Honeycomb sandwich panel surface thermal load inversion method and system and medium

Abstract

The invention relates to a honeycomb sandwich panel surface thermal load inversion method and system and a medium, and the method comprises the steps: obtaining a simulation result data set through a simulation program based on changing working condition parameters needing inversion; carrying out an actual experiment by adopting the same working condition parameters needing to be inverted, and carrying out measurement to obtain an actual measurement data set; from the actual measurement data set and the simulation result data set, features which are sensitive to thermal loads and extracted based on plate temperature field data serve as input data of a deep neural network; constructing a deep neural network by using a domain adaptive transfer learning framework; training the deep neural network based on the input data of the simulation data set and the actual measurement data set; and inputting the to-be-predicted characteristics sensitive to the thermal load into the thermal load inversion model to obtain predicted thermal load parameters. The method has very high precision in simulation data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of adaptive neural networks. In particular, it relates to a honeycomb sandwich panel surface heat load inversion method, system and medium. BACKGROUND

[0002] Honeycomb sandwich panel structures are widely used in the field of aerospace, mainly because of their good thermal protection performance. When thermal load is applied to the surface of the structure, it will cause the temperature rise of the structure surface, and further cause thermal stress. According to the different power of the thermal load, it will cause different degrees of damage, which may affect the internal electronic components of the structure, or even burn the surface material. Therefore, it is particularly important to establish a method for inverting the characteristics of the external thermal load of the structure.

[0003] The external load parameters of the structure that need to be inverted include two, which are the radiation diameter of the thermal load and the radiation power. The input parameters are the space-time variation data of the temperature field of the structure panel.

[0004] The existing methods for solving the heat conduction inversion problem are mainly based on conjugate gradient method, dynamic matrix control method and Kalman filter method. For complex structures, the method based on neural network is more suitable. However, in practical application, the neural network method needs a large amount of training data for training, and the inversion model established has dependence on these training data. When encountering input samples with too large deviation from the training data, the prediction accuracy will decrease. For example, the neural network inversion model trained based on a large number of numerical simulations performs poorly in actual environmental test data. How to solve the problem that the inversion model has dependence on these training data, and when encountering input samples with too large deviation from the training data, the prediction accuracy will decrease, is a problem that needs to be solved. SUMMARY

[0005] The present application provides a honeycomb sandwich panel surface heat load inversion method, system and medium to solve the technical difficulty that the neural network in the prior art performs poorly in actual environmental test data.

[0006] To achieve the above purpose, in a first aspect, the present application relates to a honeycomb sandwich panel surface heat load inversion method, comprising: for inverting the thermal load parameters outside the honeycomb sandwich panel, including:

[0007] Based on the need to invert the working condition parameters, a simulation program is used to establish a finite element model for the panel temperature field data, to obtain a simulation result data set, wherein the need to invert the working condition parameters is the radiation diameter of the thermal load and the radiation power of the thermal load, and the panel temperature field data is the temperature data of a plurality of preset measurement points on the honeycomb sandwich panel varying with time;

[0008] An actual experiment is carried out using the same working condition parameters that need to be inverted as the simulation result data set, and corresponding plate temperature field data are measured as the measured data set;

[0009] From the measured data set and the simulation result data set, a feature sensitive to thermal load extracted from the plate temperature field data is taken as input data of a deep neural network;

[0010] A deep neural network for inverting the working condition parameters that need to be inverted is constructed using a domain adaptation transfer learning framework, the simulation data set is taken as a source domain of transfer learning, and the measured data set is taken as a target domain of transfer learning, wherein the domain adaptation transfer learning framework comprises a feature extractor, a domain discriminator, and a predictor, and the architecture of the predictor is a ConvLSTM-based network;

[0011] The deep neural network is trained based on input data of the simulation data set and input data of the measured data set until a loss function converges to a preset threshold, and a honeycomb sandwich panel surface thermal load inversion model is obtained;

[0012] A feature sensitive to thermal load that needs to be predicted is input into the honeycomb sandwich panel surface thermal load inversion model, and the predicted thermal load parameter is obtained.

[0013] Preferably, the feature sensitive to thermal load is extracted from the plate temperature field data as input variables of the neural network, and specifically:

[0014] The plate temperature field data are respectively differentiated with respect to gradient and time to obtain temperature gradient field data T grad and temperature rate of change field data T vary , wherein the formula is:

[0015] wherein Δx is the step length of temperature data along the length direction, Δy is the step length of temperature data along the width direction, Δt is the time step of temperature data sampling, i and j respectively represent the data point serial numbers of temperature data along the x and y directions of the plate, the value range of i is an integer from 1 to 26, the value range of j is an integer from 1 to 26, and the undefined values at the boundaries are replaced by zero values;

[0016] k represents the kth time segment of sensor data, and the value range is an integer from 1 to 9;

[0017] The plate temperature field data T, the temperature gradient field data R grad , and the temperature rate of change field data T varyThe heat-sensitive feature is constructed to obtain a multi-dimensional array (S, C, W, H) corresponding to the plate temperature field data, wherein S is a time dimension, C is a heat-sensitive feature dimension, W is a width dimension, and H is a height dimension. When multiple input data are simultaneously input into the deep neural network, the multi-dimensional array is (B, S, C, W, H), and B represents a batch dimension.

[0018] Preferably, in the field adaptive transfer learning framework, during the transfer learning forward propagation process, the input data first enters the feature extractor for feature extraction. The extracted features are divided into two paths. One path enters the domain discriminator, and the domain discriminator outputs whether the input data comes from the source domain or the target domain. The other path enters the predictor based on ConvLSTM. The predictor infers the working condition parameters to be inverted according to the extracted features as the output. During the error backpropagation process, the loss of the domain predicted by the domain discriminator is the cross-entropy loss. This part of error backpropagates to the domain discriminator and then passes through the gradient reversal layer. The gradient reversal layer changes the gradient to the opposite number and propagates it to the feature extractor.

[0019] Preferably, the feature extractor uses a two-dimensional convolutional neural network. The time dimension and the feature dimension of the input data are expanded to the same dimension in the expansion layer. The feature extractor includes two two-dimensional convolutional layers and an activation function layer arranged between the two two-dimensional convolutional layers. During initialization, the middle elements of the convolution kernel of the feature extractor are set to 1, and the other elements are set to 0, so that the input and output of the feature extractor are the same in the initial state. Each two-dimensional convolutional layer divides the input channels into 27 groups, each group corresponding to one convolution kernel. The 27 input channels and the 27 output channels are one-to-one connected.

[0020] Preferably, the domain discriminator has the following structure: a batch normalization layer, three three-dimensional convolutional layers, an activation function layer arranged after each three-dimensional convolutional layer, a full connection layer, and a Softmax layer. The Softmax layer is responsible for mapping the result to valid probability values. The convolution kernel size of the three-dimensional convolutional layer is (1, 3, 3).

[0021] Preferably, the network structure of the predictor includes: a batch normalization layer, four ConvLSTM layers, an activation function layer arranged after each ConvLSTM layer, a Dropout layer, a three-dimensional convolutional layer, a three-dimensional maximum pooling layer, a three-dimensional average pooling layer, a Dropout layer, and two full connection layers. The Dropout layer randomly inactivates neurons.

[0022] To achieve the above object, the second aspect of the present application relates to a honeycomb sandwich panel surface heat load inversion system for inverting the heat load parameters outside the honeycomb sandwich panel, comprising:

[0023] A simulation module is configured to establish a finite element model for the panel temperature field data using a simulation program based on the changed inversion-required working condition parameters, to obtain a simulation result data set, wherein the inversion-required working condition parameters are the radiation diameter of the heat load and the radiation power of the heat load, and the panel temperature field data are the temperature data of a plurality of preset measuring points on the honeycomb sandwich panel varying with time;

[0024] A real data acquisition module is configured to acquire a real data set measured in an actual experiment, wherein the actual experiment is carried out using the same inversion-required working condition parameters as the simulation result data set, and the corresponding panel temperature field data measured are taken as the real data set;

[0025] An input data extraction module is configured to extract features sensitive to the heat load from the real data set and the simulation result data set as input data of a deep neural network;

[0026] A deep neural network construction module is configured to use a domain adaptive transfer learning framework to construct a deep neural network for inverting the inversion-required working condition parameters, take the simulation data set as a source domain of transfer learning, and take the real data set as a target domain of transfer learning, wherein the domain adaptive transfer learning framework comprises a feature extractor, a domain discriminator, and a predictor, and the architecture of the predictor is a network based on ConvLSTM;

[0027]

[0028] A training module is configured to train the deep neural network based on the input data of the simulation data set and the input data of the real data set until a loss function converges to a preset threshold, to obtain a honeycomb sandwich panel surface heat load inversion model;

[0029] A prediction module is configured to input features sensitive to the heat load to be predicted into the honeycomb sandwich panel surface heat load inversion model, to obtain predicted heat load parameters.

[0030] To achieve the above object, the third aspect of the present application further relates to a computer-readable storage medium, which stores instructions, and the instructions perform the above-mentioned honeycomb sandwich panel surface heat load inversion method when executed.

[0031] The honeycomb sandwich panel surface heat load inversion method, system and medium of the present application have the following beneficial effects compared with the prior art:

[0032] ​The honeycomb sandwich panel surface thermal load inversion method developed in the application can fuse simulation data and experimental data for training. The method utilizes the characteristics of the domain adaptive method, and in the neural network model training process, a large amount of simulation data and a part of measured data are organically combined to train the network together, so that the obtained network model not only has high accuracy in the simulation data, but also has good performance in the measured data. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A flowchart of the honeycomb sandwich panel surface thermal load inversion method in Example 1.

[0034] Figure 2 The domain adaptive method architecture diagram of the embedded ConvLSTM neural network of Example 1 of the honeycomb sandwich panel surface thermal load inversion method in Example 1.

[0035] Figure 3 The Abaqus modeling schematic diagram of the honeycomb sandwich panel thermal load inversion problem of Example 1 of the honeycomb sandwich panel surface thermal load inversion method in Example 1.

[0036] Figure 4 The simulation data set and experimental data set example thermal load sensitive feature visualization of Example 1 of the honeycomb sandwich panel surface thermal load inversion method in Example 1.

[0037] Figure 5 The deep transfer learning neural network model training loss curve of Example 1 of the honeycomb sandwich panel surface thermal load inversion method in Example 1.

[0038] Figure 6 The structure schematic diagram of the honeycomb sandwich panel surface thermal load inversion system in Example 2 of the application. DETAILED DESCRIPTION

[0039] The application will be further described below in conjunction with the drawings and examples. It can be understood that the specific examples described herein are only used to explain the application, and not to limit the application. In addition, it should be noted that, in order to facilitate description, only the parts related to the application are shown in the drawings, not all structures.

[0040] Example 1

[0041] A honeycomb sandwich panel surface thermal load inversion method, please refer to Figures 1-5 The honeycomb sandwich panel surface thermal load inversion method of the application is used to invert the thermal load parameters outside the honeycomb sandwich panel, and includes the following steps: S101 to S106.

[0042] S101, based on the changed need to be inverted working condition parameters, using simulation program to establish a finite element model for the plate temperature field data, get simulation result data set, wherein, the need to be inverted working condition parameters are the radiation diameter of the thermal load and the radiation power of the thermal load, the plate temperature field data are the temperature data of the preset multiple measuring points on the honeycomb sandwich panel changing with time.

[0043] In the embodiment, the simulation program can use simulation software such as Abaqus, etc., change the need to be inverted working condition parameters, and obtain a series of simulation result data.

[0044] S102 uses the same need to be inverted working condition parameters as the simulation result data set to carry out actual experiment, and measures to obtain corresponding plate temperature field data as the measured data set.

[0045] S103 extracts the features sensitive to the thermal load based on the plate temperature field data from the measured data set and the simulation result data set as the input data of the deep neural network.

[0046] In the embodiment, S103 is specifically:

[0047] In order to extract the spatial and temporal features in the data, the gradient and time derivative of the plate temperature field data are calculated to obtain the temperature gradient field data T grad and the temperature change rate field data T vary , wherein the formula is: Wherein, Δx is the step length of the temperature data along the length direction, Δy is the step length of the temperature data along the width direction, Δt is the time step of the temperature data sampling, i and j represent the data point serial number of the temperature data along the x and y directions of the plate respectively, the value range of i is an integer from 1 to 26, the value range of j is an integer from 1 to 26, and the value of the undefined boundary is replaced by a zero value; K represents the kth time segment of the sensor data, and the value range is an integer from 1 to 9, wherein Δx and Δy are the step lengths of the temperature data along the length and width directions, respectively, that is, the distances between the adjacent two measuring points in the x and y directions, respectively, and Δt is the time step of the temperature data sampling, corresponding to the time between adjacent data time steps.

[0048] Wherein, i and j correspond to the length and width direction indexes of the sensor data, respectively, in the embodiment, the value range of i is an integer from 1 to 26, and the value range of j is an integer from 1 to 26, wherein k corresponds to the time index of the sensor data, and the value range is an integer from 1 to 9.

[0049] The undefined values at the boundaries are replaced by zero values. Specifically, the boundaries refer to the boundary indexes of 1 and 26 in length and width, and the boundary indexes of 1 and 9 in time. When calculating the spatial derivative and the time derivative at the above indexes, the data corresponding to the indexes of 0, 27 and 10 are required based on the above formula, but the data at these indexes are undefined, and are replaced by zero values: i = 1 or j = 1, then T (1-1) undefined, take zero value; i = 26 or j = 26, then T (26+1) undefined, take zero value; k = 9, then T (9+1) undefined, take zero value.

[0050] The plate temperature field data T, the temperature gradient field data T grad and the temperature change rate field data T vary are constructed to obtain a four-dimensional multi-dimensional array (S, C, W, H) corresponding to the plate temperature field data, wherein S is a time dimension, C is a heat-sensitive feature dimension, W is a width dimension, and H is a height dimension.

[0051] In order to improve the prediction efficiency, the neural network usually inputs multiple input data into the deep neural network at the same time, and the input data samples are usually stacked in the first dimension to form a batch size dimension, so the multi-dimensional array dimension of each input batch can be represented as (B, S, C, W, H), and B represents a batch dimension.

[0052] In S104, a deep neural network for inverting the working condition parameters to be inverted is constructed using a domain adaptive transfer learning framework, and a simulation data set is used as a source domain of transfer learning and a measured data set is used as a target domain of transfer learning, wherein the domain adaptive transfer learning framework comprises a feature extractor, a domain discriminator and a predictor, and the architecture of the predictor is a ConvLSTM-based network.

[0053] In this embodiment, as shown in Figure 3 in the domain adaptive transfer learning framework, in the forward propagation process of transfer learning, the input data first enters the feature extractor for feature extraction, and the extracted features are divided into two paths, one path enters the domain discriminator, and the domain discriminator outputs whether the input data comes from the source domain or the target domain, and the other path enters the ConvLSTM-based predictor, and the predictor infers the working condition parameters to be inverted according to the extracted features as the output. In the error backpropagation process, the loss of the domain predicted by the domain discriminator is the cross-entropy loss, and after the error backpropagation to the domain discriminator, the gradient inversion layer will propagate the gradient to the feature extractor, which is the opposite number, that is, the optimization of the feature extractor and the domain discriminator is in two opposite directions. This method is similar to the concept of the adversarial generative network, and is the core of the domain adaptive method.

[0054] In this embodiment, the feature extractor uses a two-dimensional convolutional neural network, which expands the time dimension and the feature dimension of the input data into the same dimension in the expansion layer, includes two two-dimensional convolutional layers and an activation function layer arranged between the two two-dimensional convolutional layers, and during initialization, the middle element of the convolution kernel of the feature extractor is set to 1 and other elements are set to 0, so that in the initial state, the input of the feature extractor is the same as the output. Each two-dimensional convolutional layer divides the input channel into 27 groups, each group corresponds to one convolution kernel, and 27 input channels and 27 output channels are one-to-one connected.

[0055] The specific feature extractor uses a two-dimensional convolutional neural network, and its structure is shown in Table 1. Fatten is an expansion layer that expands the time dimension and the feature dimension of the data into the same dimension. Conv2d-i is the i-th two-dimensional convolutional layer, and ReLU is an activation function layer. During initialization, the middle element of the convolution kernel of the feature extractor is set to 1 and other elements are set to 0, so that in the initial state, the input of the feature extractor is the same as the output. The grouping of the two-dimensional convolutional layer is 27, which means that the input channel is divided into 27 groups, each group corresponds to one convolution kernel, that is, 27 input channels and 27 output channels are one-to-one connected, and the number of convolution kernels is only 27.

[0056]

[0057] Table 1

[0058] In this embodiment, the domain discriminator structure is: a batch normalization layer, three three-dimensional convolutional layers, an activation function layer arranged after each three-dimensional convolutional layer, a fully connected layer and a Softmax layer. The Softmax layer is responsible for mapping the result to a valid probability value, and the convolution kernel size of the three-dimensional convolutional layer is (1, 3, 3).

[0059] The specific domain discriminator structure is shown in Table 2. BatchNorm is a batch normalization layer, Conv3d-i is the i-th three-dimensional convolutional layer, ReLU is an activation function layer, Dense is a fully connected layer, and Softmax is responsible for mapping the result to a valid probability value. The convolution kernel size is (1, 3, 3), that is, no convolution is performed in the time dimension to avoid interference with spatial feature extraction.

[0060]

[0061] Table 2

[0062] In the embodiment, the network structure of the predictor comprises: a batch normalization layer, four ConvLSTM layers, an activation function layer arranged after each ConvLSTM layer, a Dropout layer, a three-dimensional convolution layer, a three-dimensional maximum pooling, a three-dimensional average pooling layer, a Dropout layer and two fully connected layers, and the Dropout layer randomly inactivates neurons.

[0063] The predictor deep neural network is constructed based on a ConvLSTM neural network architecture, which is a kind of neural network improved based on an LSTM recurrent neural network. Because the convolution operation and the characteristics of the recurrent network are introduced, it is suitable for extracting spatial features and time features in data. The deep neural network structure based on ConvLSTM designed by the application is shown in Table 3, and from top to bottom, it is from the input layer to the output layer of the neural network. BatchNorm represents the batch normalization layer, ConvLSTM-i represents the i-th ConvLSTM layer, LeakyReLU is the activation function layer, the Dropout layer can randomly inactivate neurons, which plays a role of regularization and accelerates training, Conv3d is a three-dimensional convolution layer, MaxPool3d and AveragePool3d are three-dimensional maximum pooling and three-dimensional average pooling layers, respectively, and Dense-i is the i-th fully connected layer.

[0064]

[0065] Table 3

[0066] S105 trains the deep neural network based on the input data of the simulation data set and the input data of the measured data set until the loss function converges to a preset threshold, and obtains a honeycomb sandwich plate surface thermal load inversion model.

[0067] S106 inputs the thermal load sensitive feature to be predicted into the honeycomb sandwich plate surface thermal load inversion model to obtain a predicted thermal load parameter.

[0068] In order to better illustrate the scheme of the application, an example is given as follows: Figures 3-5 as shown, comprising the following steps:

[0069] The simulation data set is calculated by Abaqus simulation software, and the experimental data set is measured by actual experiment. As shown in Figure 1The simulation model shown is a simulation model for the problem of honeycomb sandwich plate structure subjected to thermal load. The relevant dimensions of the plate are marked in the figure. A thermal load is applied to the surface. In the simulation, the radiation diameter of the thermal load is set to 20 mm, 25 mm and 30 mm, and the effective radiation power of the thermal load is set to 60 W, 70 W, 80 W, 90 W and 100 W. In the experiment, the honeycomb sandwich plate with the same size as in the simulation is used, the radiation diameter of the thermal load is set to 15 mm, 20 mm and 25 mm, and the effective radiation power of the thermal load is consistent with that in the simulation.

[0070] In the simulation, a 26x26 rectangular array of measurement points is set on the side that receives the thermal load radiation to measure the temperature change of the surface and obtain temperature field data. In the experimental data, 4x4 measurement points are set in the same area, and linear interpolation is performed to obtain 26x26 data format for subsequent data processing to facilitate data fusion training. After obtaining the original temperature field data, the temperature gradient field data and the time rate of change of temperature data are constructed according to the aforementioned method, and data samples with three characteristics are obtained. One sample is taken from each of the simulation data set and the experimental data set for intuitive display.

[0071] The deep transfer learning model is trained on the simulation data set and the experimental data set using the aforementioned method. The samples of the simulation data set and the experimental data set are each labeled with their own domain label, and then the two data sets are randomly mixed together. In terms of data set division, the overall data set is randomly divided, with 70% of the data set as the training data set and 30% of the data set as the test set. After parameter tuning by grid search method, the best learning rate is 0.001 and the best batch size is 16. Under the best hyperparameter configuration, the model is trained for 200 rounds, and the training loss curve is as shown in Figure 5 The loss curve normally decreases.

[0072] To compare the effect of introducing the field self-adaption transfer learning method, the aforementioned ConvLSTM-based predictor in the neural network model is additionally trained in the simulation data set alone, that is, as a control group model without transfer learning, and the learning rate and batch size and other hyperparameters are consistent with the aforementioned neural network model. The models trained with and without transfer learning are tested in the test set part of the experimental data set, and the relative errors of the predicted diameter and the predicted power are shown in Table 4. Without reference to transfer learning, the model trained in a large amount of simulation data set indeed has a large relative error on the experimental data set, mainly due to the data deviation between the two data sets. After introducing transfer learning, the relative errors of the model in predicting the diameter and the power are greatly reduced, the relative error of predicting the diameter is 9.72%, and the relative error of predicting the power is 3.50%. Randomly select 10 samples for prediction, and the corresponding prediction results are shown in Table 5. It can be seen that the deep transfer learning network model proposed in the application is effective, and the prediction error is small, which has practical application value.

[0073]

[0074] Table 4

[0075]

[0076] Table 5

[0077] Example two

[0078] A honeycomb sandwich panel surface thermal load inversion system for inverting thermal load parameters outside a honeycomb sandwich panel is realized by electronic device hardware with a central processing unit, which can be realized by a personal computer, a smart terminal, a local area network, a server, etc. In this embodiment, please refer to Figure 6 , which includes a simulation module 61, a measured data acquisition module 62, an input data extraction module 63, a deep neural network construction module 64, a training module 65, and a prediction module 66.

[0079] The simulation module 61 is used to establish a finite element model for the panel temperature field data based on the varying need to invert the working condition parameters, and obtain the simulation result data set, wherein the need to invert the working condition parameters is the radiation diameter of the thermal load and the radiation power of the thermal load, and the panel temperature field data is the temperature data of a plurality of preset measurement points on the honeycomb sandwich panel varying with time;

[0080] The measured data acquisition module 62 is used to acquire the measured data set obtained by actual experiment, wherein the actual experiment adopts the same need to invert the working condition parameters as the simulation result data set, and the corresponding panel temperature field data measured is taken as the measured data set;

[0081] The input data extraction module 63 is configured to extract the features sensitive to the thermal load from the measured data set and the simulation result data set as input data of the deep neural network.

[0082] The deep neural network construction module 64 is configured to construct a deep neural network for inverting the working condition parameters to be inverted by using a domain adaptive transfer learning framework, taking the simulation data set as a source domain of transfer learning and the measured data set as a target domain of transfer learning, wherein the domain adaptive transfer learning framework comprises a feature extractor, a domain discriminator and a predictor, and the architecture of the predictor is a ConvLSTM-based network.

[0083] The training module 65 is configured to train the deep neural network based on the input data of the simulation data set and the input data of the measured data set until a loss function converges to a preset threshold, so as to obtain a honeycomb sandwich panel surface thermal load inversion model.

[0084] The prediction module 66 is configured to input the features sensitive to the thermal load to be predicted into the honeycomb sandwich panel surface thermal load inversion model, so as to obtain predicted thermal load parameters.

[0085] In the embodiment, the input data extraction module 63 is specifically configured to:

[0086] The temperature field data of the panel is respectively differentiated to obtain temperature gradient field data T grad and temperature change rate field data T vary , wherein the formula is:

[0087] wherein Δx is the step length of the temperature data along the length direction, Δy is the step length of the temperature data along the width direction, Δt is the time step length of the temperature data sampling, i and j respectively represent the data point serial numbers of the temperature data along the x and y directions of the panel, the value range of i is an integer from 1 to 26, the value range of j is an integer from 1 to 26, and the undefined values at the boundaries are replaced by zero values;

[0088] k represents the kth time segment of the sensor data, and the value range of k is an integer from 1 to 9.

[0089] The panel temperature field data T, the temperature gradient field data T grad and the temperature change rate field data T vary are constructed into features sensitive to the thermal load, so as to obtain a multi-dimensional array (S, C, W, H) corresponding to the panel temperature field data, wherein S is the time dimension, C is the feature dimension sensitive to the thermal load, W is the width dimension, and H is the height dimension, and when multiple input data are simultaneously input into the deep neural network, the multi-dimensional array is (B, S, C, W, H), and B represents the batch dimension.

[0090] In the embodiment, in the field adaptive transfer learning framework, during the forward propagation process of transfer learning, the input data first enters the feature extractor for feature extraction, and the extracted features are divided into two paths, one path enters the domain discriminator, and the domain discriminator outputs whether the input data comes from the source domain or the target domain, and the other path enters the ConvLSTM-based predictor, and the predictor infers the working condition parameters to be inverted according to the extracted features as the output. During the error back propagation process, the loss of the domain predicted by the domain discriminator is the cross-entropy loss, and after the error back propagation to the domain discriminator, the gradient inversion layer will propagate the gradient to the feature extractor as the opposite number.

[0091] In the embodiment, the feature extractor uses a two-dimensional convolutional neural network, and the time dimension and the feature dimension of the input data are expanded to the same dimension in the expansion layer, including two two-dimensional convolutional layers and an activation function layer arranged between the two two-dimensional convolutional layers. During initialization, the middle elements of the convolution kernel of the feature extractor are set to 1, and the other elements are set to 0, so that the input and output of the feature extractor are the same in the initial state. Each two-dimensional convolutional layer divides the input channel into 27 groups, each group corresponds to one convolution kernel, and the 27 input channels and the 27 output channels are connected one by one.

[0092] In the embodiment, the domain discriminator has the following structure: a batch normalization layer, three three-dimensional convolutional layers, an activation function layer arranged after each three-dimensional convolutional layer, a full connection layer and a Softmax layer. The Softmax layer is responsible for mapping the result to a valid probability value, and the convolution kernel size of the three-dimensional convolutional layer is (1, 3, 3).

[0093] In the embodiment, the network structure of the predictor includes: a batch normalization layer, four ConvLSTM layers, an activation function layer arranged after each ConvLSTM layer, a Dropout layer, a three-dimensional convolutional layer, a three-dimensional maximum pooling layer, a three-dimensional average pooling layer, a Dropout layer and two full connection layers. The Dropout layer randomly deactivates neurons.

[0094] The implementation process method and effect of the surface thermal load inversion system of the honeycomb sandwich panel in the embodiment are the same as those of the surface thermal load inversion method of the honeycomb sandwich panel described in embodiment one, and will not be repeated here.

[0095] Embodiment three

[0096] The application relates to a computer-readable storage medium, and the storage medium stores instructions. When the instructions are executed, the surface thermal load inversion method of the honeycomb sandwich panel in embodiment one is executed. The implementation process method and effect thereof are the same as those of the surface thermal load inversion method of the honeycomb sandwich panel described in embodiment one, and will not be repeated here.

[0097] It is to be understood that the terminology "including", "comprising", or any other variation thereof, is intended to cover a non-exclusive inclusion such that process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0098] The preferred embodiments of the present application are described above with the specific language and examples. However, the patent protection scope of the present application is not limited to the preferred embodiments, and any equivalent structure or equivalent process transformation based on the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for inverting the surface thermal load of a honeycomb sandwich panel, characterized in that: Parameters used to invert the external thermal loads of honeycomb sandwich panels, including: Based on the changing operating parameters that need to be inverted, a finite element model for the panel temperature field data is established using a simulation program to obtain a simulation result data set, wherein the operating parameters that need to be inverted are the radiation diameter and the radiation power of the heat load, and the panel temperature field data are temperature data of a plurality of preset measuring points on the honeycomb sandwich panel that change with time; Conducting an actual experiment using the same operating condition parameters that need to be inverted as the simulation result data set, and measuring the corresponding plate temperature field data as a measured data set; Extracting features sensitive to thermal loads from the measured data set and the simulation result data set based on the panel temperature field data as input data for a deep neural network; A domain-adaptive transfer learning framework is used to construct a deep neural network for inverting the operating parameters that need to be inverted, with the input data of the simulation dataset used as the source domain of the transfer learning, and the input data of the measured dataset used as the target domain of the transfer learning. The domain-adaptive transfer learning framework includes: a feature extractor, a domain discriminator, and a predictor, and the architecture of the predictor is a ConvLSTM-based network; The deep neural network is trained based on the input data of the simulation data set and the input data of the measured data set until the loss function converges to a preset threshold, thereby obtaining an inversion model of the surface thermal load of the honeycomb sandwich panel; The characteristics sensitive to the thermal load to be predicted are input into the surface thermal load inversion model of the honeycomb sandwich panel to obtain the predicted thermal load parameters.

2. A method for inverting the surface thermal load of a honeycomb sandwich panel according to claim 1, characterized in that: The feature sensitive to thermal load is extracted from the plate temperature field data as the input variable of the neural network, specifically: The gradient derivative and time derivative of the plate temperature field data are calculated respectively to obtain the temperature gradient field data T grad and temperature change rate field data T vary , where the formula is: Wherein, Δx is the step length of the temperature data along the length direction, Δy is the step length of the temperature data along the width direction, Δt is the time step of the temperature data sampling, i and j represent the data point numbers of the temperature data along the x and y directions of the plate, respectively. The value range of i is an integer from 1 to 26, and the value range of j is an integer from 1 to 26. Undefined values ​​at the boundaries are filled with zero values. k represents the kth time segment of the sensor data, and its value ranges from 1 to 9; The plate temperature field data T and the temperature gradient field data T grad and the temperature change rate field data T vary Construct the heat-sensitive features to obtain a multidimensional array (S, C, W, H) corresponding to the panel temperature field data, where S is the time dimension, C is the heat-sensitive feature dimension, W is the width dimension, and H is the height dimension. When multiple input data are simultaneously input into the deep neural network, the multidimensional array is (B, S, C, W, H), where B represents batch dimension.

3. The method for inverting the surface thermal load of a honeycomb sandwich panel according to claim 1, characterized in that: In the domain adaptive transfer learning framework, during the forward propagation of the transfer learning, the input data first enters the feature extractor for feature extraction. The extracted features are divided into two paths. One path enters the domain discriminator, and the domain discriminator outputs whether the input data comes from the source domain or the target domain. The other path enters the ConvLSTM-based predictor, and the predictor infers the operating parameters that need to be inverted based on the extracted features. As output, during the error back propagation process, the domain loss predicted by the domain discriminator is the cross entropy loss. After this part of the error is back-propagated to the domain discriminator, it will pass through the gradient reversal layer. The gradient reversal layer will convert the gradient into an inverse number and propagate it to the feature extractor.

4. The method for inverting the surface thermal load of a honeycomb sandwich panel according to claim 1, characterized in that: The feature extractor uses a two-dimensional convolutional neural network, and expands the time dimension and feature dimension of the input data into the same dimension in the expansion layer. It includes two two-dimensional convolutional layers and an activation function layer arranged between the two two-dimensional convolutional layers. During initialization, the middle element of the convolution kernel of the feature extractor is set to 1, and the other elements are set to 0, so that in the initial state, the input and output of the feature extractor are the same. Each two-dimensional convolutional layer divides the input channels into 27 groups, each group corresponds to 1 convolution kernel, and the 27 input channels and 27 output channels are connected one to one.

5. The method for inverting the surface thermal load of a honeycomb sandwich panel according to claim 1, characterized in that: The domain discriminator structure is: a batch normalization layer, three three-dimensional convolution layers, an activation function layer, a fully connected layer and a Softmax layer arranged after each three-dimensional convolution layer, wherein the Softmax layer is responsible for mapping the results into valid probability values, and the convolution kernel size of the three-dimensional convolution layer is (1,3,3).

6. The method for inverting the surface thermal load of a honeycomb sandwich panel according to claim 1, characterized in that: The network structure of the predictor includes: a batch normalization layer, four ConvLSTM layers, an activation function layer arranged after each ConvLSTM layer, a Dropout layer, a three-dimensional convolution layer, a three-dimensional maximum pooling layer, a three-dimensional average pooling layer, a Dropout layer and two fully connected layers, and the Dropout layer randomly inactivates neurons.

7. A honeycomb sandwich panel surface thermal load inversion system, characterized in that: Parameters used to invert the external thermal loads of honeycomb sandwich panels, including: a simulation module for establishing a finite element model for the panel temperature field data using a simulation program based on the changing operating parameters that require inversion, and obtaining a simulation result data set, wherein the operating parameters that require inversion are the radiation diameter and the radiation power of the heat load, and the panel temperature field data are temperature data of a plurality of preset measuring points on the honeycomb sandwich panel that vary with time; a measured data acquisition module, configured to acquire a measured data set obtained by actual experimentation, wherein the actual experiment is conducted using the same operating condition parameters that need to be inverted as the simulation result data set, and the corresponding plate temperature field data is measured and obtained as the measured data set; An input data extraction module, configured to extract features sensitive to thermal load from the measured data set and the simulation result data set as input data for a deep neural network; A deep neural network construction module is used to construct a deep neural network for inverting the operating parameters that need to be inverted using a domain-adaptive transfer learning framework, using the simulation dataset as a source domain for transfer learning and the measured dataset as a target domain for transfer learning, wherein the domain-adaptive transfer learning framework includes: a feature extractor, a domain discriminator, and a predictor, and the architecture of the predictor is a ConvLSTM-based network; A training module, configured to train the deep neural network based on input data of the simulation data set and input data of the measured data set until the loss function converges to a preset threshold, thereby obtaining an inversion model of the surface thermal load of the honeycomb sandwich panel; The prediction module is used to input the characteristics sensitive to the thermal load that need to be predicted into the surface thermal load inversion model of the honeycomb sandwich panel to obtain the predicted thermal load parameters.

8. The honeycomb sandwich panel surface thermal load inversion system according to claim 7, characterized in that: The input data extraction module is specifically used to: The gradient derivative and time derivative of the plate temperature field data are calculated respectively to obtain the temperature gradient field data T grad and temperature change rate field data T vary , where the formula is: Wherein, Δx is the step length of the temperature data along the length direction, Δy is the step length of the temperature data along the width direction, Δt is the time step of the temperature data sampling, i and j represent the data point numbers of the temperature data along the x and y directions of the plate, respectively. The value range of i is an integer from 1 to 26, and the value range of j is an integer from 1 to 26. Undefined values ​​at the boundaries are filled with zero values. k represents the kth time segment of the sensor data, and its value ranges from 1 to 9; The plate temperature field data T and the temperature gradient field data T grad and the temperature change rate field data T vary Construct the heat-sensitive features to obtain a multidimensional array (S, C, W, H) corresponding to the panel temperature field data, where S is the time dimension, C is the heat-sensitive feature dimension, W is the width dimension, and H is the height dimension. When multiple input data are simultaneously input into the deep neural network, the multidimensional array is (B, S, C, W, H), where B represents batch dimension.

9. The honeycomb sandwich panel surface thermal load inversion system according to claim 7, characterized in that: The feature extractor uses a two-dimensional convolutional neural network, and expands the time dimension and feature dimension of the input data into the same dimension in the expansion layer. It includes two two-dimensional convolutional layers and an activation function layer arranged between the two two-dimensional convolutional layers. During initialization, the middle element of the convolution kernel of the feature extractor is set to 1, and the other elements are set to 0, so that in the initial state, the input and output of the feature extractor are the same. Each two-dimensional convolutional layer divides the input channels into 27 groups, each group corresponds to one convolution kernel, and the 27 input channels and 27 output channels are connected one to one; The domain discriminator structure is as follows: a batch normalization layer, three three-dimensional convolutional layers, an activation function layer, a fully connected layer, and a softmax layer after each three-dimensional convolutional layer. The softmax layer is responsible for mapping the results into valid probability values. The convolution kernel size of the three-dimensional convolutional layer is (1, 3, 3). The network structure of the predictor includes: a batch normalization layer, four ConvLSTM layers, an activation function layer arranged after each ConvLSTM layer, a Dropout layer, a three-dimensional convolution layer, a three-dimensional maximum pooling layer, a three-dimensional average pooling layer, a Dropout layer and two fully connected layers, and the Dropout layer randomly inactivates neurons.

10. A computer-readable storage medium, characterized in that: The storage medium stores instructions, which, when executed, execute a method for inverting the surface thermal load of a honeycomb sandwich panel according to any one of claims 1 to 6.

Images