Engine system weight prediction method, system and equipment and storage medium

Abstract

The invention belongs to the field of engine system weight prediction, and particularly relates to an engine system weight prediction method, system and device and a storage medium, and the method comprises the steps: building a mapping relation between engine design parameters and system weights corresponding to the design parameters, and obtaining an original data set; inputting into the generative adversarial network for training, and optimizing through an evaluation value generated by the generator and a discriminant value generated by the discriminator to obtain an optimized generative adversarial network; generating a training data set, inputting the training data set into the neural network model to obtain a loss value, obtaining an optimized neural network model by minimizing the loss value, and extracting potential features; and inputting the potential features into the prediction model, minimizing the target function to obtain an optimal prediction model, and inputting the actual engine design parameters into the optimal prediction model to obtain the actual system weight. Under the condition that engine design data is scarce, the prediction accuracy of engine weight prediction can be remarkably improved by enhancing training samples and extracting potential features.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of engine system weight prediction, and particularly relates to an engine system weight prediction method, system, device and storage medium. BACKGROUND

[0002] The essence of flight is to overcome gravity, which makes the system-level weight one of the most critical performance parameters in aircraft design. Due to the strong coupling relationship between design variables, the prediction error in the early stage may be transmitted to the judgment error of structural size, propulsion system demand and load capacity. A well-known example is the F-35B project, whose empty weight is finally about 10% higher than the predicted value, due to the underestimation of the weight of its new equipment. This directly leads to the decrease of payload capacity and the restriction of load factor. As an important part of the aircraft system, the weight of the engine system plays a core role in affecting the center of gravity position, lift, range and fuel efficiency. Therefore, accurately and efficiently predicting the weight of the engine system at the conceptual design stage is crucial to ensure the effectiveness of design and development work.

[0003] With the rapid development of artificial intelligence technology, machine learning methods have attracted widespread attention in the application of aero-engine performance prediction and analysis. Such methods exhibit significant advantages in modeling simplicity and computational efficiency. Unlike traditional regression or physical models, machine learning techniques do not require explicit construction of physical laws or feature engineering, but learn statistical laws directly from data, thereby enabling efficient and potentially more accurate prediction in high-dimensional and complex design spaces. In addition, many machine learning models have low computational overhead in the inference stage, making them very suitable for design iterations and real-time performance prediction applications that require high time efficiency.

[0004] In view of the above related technologies, machine learning requires a large number of public high-quality data sets, while the weight of the engine system is due to security reasons, and there are few public data sets, so that the model training effect is poor, and the prediction accuracy is low. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an engine system weight prediction method, system, device and storage medium, by applying constraints to the generative adversarial network, making the generated generative data set more realistic, then training and optimizing using the generative data set prediction model, obtaining an optimized prediction model, so that the prediction result is more accurate.

[0006] Obtain engine design parameters and system weight corresponding to the design parameters; Establish a mapping relationship between the engine design parameters and the system weight corresponding to the design parameters to obtain an original data set; inputting the original data set into a generative adversarial network to obtain a generated data set, the generative adversarial network comprising a generator and a discriminator, setting a constraint condition for the generator to enable the generator to evaluate the generated data set to obtain an evaluation value, the discriminator discriminating the generated data set to obtain a discrimination value, optimizing the generative adversarial network according to the discrimination value and the evaluation value to obtain an optimized generative adversarial network; obtaining an optimized generated data set according to the optimized generative adversarial network and the original data set, and combining the optimized generated data set and the original data set into a training data set; inputting the training data set into a neural network model, extracting a hidden layer feature, inputting the hidden layer feature into an activation function to obtain an output value, calculating a loss value according to the output value and a true value, optimizing model parameters of the neural network model by minimizing the loss value to obtain an optimized neural network model, and the true value being a system weight in the training data set; obtaining a latent feature according to the optimized neural network model and the training data set; inputting the latent feature into a prediction model to obtain a prediction value, constructing an objective function according to the prediction value and the true value, obtaining an adjustment parameter when a target value of the objective function is minimum, optimizing the prediction model according to the adjustment parameter to obtain an optimized prediction model; inputting an actual engine design parameter into the optimized prediction model to obtain an actual system weight.

[0007] Optionally, the inputting the original data set into a generative adversarial network to obtain a generated data set, the generative adversarial network comprising a generator and a discriminator, setting a constraint condition for the generator to enable the generator to evaluate the generated data set to obtain an evaluation value, the discriminator discriminating the generated data set to obtain a discrimination value, optimizing the generative adversarial network according to the discrimination value and the evaluation value to obtain an optimized generative adversarial network, comprises: inputting the original data set into a generative adversarial network to obtain a generated data set; the generator evaluating the generated data set to obtain an evaluation value, which is expressed as: wherein, the generator, the condition, the generator generates the generated data set under the condition , , , the prior distribution of the generated data set, the generator is sampled from the generated data set combination, It is a discriminator. This indicates that for generators The generated data set Under given conditions The probability that the generated data set is derived from the original data set is given, and the range of values ​​is given. , This indicates that under random sampling, according to generated samples Input discriminator In the middle, the discriminator gives the probability of being true and takes the expected value of the logarithm; The discriminator performs a discrimination operation on the generated data group to obtain a discrimination value, which is represented as follows: in, express It was sampled from the original dataset. For the discriminator pair The output of the probabilities from the original data set is logarithmic; For discriminator right The expected value of the logarithm of the probability of being true given from the original data set;

[0008] Based on the discriminant value and the evaluation value, the generative adversarial network is optimized to obtain an optimized generative adversarial network, represented as: in, Indicates the distribution from the original data set The real samples obtained from sampling.

[0009] Optionally, the step of inputting the training data set into the neural network model, extracting hidden layer features, passing the hidden layer features through an activation function to obtain output values, calculating a loss value based on the output value and the true value, and optimizing the model parameters of the neural network model by minimizing the loss value to obtain an optimized neural network model includes: The training data set is input into the neural network model to extract hidden layer features. , is represented as: in, For activation function, To hide the number of layers, For the first The weight matrix of the hidden layer, For the first The bias vector of the hidden layer. For the training data set; The hidden layer features are passed through an activation function to obtain the output value; Based on the output value and the true value, the loss value is calculated and expressed as: in, The loss value. This is the weight matrix. For bias vectors, For the input sample, For real labels, For the output value, Denotes the Euclidean norm; An optimized neural network model is obtained by optimizing the model parameters of the neural network model by minimizing the loss value.

[0010] Optionally, based on the optimized neural network model and the training data set, the latent feature representation is obtained as follows: in, For the first The weight matrix of the hidden layer, For the first Hidden layer output of the layer, and It is a constant. For the first The bias vector of the hidden layer. These are potential features.

[0011] Optionally, the step of inputting the latent features into the prediction model to obtain predicted values, constructing an objective function based on the predicted values ​​and the true values, obtaining adjustment parameters when the objective function's target value is minimized, and optimizing the prediction model based on the adjustment parameters to obtain an optimized prediction model includes: The latent features are input into the prediction model to obtain the predicted value, which is expressed as: in, For predicted values, Indicates the number of regression trees. For the function space of the regression tree, For the first Tree samples The predicted value; Based on the predicted and actual values, an objective function is constructed, expressed as: in, This represents the error between the predicted and actual values. This represents the total number of engine samples. for The sum of the regular terms of the tree is expanded as: wherein, is the th iteration, i.e., the th optimization of the tree, and is an adjustment parameter, is the number of leaves, is the output value of the th leaf; obtaining the adjustment parameter at which the target value of the target function is minimum, and optimizing the prediction model according to the adjustment parameter to obtain an optimized prediction model.

[0012] Optionally, the obtaining the adjustment parameter at which the target value of the target function is minimum, and optimizing the prediction model according to the adjustment parameter includes: the th tree target function can be expressed as: wherein, is the th tree target function value, is the prediction value output by the th tree, is the true value, is the prediction value of the th tree for the sample , is a constant term, is the total number of engine samples, is the regular term of the th tree, is the prediction error value of the th tree; Taylor expanding the target function expression of the th tree to obtain a Taylor expansion expression, and simplifying as follows: wherein, and respectively represent the first and second derivatives of the loss function, is a set of all samples belonging to the th leaf, is the th leaf, is the th sample, and the partial derivative of the Taylor expansion expression with respect to the th sample is taken and set to 0 to obtain the optimal output value the minimum target value calculated by the target function at this time is: obtaining the adjustment parameter at which the target value of the target function is minimum, optimizing the prediction model according to the adjustment parameter.

[0013] Optionally, the obtaining the adjustment parameter at which the target value of the target function is minimum, optimizing the prediction model according to the adjustment parameter, and obtaining the optimized prediction model according to the Taylor expansion expression include: obtaining a split evaluation value when the regression tree structure is split A , which is expressed as: wherein, and are sample sets of left and right child trees after splitting, , and the sample set before splitting is represented as: When the split evaluation value is greater than 0, the splitting of the leaf nodes of the tree is continued. When the split evaluation value is less than 0, the splitting of the leaf nodes of the tree is stopped, and the number of leaves is obtained. obtaining the optimal output value of the leaf node of the first tree according to the number of leaves and the Taylor expansion expression; obtaining the optimal output value of the leaf node of the first tree according to the number of leaves and the Taylor expansion expression; obtaining the adjustment parameter at which the target value of the target function is minimum according to the minimum target value calculated according to the optimal output value, optimizing the prediction model according to the adjustment parameter, and obtaining the optimized prediction model.

[0014] An engine system weight prediction system, comprising: an acquisition module configured to acquire engine design parameters and system weights corresponding to the design parameters; an association module configured to establish a mapping relationship between the engine design parameters and the system weights corresponding to the design parameters, and obtain a plurality of original data sets; a generative adversarial network optimization module configured to input the original data sets into a generative adversarial network, obtain generated data sets, and set a constraint condition for a generator in the generative adversarial network to enable the generator to evaluate the generated data sets and obtain an evaluation value, and a discriminator in the generative adversarial network to discriminate the generated data sets and obtain a discrimination value, and optimize the generative adversarial network according to the discrimination value and the evaluation value, and obtain an optimized generative adversarial network; a training data set generation module configured to obtain optimized generated data sets according to the optimized generative adversarial network and the original data sets, and combine the optimized generated data sets and the original data sets into a training data set; The neural network model optimization module is configured to input the training data set into a neural network model, extract hidden layer features, input the hidden layer features into an activation function to obtain output values, calculate a loss value according to the output values and true values, optimize model parameters of the neural network model by minimizing the loss value, obtain an optimized neural network model, and the true values are system weights in the training data set. The feature extraction module is configured to obtain latent features according to the optimized neural network model and the training data set. The prediction model optimization module is configured to input the latent features into a prediction model to obtain prediction values, construct an objective function according to the prediction values and true values, obtain adjustment parameters when a target value of the objective function is minimum, optimize the prediction model according to the adjustment parameters, and obtain an optimized prediction model. The prediction module is configured to input actual engine design parameters into the optimized prediction model to obtain an actual system weight.

[0015] A terminal device includes a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor loads and executes the computer program, and adopts an engine system weight prediction method.

[0016] A computer readable storage medium stores a computer program, and the computer program is loaded and executed by a processor, and an engine system weight prediction method is adopted.

[0017] The beneficial effects of the present application are: 1. By establishing a mapping relationship between engine design parameters and corresponding system weights, obtaining an original data set, then inputting it into a generative adversarial network for training, optimizing the evaluation value generated by the generator and the discrimination value generated by the discriminator, obtaining an optimized generative adversarial network, inputting the training data set into a neural network model to obtain a loss value, obtaining an optimized neural network model by minimizing the loss value, extracting latent features, inputting the latent features into a prediction model, and obtaining an optimal prediction model by minimizing the objective function, inputting actual engine design parameters into the optimized prediction model to obtain an actual system weight. The present application has the advantages of generating a large amount of similar data, i.e., a data set, under insufficient data conditions, training a prediction model after feature extraction of the generated data set, obtaining an optimized training prediction model, and improving prediction accuracy. Compared with traditional prediction model training, under the condition of insufficient engine design data, the prediction accuracy of engine weight prediction can be significantly improved by enhancing training samples and extracting latent features.

[0018] 2、Through DNN (neural network model) for capturing complex feature interaction, and XGBoost (prediction model) provides high efficiency and accuracy in regression prediction, improves the accuracy and stability of prediction. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is a whole structure schematic diagram of the engine system weight prediction method of the application; Figure 2 It is an evolution curve diagram of the loss function of the training process of the application; Figure 3 It is a comparison diagram of the true value and the predicted value of the application; Figure 4 It is a precision comparison diagram of using original data set and expanded data set in DNN model for prediction of the application; Figure 5 It is a precision comparison of using original data set and expanded data set in XGBoost model for prediction of the application; Figure 6 It is a precision comparison of using original data set and expanded data set in Int model for prediction of the application; Figure 7 It is a schematic diagram of the improvement rate of the prediction effect of DNN and XGBoost module of the application; Figure 8 It is a statistical comparison diagram of the prediction accuracy of each model of the application; Figure 9 It is a schematic diagram of the prediction accuracy comparison of three models on original data set of the application; Figure 10 It is a prediction accuracy comparison diagram of three models on expanded data set of the application; Figure 11 It is a prediction accuracy distribution diagram of different models and data sets under different threshold values of the application. DETAILED DESCRIPTION

[0020] An engine system weight prediction method, as shown in Figure 1 , comprising: S1, obtaining engine design parameters and system weight corresponding to the design parameters; Specifically, the engine design parameters are represented as: Among them, N represents the total number of engine samples, each represents the dimensional feature vector of the th engine. The feature space contains There are several independent design parameters that may affect the system's weight characteristics. Accordingly, the system weight is represented as an objective vector. ,in Indicates the first The system weight is obtained from the measurements of each engine. Given the diverse factors influencing aircraft engine weight, each input sample can be formally represented as: Each input sample can be formally represented as: in The first term used for system weight prediction Engine design parameters This indicates the total number of design parameters of this type.

[0021] S2. Establish a mapping relationship between the engine design parameters and the system weight corresponding to the design parameters to obtain the original data set (real sample). Specifically, based on the corresponding engine data The system weight is predicted using a learning function. This allows it to input the engine. Mapped to the corresponding system weight prediction value As shown in the following formula: in, Indicates the first The predicted weight of each engine is the predicted value.

[0022] Each And a corresponding real This is called a raw data set.

[0023] S3. Input the original data set into the generative adversarial network to obtain a generated data set. The generative adversarial network includes a generator and a discriminator. Set constraints on the generator to make the generator evaluate the generated data set and obtain an evaluation value. The discriminator discriminates the generated data set and obtains a discrimination value. Based on the discrimination value and the evaluation value, optimize the generative adversarial network to obtain an optimized generative adversarial network. Specifically, due to the less disclosed data of the weight parameter data of the engine, it is insufficient for model training, and therefore a conditional generative adversarial network (cGAN) model is used to enhance the engine data. The cGAN model is considered as a promising solution to approximate the distribution of real-world data and generate synthetic samples that retain the intrinsic characteristics of the original data, thereby expanding the training set and alleviating the data scarcity problem. Wherein the random noise and the condition information are jointly input into the generator G as input, and the generator generates synthetic samples and sends them to the discriminator D for evaluation. Unlike traditional GAN, the discriminator of cGAN not only needs to distinguish between true and false samples, but also needs to judge whether the generated samples meet the predetermined conditions.

[0024] The original data set is input into the generative adversarial network to obtain a generated data set, the generative adversarial network includes a generator and a discriminator, the generator evaluates the generated data set to obtain an evaluation value, the discriminator discriminates the generated data to obtain a discrimination value, and the generative adversarial network is optimized according to the discrimination value and the evaluation value to obtain an optimized generative adversarial network, including: The original data set is input into the generative adversarial network to obtain a generated data set; The generator evaluates the generated data set to obtain an evaluation value, which is expressed as:

[0025] Among them, is the generator, is the condition, indicates that under the condition , the generator generates the generated data set , is the prior distribution of the generated data set, indicates is sampled from the generated data set, is the discriminator, indicates that for the generated data set generated by the generator , the probability that the given condition is that the sample generated data set comes from the original data set, and the value range is , indicates that under random sampling, the sample generated according to is input into the discriminator , and the probability given by the discriminator is true and the expected value of the logarithm is taken;

[0026] The discriminator discriminates the generated data set to obtain a discrimination value, which is expressed as: wherein, denotes is sampled from the original data set, is the output of the discriminator from the original data set and is taken logarithm; is the discriminator for the probability that the output from the original data set is true and is taken logarithm.

[0027] According to the discriminant value and the evaluation value, the generative adversarial network is optimized, and an optimized generative adversarial network is obtained, denoted as: wherein, denotes a real sample obtained by sampling from the original data set distribution .

[0028] S4, according to the optimized generative adversarial network and the original data set, an optimized generated data set is obtained, and the optimized generated data set and the original data set are used as a training data set; S5, inputting the training data set into a neural network model, extracting hidden layer features, inputting the hidden layer features into an activation function to obtain an output value, calculating a loss value according to the output value and a true value, optimizing model parameters of the neural network model by minimizing the loss value, and obtaining an optimized neural network model, wherein the true value is a system weight in the training data set. Specifically, in order to make full use of the information contained in the input data, we use a deep neural network (DNN) model to extract features from the engine data. DNN has structural advantages by introducing multiple nonlinear transformations. By constructing a complex network architecture containing an input layer, multiple hidden layers and an output layer, DNN can deeply mine features from data layer by layer. This multi-layer structure enables DNN to learn various feature representations from simple to complex from raw data.

[0029] In addition, DNN has strong adaptive learning ability. It is trained on large-scale data sets through the backpropagation algorithm, which can continuously adjust the parameters in the network to minimize the error between the predicted output and the actual label. This process enables DNN to sensitively capture subtle potential features in the data while well fitting its underlying distribution characteristics. Even in the face of complex, high-dimensional and noisy data, DNN still has significant effect in extracting valuable potential features.

[0030] inputting the training data set into a neural network model, extracting hidden layer features, obtaining output values by passing the hidden layer features through an activation function, calculating a loss value according to the output values and true values, optimizing model parameters of the neural network model by minimizing the loss value to obtain an optimized neural network model, and the method comprises the following steps: inputting the training data set into a neural network model, extracting hidden layer features , and are represented as

[0031] wherein, is an activation function, is the number of hidden layers, is a weight matrix of an i-th hidden layer, is a bias vector of the i-th hidden layer, is a training data set; Specifically, the constructed DNN model comprises hidden layers. passing the hidden layer features through an activation function to obtain output values; Specifically, in the hidden layer, ScaledExponentialLinearUnit (SELU) is used as the activation function, and the definition is as follows:

[0032] wherein, and are constants, and the default values are used. The above is the forward propagation process of the DNN model. In order to find suitable weights and biases , the DNN uses a backpropagation algorithm to minimize the error between the predicted output and the true label, and defines a loss function. According to the output values and the true values, a loss value is calculated and represented as

[0033] wherein, is the loss value, is the weight matrix, is the bias vector, is the input sample, is the true label, is the output value, represents the Euclidean norm; optimizing the model parameters of the neural network model by minimizing the loss value to obtain an optimized neural network model.

[0034] ​​Specifically, gradient descent is used to iteratively solve for the weights and biases of each layer, improving network performance and enabling it to learn. The trained DNN model can then effectively extract high-dimensional latent features from the original data.

[0035] S6. Based on the optimized neural network model and the training data set, obtain the potential features; Based on the optimized neural network model and the training data set, the latent features are represented as follows: in, For the first The weight matrix of the hidden layer, For the first Hidden layer output of the layer, and It is a constant. For the first The bias vector of the hidden layer. These are potential features.

[0036] S7. Input the latent features into the prediction model to obtain the predicted value. Based on the predicted value and the true value, construct the objective function, obtain the adjustment parameter when the objective value of the objective function is minimized, and optimize the prediction model based on the adjustment parameter to obtain the optimized prediction model. Specifically, the prediction model in this application employs XGBoost. XGBoost is an efficient and highly scalable gradient boosting framework that uses classification and regression trees as its base learners. These base learners are combined sequentially to form a strong learner, significantly improving the model's prediction accuracy and generalization ability.

[0037] The process of inputting the latent features into the prediction model to obtain predicted values, constructing an objective function based on the predicted values ​​and the true values, obtaining adjustment parameters when the objective function's target value is minimized, and optimizing the prediction model based on the adjustment parameters to obtain the optimized prediction model includes: The latent features are input into the prediction model to obtain the predicted value, which is expressed as: in, For predicted values, Indicates the number of regression trees. For the function space of the regression tree, For the first Tree samples The predicted value; Based on the predicted and actual values, an objective function is constructed, expressed as: in, This represents the error between the predicted and actual values. This represents the total number of engine samples. for The sum of the regularization terms of the tree can be expanded as follows: in, For the first Round iteration, i.e. the first iteration Optimization of trees and To adjust the parameters, The number of leaves, For the first Output value of a leaf; The adjustment parameters are obtained when the objective function is minimized. The prediction model is then optimized based on the adjustment parameters to obtain the optimized prediction model.

[0038] The step of obtaining the adjustment parameter when the objective function is minimized, and optimizing the prediction model based on the adjustment parameter, includes: No. The objective function of the tree is: in, For the first The objective function value of each tree. For the first The predicted values ​​output by each tree. For the true value, For the first Tree pairs of samples The predicted value, For constant terms, This represents the total number of engine samples. For the first The regularization term of a tree, For the first The prediction error value for each tree.

[0039] The first The objective function expression of the tree is expanded using Taylor, resulting in the Taylor expansion expression, which is then simplified as follows: in, and Let these represent the first and second derivatives of the loss function, respectively. For belonging to the first The set of all samples of a leaf. For the first A leaf, For the first A sample, the partial derivative of the Taylor expansion expression is solved and set to 0, and the optimal output value is obtained : At this time, the minimum target value of the target function is calculated : Obtain the adjustment parameter when the target value of the target function is minimum, and optimize the prediction model according to the adjustment parameter.

[0040] Specifically, in order to further optimize the performance, XGBoost introduces a second-order Taylor expansion in the loss function to accelerate the convergence process. At the same time, in order to reduce the risk of overfitting, the complexity of the tree model is explicitly added to the regularization term. This processing improves the robustness of the model, and it is not easy to overfit even on complex data sets.

[0041] According to the Taylor expansion expression, the adjustment parameter is obtained when the target value of the target function is minimum, and the prediction model is optimized according to the adjustment parameter to obtain an optimized prediction model.

[0042] According to the Taylor expansion expression, the adjustment parameter is obtained when the target value of the target function is minimum, and the prediction model is optimized according to the adjustment parameter to obtain an optimized prediction model, including: Obtain the split evaluation value when the regression tree structure is split A , expressed as: Wherein, and are the sample sets of the left and right child trees after splitting, , and the sample set before splitting is represented by When the split evaluation value is greater than 0, continue to split the leaf nodes of the tree; When the split evaluation value is less than 0, stop splitting the leaf nodes, and obtain the number of leaves; According to the number of leaves and the Taylor expansion expression, the optimal output value of the leaf node of the first tree is obtained; According to the optimal output value, the minimum target function value is calculated, the adjustment parameter when the target value of the target function is minimum is obtained, and the prediction model is optimized according to the adjustment parameter to obtain an optimized prediction model.

[0043] According to the number of leaves and the Taylor expansion expression, the output value of the leaf node of the first tree is obtained , so as to obtain a complete XGBoost prediction model.

[0044] S8, input actual engine design parameters into the prediction model to obtain the actual system weight. Specific embodiments Dataset The dataset used in this application contains the design parameters of 144 commercial engines and 39 turbofan engines studied in various aviation projects. These commercial engines span from the mid-1960s to the late 2010s, covering more than half a century of technological progress and engineering experience, providing a realistic basis for predictive analysis.

[0046] This application focuses on 12 engine parameters, as shown in Table 1. In this dataset, turbofan engines can be further divided into three types according to system type and fan number, as shown in Table 2. These are also the condition labels when generating data using cGAN.

[0047] Table 1 Engine dataset parameters Table 2 Three types of turbofan systems The input features used for the prediction task correspond to the engine design parameters numbered 3 to 10 in Table 1. These symbols include: BPR(SLS), OPR(SLS), Thrust, IBS(SLS), CruiseMach, CruiseAlt.kft.YearCertified, SystemType, No.ofSpools. The target variable (item 12) to be predicted is the propulsion system weight (PropulsionSystemWeight).

[0048] Experimental setup

[0049] The dataset is divided into training and test sets in an 8:2 ratio. The algorithm used is based on the TensorFlow framework. All experiments are run on a high-performance server configured with dual AMD EPYC 7402 24-core 2.8 GHz processors, 256 GB of memory, and 8 GeForce RTX 3090 GPUs.

[0050] Evaluation metrics: We use three metrics to evaluate model performance: mean absolute error (MAE), coefficient of determination (R2), and accuracy (Acc).

[0051] The three are defined as follows: MAE (Mean Absolute Error) is defined as:

[0052] wherein, is the true system weight of the th engine, is the predicted value (i.e. the predicted system weight), is the total number of engine samples.

[0053] The system is defined as: wherein is the average of the true values.

[0054] The accuracy Acc is defined as: wherein is the error of the th predicted value and the reference value.

[0055] Experimental results and analysis The data augmentation and prediction tasks are sequentially performed by the Int model (corresponding to the feature extraction and prediction parts). In the data augmentation process based on cGAN, we use grid search (GridSearch) to adjust the model parameters, and compare the convergence of the loss learning curves under multiple settings. Finally, we determine the optimal hyperparameters of the two subnetworks in cGAN, and the detailed information is shown in Table 3.

[0056] Table 3 Parameter settings of the generator and discriminator in the cGAN model

[0057] The ideal situation is that the discriminator can accurately distinguish all real samples (loss is 1) and fake samples (loss is 0), and the final discriminator loss is about 0.5. The relevant trend is shown in Figure 2 .

[0058] Subsequently, the original real samples and the synthesized samples are merged and input into the DNN model for feature extraction. The extracted features are then input into the XGBoost model for prediction. Hyperparameter grid search is also performed for the DNN and XGBoost models, and the EarlyStopping strategy is introduced to prevent overfitting and excessive training time. The relevant hyperparameter settings are shown in Table 4.

[0059] Table 4 Parameter settings of the DNN and XGBoost modules

[0060] Table 5 Prediction results of each data point Table 5 shows the comparison of predicted values with the test set. Out of the 37 evaluated engines, 29 had a prediction accuracy of over 99%, and the overall accuracy was over 90%. However, there were individual cases of inaccurate predictions, such as an engine with an actual weight of 5598 that had a prediction accuracy of only 92.06%. These outliers will be further investigated in future work.

[0061] Figure 3 The direct comparison between the actual values ( ) and the predicted values ( ) for the 37 analyzed engines is shown. The graph clearly shows the high consistency of the predicted values with the true data points, with a perfect prediction line as a reference. Although there are some small deviations, most data points are close to this line, indicating that the predictions in the entire dataset have a high accuracy.

[0062] Ablation experiments The core innovation of the framework is the application of data augmentation techniques and the combination of deep neural networks (DNN) and XGBoost models. To verify the effectiveness of these operations, ablation experiments are needed. Two ablation models were designed for the study. The first experiment aims to verify the improvement of data augmentation on weight prediction performance, and the second experiment evaluates the effectiveness of the Int model.

[0063] Experiment (1) Two datasets were used: the original engine dataset (Raw_data) and the augmented engine dataset (Aug_data), as well as three models: deep neural network (DNN) model, XGBoost model, and ensemble model. First, we input Raw_data into these three models to obtain baseline prediction results. Then, we input Aug_data into these models under the same settings. Finally, the prediction results of the two datasets are compared. The results show that for each model, the use of augmented data significantly improves the prediction accuracy. Figure 4 The comparison of prediction accuracy of three models (DNN, XGBoost, and ensemble model) on the original dataset (Raw_data) and augmented dataset (Aug_data) is shown. Each subplot shows the accuracy value of a single sample under two data settings, and the dashed line connects the corresponding data points to highlight the effect of augmentation. As shown in Figure 5 and Figure 6 , the prediction accuracy of DNN and XGBoost models has significantly improved when trained with augmented data, especially on samples with initially poor performance, Figure 5 and Figure 6In this case, one x-coordinate corresponds to one engine sample, and there are 37 engine samples. This improvement is particularly evident in the case of XGBoost, indicating that it is more sensitive to data richness. Figure 6 Further emphasis is placed on the effectiveness of the ensemble model, which consistently outperforms individual models on most samples. When trained on the Aug_data dataset, the Int model not only has higher accuracy but also greater stability, with almost all predicted results exceeding 0.96. These results collectively demonstrate the effectiveness of the data augmentation method.

[0064] In addition, the improvement in prediction accuracy of the Aug_data compared to the Raw_data is quantified in the DNN model and the XGBoost model. The results, as shown in Figure 7 the bar chart intuitively presents the positive change in prediction accuracy for the 37 engine samples when switching from Raw_data to Aug_data using DNN and XGBoost models. From bottom to top, the 37 groups of bar charts sequentially display the accuracy improvement of samples numbered 1-37, where the blue bar chart represents the accuracy improvement achieved by the DNN model using Aug_data, the orange bar chart represents the accuracy improvement of the XGBoost model using Aug_data, and the gray bar represents no improvement in accuracy. It can be seen that the prediction accuracy of most samples has improved, with some samples showing an improvement of more than 10%. Notably, in some cases, the XGBoost model shows more significant improvement (e.g., samples 7, 12, and 23), while the DNN model shows consistent but moderate improvement on most data points. These results further demonstrate the effectiveness of data augmentation, especially in improving the performance of models for samples with initially low prediction accuracy.

[0065] The statistical properties of the prediction results of the DNN and XGBoost models are further analyzed. As shown in Figure 8 the box plot compares the prediction accuracy distribution of the three models (DNN, XGBoost, and ensemble model) on the original dataset and the augmented dataset. For each model, the accuracy has significantly improved in both central tendency and variability after using data augmentation. Specifically, after using augmented data, the median accuracy of all three models has increased, and the interquartile range has become more compact, which not only improves prediction performance but also enhances stability. The ensemble model (Int) outperforms individual models, especially when using augmented data, with its accuracy distribution concentrated above 0.99 and minimal variance.

[0066] Experiment (2) To validate the effectiveness of the ensemble model component in the proposed framework, we compare the prediction performance of three models - using and not using the ensemble model - on two datasets. When not using the ensemble model, predictions are made using the DNN and XGBoost models, respectively. Figure 9 and Figure 10 The performance of these three models on the original and augmented datasets, respectively, is shown. These comparisons further demonstrate the contribution of the ensemble model to overall prediction accuracy. The ensemble model consistently exhibits higher prediction accuracy than the DNN and XGBoost models alone on most data points. In particular, when using the augmented dataset (Aug_data), the ensemble model not only is overall higher, but its accuracy values are more tightly clustered near the upper limit (close to 1.0), showing better stability. In contrast, the individual models exhibit greater variability in prediction accuracy, especially on the original dataset, with several apparent outliers. These observations suggest that the integration of DNN and XGBoost effectively leverages their complementary strengths, resulting in improved overall performance and reduced sensitivity to data variability.

[0067] Subsequently, we analyze the distribution of prediction accuracy within predefined threshold intervals before and after introducing the ensemble model component. The results are presented in the form of heatmaps in Figure 11 , which show the number and percentage of predictions exceeding each threshold for comparison. The five threshold intervals considered are: >0.80, >0.85, >0.90, >0.95, and >0.99, evaluated on both the original (Raw) and augmented (cGAN) datasets for all models. Each cell in the heatmap displays the number of samples exceeding the corresponding threshold and their proportion. From Figure 9 , it is evident that the ensemble model outperforms the individual models regardless of whether the original or augmented data is used. In particular, under the augmented data setting, the proportion of samples with prediction accuracy exceeding 0.99 reaches 78%, which not only improves precision but also ensures more stable and reliable predictions across the entire dataset. This result demonstrates that the proposed framework not only enhances prediction accuracy but also strengthens model reliability, providing a more reliable foundation for practical applications requiring high-confidence predictions.

[0068] Finally, we used three evaluation metrics: mean absolute error (MAE), coefficient of determination (R²), and accuracy (ACC) to compare the prediction performance of the three models on both datasets. The results are summarized in Table 6. The ensemble model trained on augmented data, i.e., IA, consistently demonstrated the best overall performance. It achieved the lowest mean absolute error of 110.656, the highest coefficient of determination of 0.998, and a narrow range of accuracy between 0.921 and 1.000, with an average accuracy of 0.991. Compared to the models trained on the original dataset, i.e., DR, XR, and IR, all the models that benefited from augmented data, DA, XA, and IA, showed significant improvements. This demonstrates the advantage of data augmentation in improving prediction accuracy and generalization capability. Although the XGBoost model performed well in terms of coefficient of determination, reaching 0.978 in both XR and XA settings, it showed greater variability at the sample level. In particular, the lowest accuracy of XR dropped to 0.591, indicating a decrease in stability in some cases. In contrast, the ensemble models, IR and IA, not only achieved higher average accuracy but also demonstrated more consistent predictions across the entire dataset. These findings confirm that combining model ensembling with data augmentation provides significant benefits in terms of prediction precision and robustness. This approach offers a promising direction for improving the reliability of engine system weight estimation in spacecraft design.

[0069] Table 6. Comparison of prediction results of two datasets and three models. DR / XR / IR represent deep neural network (DNN), XGBoost, and ensemble model using the original dataset, respectively; DA / XA / IA denote the corresponding models trained on augmented datasets.

[0070] Table 6. Comparison of prediction results of two datasets and three models. DR / XR / IR represent deep neural network (DNN), XGBoost, and ensemble model using the original dataset, respectively; DA / XA / IA denote the corresponding models trained on augmented datasets.

[0071] The application proposes an engine system weight prediction method for predicting the weight of an aircraft engine system, aiming to solve the dual challenges of data scarcity and model complexity in the early design stage. By combining cGAN model to generate synthetic data and integrated DNN-XGBoost regression model, this framework significantly improves the accuracy and stability of the prediction. Experimental results show that this framework not only reduces the prediction error, but also improves the consistency between samples. Specifically, the data enhancement module significantly improves the accuracy of all three baseline models, with the improvement of multiple models exceeding 10%. The integrated DNN-XGBoost model further enhances the stability of the prediction, with more than 90% of the predictions reaching at least 95% accuracy, and 78% of the predictions reaching 99% accuracy. Therefore, this method has practical application value in the field of aerospace engineering, especially in the preliminary design stage, when empirical data is scarce but accurate estimation is crucial. It provides engineers with an effective tool to generate high-confidence weight predictions under uncertainty, supporting concept design research, early risk assessment, and resource planning in aircraft engine system development. Although this method has proven effective, there are still some limitations. The generated data, although diverse, may not fully capture the subtle physical limitations inherent in engine architecture, and the current model does not explicitly incorporate domain knowledge or physical priors. Although the overall prediction accuracy is high, there are still slight performance differences between individual samples, indicating that the model's sensitivity to design-specific changes can be improved. Future research will focus on improving the physical authenticity of augmented data by introducing physics-based constraints and enhancing the model's sensitivity to design-specific changes to reduce performance differences between samples. These efforts may also include exploring adaptive feature selection or attention mechanisms to improve the accuracy of local predictions while maintaining computational efficiency.

[0072] An engine system weight prediction system comprises: An acquisition module is configured to acquire engine design parameters and system weights corresponding to the design parameters. An association module is configured to establish a mapping relationship between the engine design parameters and the system weights corresponding to the design parameters to obtain a plurality of original data sets. A generative adversarial network optimization module is configured to input the original data sets into a generative adversarial network to obtain generated data sets, wherein the generative adversarial network comprises a generator and a discriminator, the generator is set with a constraint condition to evaluate the generated data to obtain an evaluation value, the discriminator discriminates the generated data sets to obtain a discrimination value, and the generative adversarial network is optimized according to the discrimination value and the evaluation value to obtain an optimized generative adversarial network. The training data set generation module is configured to generate an optimized generated data set according to the optimization generative adversarial network and the original data set, and combine the optimized generated data set and the original data set into a training data set. The neural network model optimization module is configured to input the training data set into a neural network model, extract hidden layer features, input the hidden layer features into an activation function to obtain an output value, calculate a loss value according to the output value and a true value, optimize model parameters of the neural network model by minimizing the loss value, obtain an optimized neural network model, and the true value is a system weight in the training data set. The feature extraction module is configured to obtain latent features according to the optimized neural network model and the training data set. The prediction model optimization module is configured to input the latent features into a prediction model to obtain a prediction value, construct an objective function according to the prediction value and the true value, obtain adjustment parameters when a target value of the objective function is minimum, optimize the prediction model according to the adjustment parameters, and obtain an optimized prediction model. The prediction module is configured to input actual engine design parameters into the optimized prediction model to obtain an actual system weight.

[0073] The application also discloses a terminal device, which comprises a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor loads and executes the computer program.

[0074] The terminal device can be a computer device such as a desktop computer, a notebook computer or a cloud server, and the terminal device comprises but is not limited to a processor and a memory, for example, the terminal device can also comprise an input / output device, a network access device and a bus.

[0075] The processor can be a central processing unit (CPU), and of course, according to actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), ready-to-program gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. can also be used, and the general-purpose processor can be a microprocessor or any conventional processor, etc. The application does not limit this.

[0076] The memory can be an internal storage unit of the terminal device, for example, a hard disk or a memory of the terminal device, or an external storage device of the terminal device, for example, a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD), or a flash card (FC) equipped on the terminal device, or a combination of the internal storage unit and the external storage device of the terminal device. The memory is used to store a computer program and other programs and data required by the terminal device, and can also be used to temporarily store data that has been output or will be output. The present application does not limit this.

[0077] The engine system weight prediction method in the above embodiments is stored in the memory of the terminal device, and is loaded and executed on the processor of the terminal device, so as to facilitate use.

[0078] The present application also discloses a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is executed by a processor, the engine system weight prediction method in the above embodiments is used.

[0079] The computer program can be stored in a computer readable medium, and the computer program includes computer program code. The computer program code can be in the form of source code, object code, an executable file, or some intermediate form of the above. The computer readable medium includes any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the computer readable medium includes but is not limited to the above-mentioned components.

[0080] The engine system weight prediction method in the above embodiments is stored in the computer readable storage medium, and is loaded and executed on the processor, so as to facilitate storage and application of the method.

[0081] Those skilled in the art should understand that the above discussion of any embodiment is only exemplary, and is not intended to limit the protection scope of the present application to these examples; under the idea of the present application, the above embodiments or technical features in different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of one or more embodiments of the present application as described above. In order to be brief, they are not provided in details.

[0082] It is intended that the embodiments of the application herein disclosed meet all the written requirements of the patent statutes and come within the judicial doctrines of equivalents and will not be construed to be limited to the embodiments shown and described and by the keeping within the spirit and scope of the embodiments of the application.

Claims

1. A method for predicting the weight of an engine system, characterized in that, include: Obtain the engine design parameters and the corresponding system weight; A mapping relationship is established between the engine design parameters and the corresponding system weights to obtain the original data set. The original data set is input into a generative adversarial network (GAN) to obtain a generated data set. The GAN includes a generator and a discriminator. Constraints are set on the generator to enable it to evaluate the generated data set and obtain an evaluation value. The discriminator discriminates the generated data set and obtains a discrimination value. Based on the discrimination value and the evaluation value, the GAN is optimized to obtain an optimized GAN. Based on the optimized generative adversarial network and the original data set, an optimized generated data set is obtained, and the optimized generated data set and the original data set are combined into a training data set. The training data set is input into the neural network model, hidden layer features are extracted, the hidden layer features are input into the activation function to obtain the output value, the loss value is calculated based on the output value and the true value, and the model parameters of the neural network model are optimized by minimizing the loss value to obtain the optimized neural network model. The true value is the system weight in the training data set. Based on the optimized neural network model and the training data set, latent features are obtained; The latent features are input into the prediction model to obtain the predicted values. Based on the predicted values ​​and the true values, an objective function is constructed. The adjustment parameters are obtained when the objective function's target value is minimized. The prediction model is then optimized based on the adjustment parameters to obtain the optimized prediction model. The actual engine design parameters are input into the optimization prediction model to obtain the actual system weight.

2. The engine system weight prediction method as described in claim 1, characterized in that, The process involves inputting the original data set into a generative adversarial network (GAN) to obtain a generated data set. The GAN includes a generator and a discriminator. Constraints are set on the generator to evaluate the generated data set and obtain an evaluation value. The discriminator then discriminates the generated data set to obtain a discrimination value. Based on the discrimination value and the evaluation value, the GAN is optimized to obtain an optimized GAN, comprising: The original data set is input into the generative adversarial network to obtain the generated data set; The generator evaluates the generated data set and obtains an evaluation value, which is represented as: in, For generator, As a condition, Indicates the condition Below, generator The generated data set , To generate the prior distribution of the data set, express It is sampled from the generated data set. It is a discriminator. This indicates that for generators The generated data set Under given conditions The probability that the generated data set is derived from the original data set is given, and the range of values ​​is given. , This indicates that under random sampling, according to generated samples Input discriminator In the middle, the discriminator gives the probability of being true and takes the expected value of the logarithm; The discriminator performs a discrimination operation on the generated data group to obtain a discrimination value, which is represented as follows: in, express It was sampled from the original dataset. For the discriminator pair The output of the probabilities from the original data set is then logarithmic. For discriminator right The expected value of the logarithm of the probability of being true given from the original data set; Based on the discriminant value and the evaluation value, the generative adversarial network is optimized to obtain an optimized generative adversarial network, represented as: in, Indicates the distribution from the original data set The real samples obtained from sampling.

3. The engine system weight prediction method as described in claim 1, characterized in that, The steps of inputting the training data set into the neural network model, extracting hidden layer features, passing the hidden layer features through an activation function to obtain output values, calculating a loss value based on the output value and the true value, and optimizing the model parameters of the neural network model by minimizing the loss value to obtain an optimized neural network model include: The training data set is input into the neural network model to extract hidden layer features. , is represented as: in, For activation function, To hide the number of layers, For the first The weight matrix of the hidden layer, For the first The bias vector of the hidden layer. For the training data set; The hidden layer features are passed through an activation function to obtain the output value; Based on the output value and the true value, the loss value is calculated and expressed as: in, This is the loss value. This is the weight matrix. For bias vectors, For the input sample, For real labels, For the output value, Denotes the Euclidean norm; An optimized neural network model is obtained by optimizing the model parameters of the neural network model by minimizing the loss value.

4. The engine system weight prediction method as described in claim 1, characterized in that, Based on the optimized neural network model and the training data set, the latent features are represented as follows: in, For the first The weight matrix of the hidden layer, For the first Hidden layer output of the layer, and It is a constant. For the first The bias vector of the hidden layer. These are potential features.

5. The engine system weight prediction method as described in claim 1, characterized in that, The process of inputting the latent features into the prediction model to obtain predicted values, constructing an objective function based on the predicted values ​​and the true values, obtaining adjustment parameters when the objective function's target value is minimized, and optimizing the prediction model based on the adjustment parameters to obtain the optimized prediction model includes: The latent features are input into the prediction model to obtain the predicted value, which is expressed as: in, For predicted values, Indicates the number of regression trees. For the function space of the regression tree, For the first Tree samples The predicted value; Based on the predicted and actual values, an objective function is constructed, expressed as: in, This represents the error between the predicted and actual values. This represents the total number of engine samples. for The sum of the regularization terms of the tree can be expanded as follows: in, For the first Round iteration, i.e. the first iteration Optimization of trees and To adjust the parameters, The number of leaves, For the first Output value of a leaf; The adjustment parameters are obtained when the objective function is minimized. The prediction model is then optimized based on the adjustment parameters to obtain the optimized prediction model.

6. The engine system weight prediction method as described in claim 5, characterized in that, The step of obtaining the adjustment parameter when the objective function is minimized, and optimizing the prediction model based on the adjustment parameter, includes: No. The objective function of the tree is expressed as: in, For the first The objective function value of each tree. For the first The predicted values ​​output by each tree. For the true value, For the first Tree pairs of samples The predicted value, For constant terms, This represents the total number of engine samples. For the first The regularization term of a tree, For the first The prediction error value for each tree; The first The objective function expression of the tree is expanded using Taylor, resulting in the Taylor expansion expression, which is then simplified as follows: in, and Let these represent the first and second derivatives of the loss function, respectively. For belonging to the first The set of all samples of a leaf. For the first A leaf, For the first Taking the partial derivative of the Taylor expansion with respect to a given sample and setting it to zero, the optimal output value is obtained. : At this point, the minimum objective value calculated using the objective function is... for: Obtain the adjustment parameters when the objective value of the objective function is minimized, and optimize the prediction model based on the adjustment parameters.

7. The engine system weight prediction method as described in claim 6, characterized in that, The step of obtaining the adjustment parameter when the objective function is minimized based on the Taylor expansion expression, and optimizing the prediction model based on the adjustment parameter to obtain the optimized prediction model includes: Obtaining the segmentation evaluation value when the regression tree structure splits. A , is represented as: in, and These are the sample sets of the left and right subtrees after the split, respectively. This represents the sample set before the split; If the segmentation evaluation value is greater than 0, then continue to split the leaf nodes of the tree; When the segmentation evaluation value is less than 0, the splitting of the logarithmic leaf nodes is stopped, and the number of leaves is obtained; Based on the number of leaves and the Taylor expansion expression, we obtain the first... The optimal output value of the leaf nodes of the tree; The minimum objective function value is calculated based on the optimal output value. The adjustment parameter is obtained when the objective function has the minimum objective value. The prediction model is then optimized based on the adjustment parameter to obtain the optimized prediction model.

8. An engine system weight prediction system, characterized in that, include: The acquisition module is used to acquire engine design parameters and the corresponding system weight. The association module is used to establish a mapping relationship between the engine design parameters and the system weights corresponding to the design parameters, and to obtain several sets of raw data. A generative adversarial network (GAN) optimization module is used to input the original data set into a GAN to obtain a generated data set. The GAN includes a generator and a discriminator. Constraints are set on the generator to enable it to evaluate the generated data set and obtain an evaluation value. The discriminator discriminates the generated data set to obtain a discrimination value. Based on the discrimination value and the evaluation value, the GAN is optimized to obtain an optimized GAN. The training data set generation module is used to obtain an optimized generated data set based on the optimized generative adversarial network and the original data set, and to combine the optimized generated data set and the original data set into a training data set. The neural network model optimization module is used to input the training data set into the neural network model, extract hidden layer features, input the hidden layer features into the activation function to obtain the output value, calculate the loss value based on the output value and the true value, optimize the model parameters of the neural network model by minimizing the loss value, and obtain an optimized neural network model, wherein the true value is the system weight in the training data set; The feature extraction module is used to obtain potential features based on the optimized neural network model and the training data set; The prediction model optimization module is used to input the latent features into the prediction model to obtain the predicted value, construct an objective function based on the predicted value and the true value, obtain the adjustment parameter when the objective function's target value is minimized, and optimize the prediction model based on the adjustment parameter to obtain the optimized prediction model. The prediction module is used to input the actual engine design parameters into the optimization prediction model to obtain the actual system weight.

9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.

Images