional evaluation mechanism avoids the bias caused by single-index optimization, enabling the final model to achieve a good balance between accuracy, stability, and interpretability, significantly improving the predictive reliability and practical value of neural networks in complex engineering scenarios.

[0078] Please continue reading. Figure 3 As shown, the model training module further includes:

[0079] The elite retention unit is used to determine elite individuals using an elite retention strategy and select the best individual using a tournament selection strategy. The elite individuals and the best individual are then used as the iteration results to enter the next generation population. The next generation population is the input data of the next iteration layer of the genetic algorithm.

[0080] Specifically, the elite retention unit employs a combination of elite retention and tournament selection strategies to ensure that high-quality individuals are not lost during evolution and to efficiently guide the population towards a better solution. This unit directly retains the most fit elite individuals from each generation to the next, while simultaneously using a tournament mechanism to select high-performing individuals from the remaining population for crossover and mutation. This ensures both the stability of the optimal solution and the diversity of the population. This hybrid selection mechanism effectively avoids the premature convergence problem common in genetic algorithms, accelerating the optimization process while enhancing global search capabilities. Through the synergistic effect of this unit, the model training module can obtain a high-precision, highly robust neural network architecture and hyperparameter combination within a fewer iterations.

[0081] Please continue reading. Figure 1 The aforementioned automated neural network training system further includes:

[0082] The inference and prediction module is used to load the optimal architecture and hyperparameter combination of the neural network model, input the received user data into the neural network model, and then return the prediction result; the received user data is a set of independent variable data for prediction input by user interaction.

[0083] Specifically, the inference and prediction module can load the optimal neural network model optimized by a genetic algorithm, receive user-input independent variable data, and quickly return prediction results, achieving one-click deployment from training to application. This module supports seamless integration with a visual interface, allowing users to complete prediction operations without programming, greatly improving model usability and engineering implementation efficiency. More importantly, prediction results can be fed back into subsequent training processes, forming a closed-loop iterative mechanism of "prediction-validation-retraining" to continuously optimize model performance. This module is particularly suitable for rapid performance evaluation and design parameter optimization scenarios after CAE simulation, significantly shortening the engineering decision-making cycle and promoting the efficient application of intelligent design in real-world projects.

[0084] This application also provides an automated training method for neural networks, the flowchart of which can be found in the document. Figure 4 As shown, it includes:

[0085] Step S1: Obtain the dataset and divide it into an independent variable data set and a dependent variable data set. Then, further divide the dataset into a training set and a validation set. The independent variable data set is set as X={x1,x2......xn}, and the dependent variable data set is set as Y={y1,y2...,ym}, where x1, x2, and xn are the 1st, 2nd, and nth independent variable data, respectively, and y1, y2, and ym are the 1st, 2nd, and mth dependent variable data, respectively.

[0086] Please continue reading. Figure 4 As shown, the automated neural network training method further includes:

[0087] Step S2: Load the configuration file using a visual interactive interface and parse the configuration file.

[0088] The configuration file stores data in a key-value pair format of ["keyword", "value", "remark"]. Each key-value pair represents a configuration file parameter, where the keyword is the name of the configuration file parameter, the value is the numerical value of the configuration file parameter, and the remarks are explanations of the configuration file parameter to help users understand it. It is worth noting that the numerical values ​​of the configuration file parameters in this application are numerical values ​​in the computer field, which can be non-numeric values, such as numerical ranges, file paths, string sets, etc.

[0089] For example, the configuration file parameters described in this application include:

[0090] Please continue reading. Figure 4 As shown, the automated neural network training method further includes:

[0091] Step S3: Train the neural network model through the visual interface, the partitioned dataset, and the parsed configuration file, and obtain the training logs, then display the training logs on the visual interface.

[0092] Specifically, the training process for the neural network model has already been described in the aforementioned automated neural network training system, and will not be repeated here.

[0093] Please continue reading. Figure 4 As shown, the automated neural network training method further includes:

[0094] Step S4: Based on the trained neural network model and the parsed configuration file, the received user data is fed into the neural network model, and then the prediction result is returned; the prediction result is a set of dependent variable data obtained by predicting the received user data, and its format is a predictions.csv file.

[0095] Please continue reading. Figure 4 As shown, the automated neural network training method further includes:

[0096] Step S5: Add the prediction results to the received user data for the next prediction process and repeat the prediction.

[0097] The neural network automated training method provided in this application can execute the neural network automated training system provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the execution method.

[0098] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An automated neural network training system, characterized in that, include: The parameter configuration module is used to load configuration files and display them in a tree structure within a visual interactive interface, allowing users to interactively edit and save the edited configuration files. The data preparation module is used to extract actual data based on the saved configuration file, interactively obtain the dataset, and then integrate the data for model training preparation. The model training module is used to search for the optimal architecture and hyperparameter combination of a neural network model using a genetic algorithm. The inference and prediction module is used to load the optimal architecture and hyperparameter combination of the neural network model, input the received user data into the neural network model, and then return the prediction results.

2. The automated neural network training system according to claim 1, characterized in that, In the parameter configuration module, the loaded configuration file is a YAML format configuration file, and the configuration file is parsed to display the parsed configuration file in a tree structure on the visual interactive interface.

3. The automated neural network training system according to claim 1, characterized in that, The data preparation module includes: The path parsing unit is used to extract the actual data of the path parameters based on the values ​​of the path parameters, and to use the actual data of the path parameters and the determined parameters as a preloaded training file; The data preprocessing unit is used to interactively obtain the dataset and divide the dataset into a set of independent variable data and a set of dependent variable data, and then divide the dataset into a training set and a validation set.

4. The automated neural network training system according to claim 1, characterized in that, The model training module includes: The population initialization unit is used to initialize the population based on the data prepared for model training, drive the iterative optimization of the population, and use the iteratively optimized population as the optimal architecture and hyperparameter combination of the neural network model. The multi-objective optimization unit is used to calculate the individual fitness of the neural network model and optimize the process of searching for the optimal architecture of the neural network model based on the calculation results. The population initialization unit is used to complete the population initialization and drive the population iterative optimization. The elite retention unit is used to identify elite individuals using an elite retention strategy and select the best individual using a tournament selection strategy. The elite individuals and the best individual are then used as the iteration results to enter the next generation of the population.

5. The automated neural network training system according to claim 3, characterized in that, In the data preprocessing unit, the set of independent variable data is set as X={x1,x2......xn}, and the set of dependent variable data is set as Y={y1,y2...,ym}, where x1, x2, and xn are the 1st, 2nd, and nth independent variable data, respectively, and y1, y2, and ym are the 1st, 2nd, and mth dependent variable data, respectively. The dataset is an engineering design dataset generated by CAE simulation technology. In the data preprocessing unit, the primary data is used as the independent variable data, and the secondary data obtained by applying the primary data is used as the dependent variable data. The primary data is the basic data to ensure the construction of the project, and the secondary data is the data obtained by CAE experimental verification based on the primary data.

6. The automated neural network training system according to claim 4, characterized in that, In the multi-objective optimization unit, the fitness function is set by weight combination, and the training set is trained with the multi-objective fitness function. Then, the multi-objective fitness function index between the test set is calculated, and the calculation result is fed back to the genetic algorithm as the individual fitness, thereby optimizing the process of searching for the optimal architecture of the neural network model.

7. The automated neural network training system according to claim 6, characterized in that, The process of setting the multi-objective fitness function is as follows: minimizing the mean squared error (MSE) and the coefficient of determination (R²) 2 The results of maximizing the objective and minimizing the mean absolute error (MAPE) are used as weights to construct a multi-objective fitness function.

8. The automated neural network training system according to claim 7, characterized in that, The population initialization unit is used to treat each configuration file parameter as a set of candidate parameters, and to define the population search space with the genetic algorithm-related parameters in the saved configuration file, and then complete the population initialization with the population search space and candidate parameters.

9. The automated neural network training system according to claim 8, characterized in that, The process of driving population iterative optimization is as follows: For integer parameters among the candidate parameters, the mutation amplitude is adjusted according to the gradient based on the search space range and the user-specified mutation scaling factor; for floating-point parameters, the mutation amplitude is dynamically adjusted by combining random perturbation with the user-specified mutation scaling factor.

10. An automated neural network training method, applied to the automated neural network training system as described in any one of claims 1-9, characterized in that, include: Step S1: Obtain the dataset and divide the dataset into a set of independent variable data and a set of dependent variable data; Step S2: Load the configuration file using a visual interactive interface and parse the configuration file; Step S3: Train the neural network model through the visual interface, the partitioned dataset and the parsed configuration file, and obtain the training log, then display the training log on the visual interface. Step S4: Based on the trained neural network model and the parsed configuration file, the received user data is fed into the neural network model, and then the prediction result is returned. Step S5: Add the prediction results to the received user data for the next prediction process and repeat the prediction.

Images