Method for predicting wall heat flux considering material ablation response under semi-supervised learning framework

Abstract

The application discloses a wall heat flow prediction method considering material ablation response under a semi-supervised learning framework, and relates to the field of aerospace. The method comprises the following steps: constructing a data set; building a neural network model and introducing an ablation physical information loss term into a loss function of the neural network model; pre-training the neural network model; predicting unlabelled data through the pre-trained neural network model to generate pseudo labels; mixing labelled data and unlabelled data to semi-supervise and train the pre-trained neural network model to obtain a final neural network model; and predicting wall heat flow of material ablation response through the final neural network model. The method reduces the demand of the model on the labelled data set, enhances the model generalization ability under the condition of small sample training, increases the physical interpretability of the network model, and improves the wall heat flow prediction precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aerospace, and particularly relates to a wall heat flux prediction method considering material ablation response under a semi-supervised learning framework. BACKGROUND

[0002] During flight, the surface heat flux of the aircraft increases rapidly with the 3rd power of the flight speed, which leads to a more severe aerodynamic heating environment of the aircraft. In order to protect the aircraft from the severe aerodynamic heating under the flight condition, an ablation heat protection system is often added to the surface of the aircraft, which can absorb a large amount of heat through the injection of ablation gas generated by the ablation material, thereby reducing the heat flux density of the aircraft wall and effectively avoiding the aircraft from being burned due to the excessively high heat flux. Different ablation protection materials and different structural forms will directly affect the wall heat flux of the aircraft, and therefore, in the process of aircraft design, accurate prediction of the wall heat flux is the key to the design of the ablation heat protection system.

[0003] The current methods commonly used to predict wall heat flux are: ablation wind tunnel test, empirical formula prediction, CFD-ablation coupling calculation and full-supervised neural network prediction. Among them, the ablation wind tunnel test is too high in cost and it is difficult to obtain a large amount of experimental data; the empirical formula prediction result deviates too much from the true result, and cannot accurately predict the wall heat flux data of the aircraft; the CFD-ablation coupling calculation can realize fine prediction of the ablation heat flux of the aircraft wall, but this method is high in cost and time-consuming, and it often takes 1-2 months or even longer to calculate along the flight trajectory (the flight trajectory is hundreds of points, and the incoming flow conditions of each point are different, so it is difficult to obtain the coupling calculation); therefore, the full-supervised neural network is used to quickly predict the wall ablation heat flux, which is a common method for predicting the wall ablation heat flux at present, which takes the wall heat flux without considering ablation, the boundary layer outer edge parameter, the incoming flow parameter and the ablation mass flow rate as the input features of the full-supervised neural network, considers the wall heat flux under the ablation condition as the output feature, constructs a full connection neural network model, and after a large amount of data is trained, the wall heat flux data under the ablation condition can be obtained only by inputting the data without ablation and the mass flow rate data. However, through long-term research, it is found that this method also has some shortcomings, one of which is that the data dependency is prominent. The model training needs a large amount of training data set with strict corresponding relationship (the ablation heat flux, the ablation mass flow rate and the target ablation heat flux need to be one-to-one corresponding, that is, the labeled data set), and each data sample needs to be obtained by coupling simulation of computational fluid dynamics (CFD) and ablation model, which takes several weeks to generate a single data, and the cost of large-scale data set construction increases exponentially. The second is that the small sample generalization ability is weak, when the training data is too small, the model prediction error generally exceeds the engineering allowable threshold of 15%, and the local area heat flow prediction distortion phenomenon is easy to occur, which is difficult to meet the high-precision aerodynamic thermal protection design requirement; the third is the lack of physical interpretability. Although the existing network architecture has strong non-linear fitting ability, the "black box" characteristic makes it impossible to analyze the physical relationship between the mass flow rate and the heat flux, which seriously restricts the reliability verification of the model in the design of new aircraft. Therefore, it is urgent to construct a prediction model with the following characteristics: (1) reduce the demand of the model for labeled data set, (2) reduce the training data under the premise of ensuring the prediction accuracy, that is, enhance the model generalization ability under the condition of small sample amount training (3) increase the physical interpretability of the network model. SUMMARY

[0004] In view of the above problems in the prior art, the wall heat flux prediction method considering material ablation response under the semi-supervised learning framework provided by the present application solves the problems of large demand for labeled data, weak generalization ability of small sample amount data and poor physical interpretability of the existing full-supervised network model.

[0005] In order to achieve the above invention purposes, the technical scheme adopted by the present application is:

[0006] A wall heat flux prediction method considering material ablation response under a semi-supervised learning framework is provided, comprising the following steps:

[0007] Smooth wall heat flux distribution data and ablation mass flow rate are obtained through CFD calculation and ablation calculation method; part of the ablation wall heat flux distribution data is obtained through CFD-ablation coupling calculation and is taken as the true label; wherein when the smooth wall heat flux distribution data and the ablation mass flow rate have corresponding ablation wall heat flux distribution data, the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as labeled data, otherwise the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as unlabeled data;

[0008] A neural network model is built, the ablation physical information loss term is calculated, and the sum of the ablation physical information loss term and the mean square error loss function is taken as the loss function of the neural network model;

[0009] Based on the loss function of the neural network model, the neural network model is pre-trained using the labeled data to obtain a pre-trained neural network model;

[0010] The unlabeled data is predicted by the pre-trained neural network model to generate pseudo labels;

[0011] Based on the loss function of the neural network model, the pre-trained neural network model is semi-supervised self-trained using a data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo labels to obtain a final neural network model;

[0012] The wall heat flux prediction of the material ablation response is performed through the final neural network model.

[0013] Further, the smooth wall heat flux distribution data includes boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, total enthalpy of incoming flow, wall temperature, wall pressure and smooth wall heat flux; the ablation mass flow rate is obtained through ablation calculation.

[0014] Further, the neural network model is a deep neural network model.

[0015] Further, the expression of the loss function of the neural network model is:

[0016]

[0017]

[0018] wherein is the loss function of the neural network model; is the mean square error loss function; is an ablation physical information loss term; n is a total number of labeled data; is a label prediction result of the neural network model on the labeled data; is a true label.

[0019] Further, the calculation expression of the ablation physical information loss term is:

[0020]

[0021] wherein is a non-ablation wall heat flux; is a gas injection mass flow rate; is a boundary layer outer edge static enthalpy; is a recovery coefficient; is a boundary layer outer edge velocity.

[0022] Further, the specific method for pre-training the neural network model by using the labeled data is:

[0023] The labeled data is taken as the input of the deep neural network model, the label of the input labeled data is predicted by the deep neural network model, the predicted label and the corresponding true label are taken as the calculation basis of the mean square error loss function, the ablation physical information loss term is calculated, and then the loss function value of the neural network model is obtained; the deep neural network model is optimized by the loss function value of the neural network model, and the pre-training is completed.

[0024] Further, the specific method for semi-supervised self-training of the pre-trained neural network model by using the data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo label includes the following steps:

[0025] The labeled data and the unlabeled data are mixed to obtain mixed data;

[0026] The current neural network model is trained by the mixed data, and the normalized loss of the unlabeled data is calculated; wherein the current neural network model before training by the mixed data is the pre-trained neural network model;

[0027] The unlabeled data with a normalized loss less than 0.5 is divided into labeled data and the mixed data is updated; the current neural network model is updated by the loss function value of the neural network model;

[0028] The semi-supervised self-training of the current neural network model by the mixed data is repeated until the amount of remaining unlabeled data is less than one fifth of the initial total amount of unlabeled data or the pre-set training round number is reached, and the final neural network model is obtained.

[0029] Further, the specific method for predicting the wall heat flux of material ablation response by the final neural network model is as follows:

[0030] The smooth wall heat flux distribution data and the ablation mass flow rate of the object to be predicted are obtained by the CFD calculation and the ablation calculation method and are taken as the input of the final neural network model, and the ablation wall heat flux distribution data output by the final neural network model is taken as the prediction result of the wall heat flux of material ablation response.

[0031] A computer device is provided, which comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to enable the processor to execute the wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0032] A computer readable storage medium is provided, which stores a computer program, and the computer program is executed by the processor to enable the processor to execute the wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0033] The beneficial effects of the present application are as follows:

[0034] 1. The demand of the model for the labeled data set is reduced. The present neural network model fully utilizes the unlabeled data, and the CFD-ablation coupling calculation is not required for each working condition during the calculation along the flight trajectory, so that the cost of constructing the data set is greatly reduced.

[0035] 2. The generalization ability of the model under the small sample training condition is enhanced. Compared with the traditional fully supervised neural network model, the wall heat flux prediction accuracy of the neural network model under the small sample training condition is improved by 30%.

[0036] 3. The physical interpretability of the network model is increased. The ablation physical information loss term is introduced into the loss function of the neural network model, the relationship between the ablation wall heat flux, the ablation mass flow rate and the non-ablation wall heat flux is accurately described, and the physical meaning of the network model is increased. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is a flowchart of the present method;

[0038] Figure 2 is a schematic diagram of the stardust calculation model;

[0039] Figure 3 is a comparison diagram of the calculation result and the machine learning prediction result of T66;

[0040] Figure 4 is a comparison diagram of the calculation result and the machine learning prediction result of T76. DETAILED DESCRIPTION

[0041] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0042] As shown in the figure, the wall heat flux prediction method considering material ablation response under the semi-supervised learning framework includes the following steps: Figure 1

[0043] S1, obtain smooth wall heat flux distribution data and ablation mass flow rate through CFD calculation and ablation calculation method; obtain part of the ablation wall heat flux distribution data through CFD-ablation coupling calculation and take it as the true label; wherein when the smooth wall heat flux distribution data and the ablation mass flow rate have corresponding ablation wall heat flux distribution data, the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as labeled data, otherwise the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as unlabeled data;

[0044] S2, build a neural network model, calculate the ablation physical information loss term and take the sum of the ablation physical information loss term and the mean square error loss function as the loss function of the neural network model;

[0045] S3, based on the loss function of the neural network model, pre-train the neural network model with labeled data to obtain a pre-trained neural network model;

[0046] S4, predict the unlabeled data through the pre-trained neural network model to generate pseudo labels;

[0047] S5, based on the loss function of the neural network model, use the data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo labels to semi-supervisedly self-train the pre-trained neural network model to obtain a final neural network model;

[0048] S6, predict the wall heat flux of the material ablation response through the final neural network model.

[0049] In the specific implementation process, the smooth wall heat flux distribution data includes boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, total enthalpy of incoming flow, wall temperature, wall pressure and smooth wall heat flux; the ablation mass flow rate is obtained by ablation calculation. The neural network model is a deep neural network model (DNN). The parameters input into the neural network model are used to participate in the prediction of the label by the neural network model, and affect the calculation of the mean square error loss function, thereby promoting the prediction accuracy of the label by the neural network model.​

[0050] The expression of the loss function of the neural network model in step S2 is as follows:

[0051]

[0052]

[0053] wherein is the loss function of the neural network model; is the mean square error loss function; is the ablation physical information loss term; n is the total number of labeled data; is the label prediction result of the neural network model on the labeled data; is the real label.

[0054] Ablation physical information loss term The calculation expression of the ablation physical information loss term is as follows:

[0055]

[0056] wherein is the ablation wall heat flux; is the gas injection mass flow rate; is the outer edge static enthalpy of the boundary layer; is the recovery coefficient; is the outer edge velocity of the boundary layer.

[0057] The specific method of pre-training the neural network model by using the labeled data in step S3 is as follows:

[0058] The labeled data is taken as the input of the deep neural network model, the label of the input labeled data is predicted by the deep neural network model, the predicted label and the corresponding real label are taken as the calculation basis of the mean square error loss function, the ablation physical information loss term is calculated, and then the loss function value of the neural network model is obtained; the deep neural network model is optimized by the loss function value of the neural network model, and the pre-training is completed.

[0059] In the embodiment, the specific method of semi-supervised self-training the pre-trained neural network model by using the data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo label includes the following steps:

[0060] S5-1, mixing the labeled data and the unlabeled data to obtain mixed data;

[0061] S5-2, training the current neural network model by mixed data and calculating the normalized loss of the unlabeled data; wherein the current neural network model before training by mixed data is a pre-trained neural network model;

[0062] S5-3, dividing the unlabeled data with a normalized loss less than 0.5 into labeled data and updating the mixed data; updating the current neural network model by the loss function value of the neural network model;

[0063] S5-4, repeating the semi-supervised self-training of the current neural network model by mixed data until the amount of remaining unlabeled data is less than one fifth of the total amount of initial unlabeled data or reaches a pre-set training round number, and obtaining a final neural network model.

[0064] The specific method of wall heat flux prediction of material ablation response by the final neural network model in step S6 is as follows:

[0065] The smooth wall heat flux distribution data and ablation mass flow rate of the object to be predicted are obtained by CFD calculation and ablation calculation method and used as the input of the final neural network model, and the ablation wall heat flux distribution data output by the final neural network model is used as the wall heat flux prediction result of material ablation response.

[0066] The embodiment also provides a computer device comprising a memory and a processor, the memory storing a computer program, and the computer program is executed by the processor to make the processor execute the wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0067] The embodiment also provides a computer readable storage medium storing a computer program, and the computer program is executed by the processor to make the processor execute the wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0068] In an embodiment of the present application, the accuracy of the method is verified by taking the stardust model as an example. The schematic diagram of the stardust model is shown in Figure 2 The inflow conditions and the required calculation methods are shown in Table 1.

[0069] Table 1: Inflow conditions and required calculation methods

[0070]

[0071] Among them, the ablation wall heat flux of working conditions T34 and T60 is obtained by CFD-ablation coupling calculation, that is, there is ablation heat flux data that can be used as the output data of the neural network, so the working conditions T34 and T60 are called labeled data and divided into a training set; T42, T48 and T54 are divided into an unlabeled data set; T66 and T76 are divided into a test data set.

[0072] The ablation heat flow comparison results of the CFD-ablation coupling calculation results, the traditional full-supervised neural network prediction results and the results of the method are shown in FIGS. Figure 3 and Figure 4 As shown in the figures, the ablation heat flow results obtained by the method have a higher degree of agreement with the CFD-ablation coupling calculation results, the prediction error is within 5% under the same training data (a small amount of training data), and the prediction accuracy is significantly better than that of the traditional full-supervised neural network model.

[0073] In summary, the present application proposes a wall heat flow prediction method considering material ablation response under a semi-supervised learning framework, which reduces the demand of the model for labeled data sets, enhances the model generalization ability under small sample training conditions, increases the physical interpretability of the network model, and improves the wall heat flow prediction accuracy.

Claims

1. A wall heat flux prediction method considering material ablation response under a semi-supervised learning framework, characterized in that, The method comprises the following steps: obtain smooth wall heat flux distribution data and ablation mass flow rate through CFD calculation and ablation calculation method; obtain part of the ablation wall heat flux distribution data through CFD-ablation coupling calculation and take it as the real label; wherein when the smooth wall heat flux distribution data and the ablation mass flow rate have corresponding ablation wall heat flux distribution data, the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as labeled data, otherwise the corresponding smooth wall heat flux distribution data and the ablation mass flow rate are recorded as unlabeled data; build a neural network model, calculate the ablation physical information loss term and take the sum of the ablation physical information loss term and the mean square error loss function as the loss function of the neural network model; based on the loss function of the neural network model, pre-train the neural network model using the labeled data to obtain a pre-trained neural network model; predict the unlabeled data through the pre-trained neural network model to generate pseudo labels; based on the loss function of the neural network model, perform semi-supervised self-training on the pre-trained neural network model using a data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo labels to obtain a final neural network model; predict the wall heat flux of the material ablation response through the final neural network model; the smooth wall heat flux distribution data includes boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, incoming flow total enthalpy, wall temperature, wall pressure and smooth wall heat flux; the ablation mass flow rate is obtained through ablation calculation; the expression of the loss function of the neural network model is: wherein is a loss function of the neural network model; is a mean square error loss function; is an ablation physical information loss term; n is a total number of labeled data; is a label prediction result of the neural network model on the labeled data; is a real label; Ablation physical information loss term The calculation expression is: where is the ablated wall heat flux; is the gas injection mass flow rate; is the free-stream static enthalpy; is the recovery coefficient; is the free-stream velocity.

2. The method according to claim 1, wherein, the neural network model is a deep neural network model.

3. The method of claim 2, wherein the material ablation response is considered in the semi-supervised learning framework for wall heat flux prediction. The specific method for pre-training the neural network model using the labeled data is: take the labeled data as the input of the deep neural network model, predict the label of the input labeled data through the deep neural network model, take the predicted label and the corresponding real label as the calculation basis of the mean square error loss function, calculate the ablation physical information loss term, and then obtain the loss function value of the neural network model; optimize the deep neural network model through the loss function value of the neural network model to complete the pre-training.

4. The method of claim 1, wherein the material ablation response is considered in the semi-supervised learning framework. The specific method for performing semi-supervised self-training on the pre-trained neural network model using the data set formed by the labeled data, the ablation wall heat flux distribution data, the unlabeled data and the pseudo labels comprises the following steps: mix the labeled data and the unlabeled data to obtain mixed data; train the current neural network model through the mixed data and calculate the normalized loss of the unlabeled data; wherein the current neural network model before training through the mixed data is the pre-trained neural network model; divide the unlabeled data with a normalized loss less than 0.5 into labeled data and update the mixed data; update the current neural network model through the loss function value of the neural network model; repeat the semi-supervised self-training of the mixed data on the current neural network model until the remaining amount of the unlabeled data is less than one fifth of the initial total amount of the unlabeled data or the pre-set training number of rounds is reached to obtain the final neural network model.

5. The method of claim 1, wherein, A specific method for predicting the wall heat flux of material ablation response through the final neural network model is as follows: The smooth wall heat flux distribution data and the ablation mass flow rate of the object to be predicted are obtained through CFD calculation and ablation calculation method and are used as the input of the final neural network model, and the ablation wall heat flux distribution data output by the final neural network model is used as the prediction result of the wall heat flux of material ablation response.

6. A computer device, comprising: The memory stores a computer program, and the computer program is executed by the processor to make the processor execute the wall heat flux prediction method considering material ablation response in the semi-supervised learning framework according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The memory stores a computer program, and the computer program is executed by the processor to make the processor execute the wall heat flux prediction method considering material ablation response in the semi-supervised learning framework according to any one of claims 1-5.

Images