Panel aeroelasticity nonlinear response prediction method based on data driving

Abstract

The invention provides a wallboard aeroelasticity nonlinear response prediction method based on data driving, and relates to the technical field of fluid-solid coupling numerical calculation. And high-dimensional correction of the cross characteristics and the unsteady aerodynamic load is realized by adopting a residual network. And the prediction precision of the model is improved by combining the input characteristics of the neural network model and an adaptive model switching mechanism. Compared with a traditional numerical simulation method, the calculation efficiency of the wallboard aeroelasticity nonlinear response prediction method based on data driving is obviously improved, and meanwhile the model has good prediction precision and generalization performance.

Description

Technical Field

[0001] The present application relates to the technical field of fluid-solid coupling numerical calculation, and in particular to a data-driven method for predicting the aeroelastic nonlinear response of a wall panel. Background Art

[0002] The aeroelastic response of panels is a typical fluid-structure interaction problem. When one or both sides of a panel are exposed to supersonic flow, it will experience dynamic vibrations such as flutter. This leads to nonlinear aeroelastic response, causing structural fatigue, noise generation, and impacting flight safety. With the pursuit of lightweight and high-speed aircraft, the nonlinear aeroelastic response of elastic panels has become a major challenge in aircraft design and optimization. Fluid-structure interaction numerical solutions are the mainstream approach to solving the nonlinear aeroelastic response of panels.

[0003] The numerical calculation method of fluid-structure coupling mainly consists of three parts: fluid dynamics solution, structural dynamics solution and fluid-structure coupling equation solution. The fluid solver is used to solve the fluid dynamics equation and calculate the flow field information at the current moment based on the boundary conditions formed by the current structure. The structural solver is used to solve the structural dynamics equation and calculate the structural motion information based on the current structural surface pressure load. The fluid-structure coupling solver is used to solve the equilibrium equations of the fluid and solid interface and realize data exchange between the fluid solver and the structural solver. The loose coupling solution method is usually used for numerical calculation of fluid-structure coupling. This method simplifies the problem solving process by exchanging data only once between the fluid solver and the structural solver in each time step, and has been widely used in the field of fluid-structure coupling solution. However, due to the huge number of aerodynamic grids, the fluid solver needs to consume a lot of computing resources and time, resulting in low solution efficiency of the loose coupling solution method. Summary of the Invention

[0004] This application aims to address the limitations of existing numerical methods for calculating the nonlinear aeroelastic response of siding panels by providing a data-driven method for predicting the nonlinear aeroelastic response of siding panels. This method first establishes a neural network model based on a data-driven approach, then combines it with a structural solver to predict the nonlinear aeroelastic response of siding panels. This method effectively improves computational efficiency and exhibits good robustness and generalization performance.

[0005] To achieve the above objectives, the technical solutions adopted in this application are as follows: The present application provides a data-driven method for predicting the aeroelastic nonlinear response of a panel, comprising the following steps: S101: Constructing a neural network model for predicting the aeroelastic nonlinear response of the panel, wherein the neural network model includes input features, a cross network, a residual network, a fully connected network, and a predicted value. The data transmission process of the neural network model is as follows: input features are input into the cross network and the residual network in parallel, and then combined through the fully connected network to obtain a predicted value; S102: Based on the numerical coupling solution method, various structural responses of the panel's aeroelastic nonlinear response and the corresponding unsteady aerodynamic loads are obtained, and multiple databases are constructed. Based on the adaptive model switching mechanism, different databases are selected to train the neural network model into different sub-neural network models. S103: selecting at least one structural response and at least one unsteady aerodynamic load as input features and target values, respectively, establishing a sample data set containing the input features and target values ​​according to each database, preprocessing the sample data set, setting model hyperparameters, and training the neural network model into a sub-neural network model according to the sample data set and the model hyperparameters; S104: Predict the aeroelastic nonlinear response of the panel based on multiple sub-neural network models.

[0006] Furthermore, in S101, the cross network is used to quickly generate explicit cross features and construct a low-dimensional mapping between the cross features and the unsteady aerodynamic loads. The cross network includes multiple cross network layers and multiple first fully connected layers. i The expression of the layer-cross network layer is:

[0007] in, x in Is the input feature, which is a vector, ⊙ represents the vector dot product, x c,i and y c,i It is i Layer crosses the input and output of the network layer, W c,i and b c,i It is i The weight parameters and bias parameters of the cross network layer. i =1, x c,1 =x in .

[0008] No. i The expression of the first fully connected layer is:

[0009] in, xfc,i and y fc,i It is i The input and output of the first fully connected layer, W fc,i and b fc,i It is i The weight parameters and bias parameters of the first fully connected layer; The expression of the cross network is:

[0010] in, is the output value of the crossover network.

[0011] Furthermore, in S101, the residual network is used to construct a high-dimensional correction between cross features and unsteady aerodynamic loads. The residual network includes a feature cross layer, a convolutional layer, a residual layer, a pooling layer, and a second fully connected layer. The expression of the feature cross layer is:

[0012] in, is the input feature with the added feature 1, ⊗ represents the matrix multiplication, y inter is the two-dimensional output matrix of the feature cross layer; No. i The expression of the residual layer is:

[0013] in, W i,1 , W i,2 and W i,3 They are i Different weight parameters of the residual layer, x R,i and y R,i They are i The input and output of the residual layer, f(∙) is the activation function ReLU , BN(∙) is batch normalization; The expression of the residual network is:

[0014] in, is the expression of the convolutional layer, is the expression of the pooling layer. The convolutional layer and the pooling layer use common techniques of neural networks, and their expressions are not repeated here. The expression of the second fully connected layer is the same as that of the first connected layer. is the output value of the residual network.

[0015] Furthermore, in S101, the fully connected network combines the cross network and the residual network to generate the predicted value of the neural network model. The expression of the predicted value is:

[0016] in, w 1 and w 2 are the weight parameters corresponding to the cross network and the residual network, is the predicted value.

[0017] Furthermore, S102 is specifically as follows: considering the relevant parameters that affect the movement of the wall panel, selecting the value range of the parameters, using Latin hypercube sampling for random sampling to obtain multiple groups of parameter combinations, and then combining the fluid-solid coupling numerical solution method, selecting observation points, and generating a database.

[0018] Furthermore, in S102, the plurality of databases include a micro-vibration database and a large-vibration database; When the amplitude of the wall panel is less than or equal to the amplitude threshold, the vibration data of the wall panel related observation points constitute a micro-vibration database; when the amplitude of the wall panel is greater than the amplitude threshold for the first time, the vibration data of the wall panel related observation points constitute a large-amplitude vibration database.

[0019] Furthermore, in S102, the control parameters of the adaptive model switching mechanism are κ , when the amplitude of the panel at 3 / 4 of its length from the left end to the right end is less than or equal to the amplitude threshold, κ =0, select the micro-amplitude vibration database for training the neural network model as a micro-amplitude vibration neural network model; when the amplitude of the wall panel at 3 / 4 of the length from the left end to the right end along its length direction is greater than the amplitude threshold for the first time, κ =1, select the large vibration database to train the neural network model into a large vibration neural network model; Control parameters κ The expression is:

[0020] in, |W 3 / 4 | is the amplitude of the panel at 3 / 4 of its length from the left end to the right end, with the subscript " first "Representative for the first time, δ is the amplitude threshold.

[0021] Furthermore, the amplitude threshold is equal to 0.1.

[0022] Furthermore, S103 is specifically as follows: the motion characteristics of a certain point and the motion characteristics (structural response) of the neighborhood of the point are selected as input characteristics, and the surface pressure (unsteady aerodynamic load) at the point is selected as the target value. A sample data set containing input characteristics and target values ​​is established according to each database, and the sample data set is preprocessed. The model hyperparameters include learning rate, batch size, and number of training times. The training of the neural network model includes initializing the weight parameters of the neural network model, optimizing the parameters of the neural network model, and defining the loss function of the neural network model. The parameters of the neural network model are optimized by combining the predicted value and the target value. A micro-vibration neural network model is trained based on the micro-vibration database of S102, and a large-vibration neural network model is trained based on the large-vibration database of S102. Thus, two different sub-neural network models are trained.

[0023] Furthermore, in S103, a maximum-minimum normalization method is used to preprocess at least part of the input features. The expression of the maximum-minimum normalization method is:

[0024] in, φ is a parameter, φ max and φ min They are φ The maximum and minimum values ​​of are the preprocessed parameters.

[0025] Furthermore, in S103, a standardization method is used to preprocess at least part of the input features and at least part of the target value. The expression of the standardization method is:

[0026] in, φ is a parameter, ε and σ They are φ The mean and standard deviation of are the preprocessed parameters.

[0027] Furthermore, in S103, the loss function is L 2 The mean square error function of the regularization parameter is expressed as:

[0028] in, N in is the number of input samples, y i yes The corresponding target value, η yes L2 regularization parameter, W l It is l The weight parameters of the layer, Loss is the output value of the loss function.

[0029] Finally, the structural solver and the two trained sub-neural network models (i.e., the micro-vibration neural network model and the large-amplitude vibration neural network model) are combined to realize the prediction of the aeroelastic nonlinear response of the wall panel.

[0030] The beneficial effects of this application include: This application provides a data-driven method for predicting the nonlinear aeroelastic response of siding. This method uses a cross network to establish a low-dimensional mapping relationship between cross features and unsteady aerodynamic loads, and implements high-dimensional correction of cross features and unsteady aerodynamic loads through a residual network. Combining the input features of the neural network model with an adaptive model switching mechanism improves the model's prediction accuracy. Compared with traditional numerical simulation methods, the data-driven method for predicting the nonlinear aeroelastic response of siding significantly improves computational efficiency, while also exhibiting good prediction accuracy and generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 A flow chart of a data-driven aeroelastic nonlinear response prediction method for a panel provided in this application; Figure 2 Schematic diagram of the neural network model provided for this application; Figure 3 Cross-network flow chart provided for this application; Figure 4 The residual network flowchart provided for this application; Figure 5 Schematic diagram of the two-dimensional simply supported wall panel system provided for this application; Figure 6 A flowchart for solving the wall panel using a loosely coupled method provided for this application; Figure 7 Schematic diagram of the input features of the neural network model provided in this application; Figure 8 The surface pressure variation and relative error of the numerical simulation using the fluid solver and the neural network model provided in this application; Figure 9 The surface displacement changes and errors of numerical simulations using the fluid solver and the neural network model provided in this application; Figure 10 The wall plate stability boundary is solved using piston theory, fluid solver and the neural network model provided in this application. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of this application, not all of them. Generally, the components of this application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0034] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application as claimed, but merely represents selected embodiments of the present application. It should be noted that, unless there is a conflict, the various features of the embodiments of the present application may be combined with each other, and the combined embodiments are still within the scope of protection of the present application.

[0035] It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures. In addition, it should be noted that in the description of this application, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0036] The technical solution of the present application is described in detail below with reference to specific embodiments.

[0037] Figure 1 A flow chart of a data-driven aeroelastic nonlinear response prediction method for a panel is provided in this application, such as Figure 1 As shown, the method comprises the following steps: S101: Build a neural network model for predicting the nonlinear aeroelastic response of panels. Figure 2 The neural network model diagram provided for this application is as follows: Figure 2 As shown in Figure 1, the neural network model includes input features, a cross network, a residual network, a fully connected network, and a predicted value. The data transmission process of the neural network model is as follows: the input features are input to the cross network and the residual network in parallel, and then combined through the fully connected network to obtain the predicted value.

[0038] Figure 3 The cross-network flow chart provided for this application is as follows: Figure 3As shown in Figure 1, the cross network is used to quickly generate explicit cross features and construct a low-dimensional mapping between cross features and unsteady aerodynamic loads. The cross network includes multiple cross network layers and multiple first fully connected layers. i The expression of the layer-cross network layer is:

[0039] in, x in Is the input feature, which is a vector, ⊙ represents the vector dot product, x c,i and y c,i It is i Layer crosses the input and output of the network layer, W c,i and b c,i It is i The weight parameters and bias parameters of the cross network layer. i =1, x c,1 =x in .

[0040] No. i The expression of the first fully connected layer is:

[0041] in, x fc,i and y fc,i It is i The input and output of the first fully connected layer, W fc,i and b fc,i It is i The weight parameters and bias parameters of the first fully connected layer.

[0042] In this embodiment, the cross network layer is 2 layers, the first full link layer is 2 layers, and the specific expression of the cross network is as follows:

[0043] in, is the output value of the crossover network.

[0044] Figure 4 The residual network flow chart provided for this application is as follows: Figure 4 As shown in Figure 2, the residual network is used to construct a high-dimensional correction between cross features and unsteady aerodynamic loads. The residual network includes a feature cross layer, a convolutional layer, a residual layer, a pooling layer, and a second fully connected layer. The expression of the feature cross layer is:

[0045] in, is the input feature with the added feature 1, ⊗ represents the matrix multiplication, y inter is the two-dimensional output matrix of the feature cross layer.

[0046] No. i The expression of the residual layer is:

[0047] in, W i,1 , W i,2 and W i,3 They are i Different weight parameters of the residual layer, x R,i and y R,i They are i The input and output of the residual layer, f(∙) is the activation function ReLU , BN(∙) It is batch normalization.

[0048] In this embodiment, the residual layer is 3 layers, the second fully connected layer is 3 layers, and the specific expression of the residual network is:

[0049] in, is the expression of the convolutional layer, is the expression of the pooling layer. The convolutional layer and the pooling layer use common techniques of neural networks, and their expressions are not repeated here. The expression of the second fully connected layer is the same as that of the first connected layer, so they are not repeated here. is the output value of the residual network.

[0050] The parameters of the neural network model in this example are shown in the following table:

[0051] S102: Based on the numerical coupling solution method, various structural responses of the aeroelastic nonlinear response of the wall panel and the corresponding unsteady aerodynamic loads are obtained, and multiple databases are constructed; based on the adaptive model switching mechanism, different databases are selected to train the neural network model into different sub-neural network models.

[0052] Figure 5 The schematic diagram of the two-dimensional simply supported wall panel system provided for this application is as follows: Figure 5 As shown, the Mach number M =2, wall panel lengtha =0.1m, the wall thickness is h =0.005m, p s is the surface pressure on the wall, p c is the cavity pressure, x and y is the Cartesian coordinate. The numerical solution method is used to solve Figure 5 The two-dimensional simply supported panel system shown is solved. Figure 6 The flowchart for solving the wall panel using the loosely coupled method provided in this application is as follows: Figure 6 As shown, it includes the following steps: (1) The structural solver solves the structural motion equation to obtain the structural response w ;(2)The fluid solver is based on the structural response w , solve the unsteady aerodynamic load and obtain the surface pressure of the structure P ; (3) The surface pressure of the structure P Passed to the structure for the next iteration; where the subscript n represents the number of iterations.

[0053] In this example, the dimensionless mass ratio is selected μ , dimensionless dynamic pressure λ , infinite rigid in-plane load r x As relevant parameters for controlling the movement of the wall panels. μ The values ​​of are as follows: [0.02, 0.10, 0.20, 0.30, 0.40, 0.50]. λ The value range of is [0, 400], r x The range of is [0, 5.0]. Latin hypercube sampling is used to λ and r x Random sampling, combined μ The value of , multiple sets of parameter combinations are obtained. Each parameter combination [ μ, λ, r x ] Iterate 10000 steps. The dimensionless number of observation points x The coordinates are [0.025, 0.075, ..., 0.975]. The split point is the point where the displacement first exceeds the amplitude threshold (0.1 in this example) at 3 / 4 of the length of the panel from the left end to the right end. The motion data of each iteration step for all observation points before the split point form a micro-vibration database. The motion data of each iteration step for all observation points after the split point form a large-vibration database.

[0054] Then, the adaptive model switching mechanism is introduced. The control parameters of the adaptive model switching mechanism are κ, when the amplitude of the wall panel at 3 / 4 of its length from the left end to the right end is less than or equal to 0.1, κ =0, select the micro-vibration database for training the neural network model into a micro-vibration neural network model; when the amplitude of the wall panel at 3 / 4 of its length from the left end to the right end is greater than 0.1 for the first time, κ When the value is equal to 1, the large-scale vibration database is selected for training the neural network model as a large-scale vibration neural network model.

[0055] Control parameters κ The expression is:

[0056] in, |W 3 / 4 | is the amplitude of the panel at 3 / 4 of its length from the left end to the right end, with the subscript " first "Represents the first time.

[0057] S103: Select at least one structural response and at least one unsteady aerodynamic load as input features and target values, respectively, establish a sample data set containing the input features and target values ​​according to each database, preprocess the sample data set, set model hyperparameters, and train the neural network model into a sub-neural network model according to the sample data set and the model hyperparameters.

[0058] Figure 7 The input feature diagram of the neural network model provided in this application is as follows: Figure 7 As shown, the input features of the neural network model in this example are , , , , , , , , , , , , , , , , The target value in this example is the surface pressure P at a certain location. For the i The degree of wall deformation at the center of each unit, For the i The deformation velocity of the unit center, P i It is i The surface pressure of a unit.

[0059] A sample dataset containing input features and target values ​​is established based on the database. 500 iteration steps of data are randomly selected from each observation point in the micro-vibration database as the micro-vibration sample dataset, and 2000 iteration steps of data are randomly selected from each observation point in the large-vibration database as the large-vibration sample dataset.

[0060] The data of the sample data set is preprocessed, and the preprocessing methods include maximum and minimum normalization and standardization methods.

[0061] The expression of the maximum and minimum normalization method is:

[0062] in, φ is a parameter, φ max and φ min They are φ The maximum and minimum values ​​of are the preprocessed parameters.

[0063] The expression of the normalization method is:

[0064] in, φ is a parameter, ε and σ They are φ The mean and standard deviation of are the preprocessed parameters.

[0065] In this embodiment, μ and λ The maximum and minimum normalization method is used for data preprocessing. , and P Data preprocessing was performed using standardized methods.

[0066] The model hyperparameters in this example are set as follows: learning rate is 0.001, batch size is 256, and training is 15 iterations. L 2 The regularization parameter is set to 0.0001. The weight parameter initialization method of the neural network model is the Xavier method, the parameter optimization method of the neural network model is the Adam optimizer, and the loss function of the neural network model is L 2 is the mean squared error function with regularization parameter.

[0067]

[0068] in, N in is the number of input samples,y i yes The corresponding target value, η yes L 2 regularization parameter, W l It is l The weight parameters of the layer, Loss is the output value of the loss function.

[0069] The neural network model is trained by combining the model parameter optimization method and the sample data set described in S103. The large-amplitude vibration neural network model is trained based on the large-amplitude vibration sample data set in S102, and the small-amplitude vibration neural network model is trained based on the small-amplitude vibration sample data set in S102. Thus, two different sub-neural network models are trained.

[0070] S104: Predict the aeroelastic nonlinear response of the panel based on multiple sub-neural network models.

[0071] Finally, the two trained sub-neural network models (i.e., micro-amplitude vibration neural network model and large-amplitude vibration neural network model) are used to replace Figure 6 The fluid solver is combined with the structural solver to predict the unsteady aerodynamic behavior of the panel.

[0072] Figure 8 for When the fluid solver and the neural network model provided in this application are used to perform numerical simulation on the surface pressure change and relative error, Figure 9 for When the fluid solver and the neural network model provided by this application are used to perform numerical simulation on the surface displacement changes and errors. Figure 8 and Figure 9 As shown in the figure, the displacement and pressure changes predicted by the neural network model are consistent with the results calculated using the fluid solver. The aeroelastic calculation of the panel takes 120 minutes using the fluid solver, while the neural network model only takes 3 minutes, a 40-fold increase in computational efficiency. Figure 10 for μ = 0.02, the wall plate stability boundary is solved using piston theory, fluid solver and the neural network model provided by this application. Figure 10 As shown, the stability boundary of the panel solved by the neural network model is basically consistent with that solved by the fluid solver and the piston theory. The prediction results show that the proposed method for predicting the aeroelastic nonlinear response of the panel has good generalization performance and prediction accuracy, and improves computational efficiency.

[0073] The foregoing description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A data-driven method for predicting the nonlinear aeroelastic response of a panel, characterized by: The following steps are involved: S101: Constructing a neural network model for predicting the aeroelastic nonlinear response of a panel, wherein the neural network model includes input features, a cross network, a residual network, a fully connected network, and a predicted value. The data transmission process of the neural network model is as follows: inputting the input features into the cross network and the residual network in parallel, and then combining them through the fully connected network to obtain the predicted value; S102: Based on a numerical coupling solution method, multiple structural responses of the aeroelastic nonlinear response of the panel and corresponding aerodynamic loads are obtained, and multiple databases are constructed; based on an adaptive model switching mechanism, different databases are selected to train the neural network model into different sub-neural network models; S103: selecting at least one structural response and at least one aerodynamic load as input features and target values, respectively, establishing a sample data set including the input features and the target values ​​according to each database, preprocessing the sample data set, setting model hyperparameters, and training the neural network model into the sub-neural network model according to the sample data set and the model hyperparameters; S104: Predicting the aeroelastic nonlinear response of the panel based on the multiple sub-neural network models.

2. The method according to claim 1, characterized in that In the S101, the cross network includes multiple cross network layers and multiple first full link layers. i The expression of the cross network layer is: in, x in is the input feature, ⊙ represents the vector dot product, x c,i and y c,i It is i The input and output of the cross network layer, W c,i and b c,i It is i The weight parameters and bias parameters of the cross network layer; No. i The expression of the first fully connected layer is: in, x fc,i and y fc,i It is i The input and output of the first fully connected layer, W fc,i and b fc,i It is i The weight parameters and bias parameters of the first fully connected layer; The expression of the cross network is: in, is the output value of the crossover network.

3. The method according to claim 2, characterized in that In S101, the residual network includes a feature cross layer, a convolutional layer, a residual layer, a pooling layer, and a second fully connected layer. The expression of the feature cross layer is: in, is the input feature with the added feature 1, ⊗ represents the matrix multiplication, y inter is the two-dimensional output matrix of the feature cross layer; No. i The expression of the residual layer is: in, W i,1 , W i,2 and W i,3 They are i The different weight parameters of the residual layer described in the layer, x R,i and y R,i They are i The input and output of the residual layer, f(∙) is the activation function ReLU , BN(∙) is batch normalization; The expression of the residual network is: in, is the expression of the convolutional layer, is the expression of the pooling layer, the expression of the second fully linked layer is the same as that of the first linked layer, is the output value of the residual network.

4. The method according to claim 3, characterized in that In S101, the fully linked network combines with the cross network and the residual network to generate the predicted value of the neural network model. The expression of the predicted value is: in, w 1 and w 2 are the weight parameters corresponding to the cross network and the residual network, is the predicted value.

5. The method according to any one of claims 1 to 4, characterized in that In said S102, the plurality of databases include a micro-vibration database and a large-amplitude vibration database; When the amplitude of the wall panel is less than or equal to the amplitude threshold, the vibration data of the wall panel constitutes the micro-vibration database; when the amplitude of the wall panel is greater than the amplitude threshold for the first time, the vibration data of the wall panel constitutes the large-amplitude vibration database.

6. The method according to claim 5, characterized in that In S102, the control parameters of the adaptive model switching mechanism are κ , when the amplitude of the wall panel is less than or equal to the amplitude threshold, κ is equal to 0, the micro-amplitude vibration database is selected to train the neural network model into a micro-amplitude vibration neural network model; when the amplitude of the wall panel is greater than the amplitude threshold for the first time, κ is equal to 1, selecting the large-scale vibration database for training the neural network model into a large-scale vibration neural network model; The control parameters κ The expression is: in, |W | is the amplitude of the panel, subscript " first "Representative for the first time, δ is the amplitude threshold.

7. The method according to claim 5, characterized in that The amplitude threshold is equal to 0.

1.

8. The method according to any one of claims 1 to 4, characterized in that In S103, a maximum-minimum normalization method is used to preprocess at least part of the input features. The expression of the maximum-minimum normalization method is: in, φ is a parameter, φ max and φ min They are φ The maximum and minimum values ​​of are the preprocessed parameters.

9. The method according to any one of claims 1 to 4, characterized in that In S103, a standardization method is used to preprocess at least part of the input features and at least part of the target value. The expression of the standardization method is: in, φ is a parameter, ε and σ They are φ The mean and standard deviation of are the preprocessed parameters.

10. The method according to claim 4, characterized in that In S103, the training of the neural network model includes initializing the weight parameters of the neural network model, optimizing the parameters of the neural network model, and defining the loss function of the neural network model, wherein the loss function is a function with L 2 The mean square error function of the regularization parameter is expressed as: in, N in is the number of input samples, y i yes The corresponding target value, η yes L 2 regularization parameter, W l It is l The weight parameters of the layer, Loss is the output value of the loss function.

Images