A method and system for identifying and predicting surface heat flux of transpiration cooling

Abstract

The application discloses a kind of perspiration cooling surface heat flux identification prediction method and system.For the thermal lag caused by internal micropore seepage, gas-liquid phase change of perspiration cooling thermal protection system, leading to the problem that there is feedback delay in the temperature inversion of inner wall temperature to outer surface heat flux, the application is by arranging sensor array in the thickness direction inside porous material, collect internal discrete temperature data at continuous multiple time steps, construct space-time history temperature evolution sequence;Based on the sequence, the nonlinear physical mapping law between internal temperature change and outer surface heat flux is learned using deep learning network.The method can realize the online reconstruction of the current heat flux of the outer surface, and predict the heat flux evolution trend in the future time window using the historical temperature evolution information, which provides a data basis for the optimization design and online feedforward control of the cooling working medium mass flow rate, and has engineering application value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of aerospace thermal protection technology and artificial intelligence, and specifically relates to a method and system for identifying and predicting the heat flux of a sweating cooling surface. Background Technology

[0002] High-speed aircraft experience high surface temperatures during high-speed flight or atmospheric reentry, which places high demands on the reliability of thermal protection systems. Accurate and timely acquisition of surface heat flux information is fundamental to the fine-tuning and optimized design of the cooling fluid mass flow rate in active thermal protection systems (such as sweating cooling systems).

[0003] Current engineering practices typically embed temperature sensors within the heat shield layer to invert the heat flux to the outer surface using the inner wall temperature. However, systems like perspiration cooling involve complex microchannel percolation heat transfer and phase change processes, resulting in significant time delays in heat conduction. The temperature response lags severely behind changes in the heat flux to the outer surface, a phenomenon known as "thermal hysteresis." Existing classical theoretical methods (such as the Fay-Riddell formula and the reference enthalpy method) struggle to handle complex geometries and time-varying flow field conditions; traditional inverse solving algorithms (such as the conjugate gradient method) are unable to overcome the time delay and respond promptly to time-varying thermal loads. Existing deep learning-based aerodynamic thermal prediction methods (such as U-Net) primarily focus on static input features and lack specificity in addressing dynamic thermal hysteresis effects, making it difficult to predict subsequent heat flux evolution trends. Insufficient predictive capability can lead to lags in the dynamic adjustment of the cooling fluid mass flow rate, potentially causing localized thermal runaway or even structural failure. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method and system for identifying and predicting the heat flux of a sweating cooling surface, which mainly solves at least one of the technical problems of traditional methods in dealing with thermal hysteresis effects, such as insufficient response timeliness, large computational overhead, and difficulty in providing feedforward control data for the cooling working fluid supply system.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying and predicting the heat flux of a sweating cooling surface includes the following steps: Step 1: Inside the porous medium material of the sweating cooling system, a multi-layer temperature sensor array is arranged along the thickness direction; discrete internal temperature data of multiple consecutive time steps are collected at a preset sampling frequency, and constructed into a four-dimensional temperature feature input tensor with temporal and spatial topological structure, thereby obtaining the multi-dimensional spatiotemporal temperature feature sequence of the sweating cooling system; in order to realize the prediction of heat flux fluctuation trend within a future preset time window, the time span corresponding to the number of consecutively collected time steps is greater than or equal to the inherent thermal response hysteresis time of the sweating cooling system. Step 2: Construct a deep neural network based on a symmetric encoder-decoder architecture. Embed a convolutional long short-term memory network module in the bottleneck layer of the network to extract the dynamic evolution features of the time series. Configure compression and activation channel attention modules in series at the skip connections corresponding to the encoder and decoder to adaptively calibrate the channel weights in the spliced ​​features across time frames. This results in a spatiotemporal feature extraction and prediction network model with a channel attention mechanism. Step 3: Train the network model based on historical operating data; Step 4: Use the trained network model to perform online identification and multi-step prediction.

[0006] In one embodiment of the present invention, the dimension of the four-dimensional temperature feature input tensor mentioned in step 1 is represented as follows: t × W × H × n ,in, t This represents the number of time steps for continuous data collection. W and H These represent the horizontal and vertical spatial sampling node dimensions of each layer of the temperature sensor array, respectively. n The number of layers in the temperature sensor array arranged along the thickness direction.

[0007] In one embodiment of the present invention, the bottleneck layer embedded with a convolutional long short-term memory network module in step 2 has the following specific operating mechanism: Receive the high-dimensional spatial feature set output by the encoder; The matrix multiplication operation inside the traditional long short-term memory unit is replaced by two-dimensional convolution operation, which preserves the spatial topological structure information of the input feature set during the transmission of hidden state information in the time series.

[0008] In one embodiment of the present invention, the compression and excitation channel attention module described in step 2 has an operating mechanism comprising three sub-steps: compression, excitation, and recalibration. The compression employs a global average pooling operation to reduce the dimensionality of the cross-time frame feature map (i.e., the feature map formed by splicing the corresponding level jump features extracted by the encoder in the channel dimension according to the time frame order) at the jump connection along the spatial dimension, aggregates global spatial information, and generates channel-level statistical feature vectors. The stimulus inputs the channel-level statistical feature vector into a bottleneck structure multilayer perceptron containing a dimensionality reduction layer and an dimensionality increase layer. After mapping by a nonlinear activation function, a normalized channel weight vector with values ​​between 0 and 1 is output. The recalibration involves performing a channel-by-channel element-wise multiplication operation between the channel weight vector and the original input cross-time frame feature map, which is used to adjust the cross-time frame feature map according to the channel weights.

[0009] In one embodiment of the present invention, step 3 involves establishing a sweating cooling aerodynamic thermal sample dataset covering various flight profile parameters, dividing it into a training set, a validation set, and a test set; using the training set to iteratively train the network model constructed in step 2, optimizing the network parameters until the loss function converges, and obtaining a trained prediction model. Each historical aerodynamic thermal sample in the sample dataset generates a corresponding four-dimensional temperature feature input tensor using the same acquisition and construction method as in step 1, and uses time-series data of external surface heat flux under the same flight conditions as the supervision label; the four-dimensional temperature feature input tensor during the training phase is used to learn the mapping relationship between the internal temperature hysteresis response and the external surface heat flux, while the four-dimensional temperature feature input tensor constructed in real-time according to step 1 during the online phase serves as the input to the trained network model.

[0010] In one embodiment of the present invention, the flight profile parameters include Mach number, flight altitude, and angle of attack.

[0011] In one embodiment of the present invention, the sample dataset described in step 3 covers a range of flight altitudes from 40 km to 60 km, Mach numbers from 5 to 20, and angles of attack from 0° to 20°, and the Adam optimizer is used during the network model training process.

[0012] In one embodiment of the present invention, step 4 involves embedding the four-dimensional temperature feature input tensor, which is collected online under actual flight conditions, into the trained network model. The model then synchronously outputs the online identified heat flux of the outer surface of the sweating cooling system at the current moment, as well as the evolution prediction curve of time and heat flux within a preset future time window.

[0013] The present invention also provides a system for identifying and predicting the heat flux of a sweating cooling surface, used to implement the aforementioned method for identifying and predicting the heat flux of a sweating cooling surface, comprising: The data acquisition module is used to acquire transient temperature sequence data inside the porous media material of the sweating cooling system online; The preprocessing module is used to reconstruct the transient temperature sequence data into a four-dimensional temperature feature input tensor format with spatiotemporal dimensions; The core computing module is equipped with a trained channel attention temporal U-Net model to extract the nonlinear spatiotemporal mapping relationship between internal temperature and external surface heat flux. The output prediction module is used to simultaneously output the online identified heat flux value of the outer surface of the sweating cooling system and the predicted heat flux curve within a preset time window.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1) Implement an active prediction paradigm to support feedforward optimization design.

[0015] This method transforms passive inversion into active prediction, simultaneously identifying the current heat flux and predicting its fluctuation trend over the next 10 to 30 seconds. This helps reduce the impact of thermal hysteresis and provides a certain system response time for the optimized design of the cooling medium mass flow rate supply strategy and online feedforward control in the sweating cooling system.

[0016] 2) Improve computational efficiency and prediction accuracy.

[0017] Unlike approaches that place the attention mechanism inside the encoder, this invention deploys it lightweightly at the skip connections. Test results show that the root mean square error (RMSE) is as low as 0.0189 MW / m. 2 (84.7% lower than the baseline model), coefficient of determination (R²) 2 Assuming a value of 0.9984, the model parameter count is 14.5 × 10⁻⁶. 6 The required computation is 25.94 GMac, and a single inference takes about 5.04 seconds.

[0018] 3) Robustness and noise resistance.

[0019] Because the Conv-LSTM module has timing smoothing capabilities, and with the filtering of the channel attention mechanism, this invention can reduce the impact of noise transmission even under set high-frequency noise interference conditions, with a maximum relative error of 1.6%, meeting the engineering testing requirements of the preset flight environment. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the three-dimensional physical model structure and method of the sweating cooling system of the present invention.

[0021] Figure 2 This is the overall architecture diagram of the Channel Attention Temporal U-Net (CATU), which is the core of this invention.

[0022] Figure 3 This is a schematic diagram of the internal workflow of the Conv-LSTM timing module in this invention.

[0023] Figure 4 This is a schematic diagram of the embedded channel attention mechanism (SE Block) structure in this invention.

[0024] Figure 5 The results are model identification and prediction in the embodiments of the present invention.

[0025] Figure 6 This refers to the relative error of model identification and prediction in the embodiments of the present invention. Detailed Implementation

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0027] This invention provides a method for identifying and predicting the heat flux of a sweating cooling surface, mainly including the following steps: Step 1: Obtain the multidimensional spatiotemporal temperature feature sequence of the sweating cooling system.

[0028] The three-dimensional physical model of the sweating cooling system targeted by this invention is as follows: Figure 1 As shown, the sweating cooling system utilizes a porous material to coat the liquid storage layer. To achieve data acquisition, a multi-layer temperature sensor array is arranged along the thickness direction as a sensor matrix inside the porous medium material in a predetermined region of the sweating cooling system. Subsequently, discrete temperature data inside the porous material at multiple consecutive time steps are collected at a predetermined sampling frequency, and these data are constructed into a four-dimensional temperature feature input tensor, the dimensions of which can be represented as... t × W × H × n .in, t This represents the number of time steps for continuous data collection. W and H These represent the horizontal and vertical spatial sampling node dimensions of each layer of the temperature sensor array, respectively. n This represents the number of layers in the temperature sensor array arranged along the thickness direction. Thus, this tensor simultaneously possesses both temporal and spatial topological structure, serving as the input feature of the network model of this invention.

[0029] To predict the heat flux fluctuation trend over the next 10 to 30 seconds, the time step... t The product of the sampling time interval and the sampling time interval (i.e., the total physical duration covered by the input sequence) is preferably greater than or equal to the inherent thermal response hysteresis time of the sweating cooling system. In a specific embodiment of the invention, a long short-term memory network is used to extrapolate historical trends, and an initial 80 seconds is selected as the input duration of the historical temperature sequence (i.e., t × sampling interval = 80 s), thereby enabling the prediction of the heat flux evolution trend in the subsequent 20 s (i.e., the 80s to 100s window), with a maximum prediction error of 2.5%.

[0030] Step 2: Construct a spatiotemporal feature extraction and prediction network with channel attention mechanism.

[0031] To address the issue of feedback delays in conventional temperature inversion methods caused by thermal hysteresis effects resulting from complex physical processes such as internal micropore seepage and gas-liquid phase transitions in existing active thermal protection systems, this invention proposes a CATU deep learning architecture embedding a channel attention mechanism, such as... Figure 2As shown, the deep neural network is based on a symmetric encoder-decoder architecture, with U-Net as the backbone, and extracts the spatial features of each single frame of data during the downsampling stage. The distinguishing feature of the network structure in this invention is that a Convolutional Long Short-Term Memory (Conv-LSTM) network module is embedded in the bottleneck layer of the network. This module is used to extract the dynamic evolution features of the time series and learn long-term temporal evolution patterns. Furthermore, a compression and excitation channel attention module (SE Block) is configured in series at the skip connections corresponding to the encoder and decoder. This module is used to adaptively calibrate the channel weights in the spliced ​​features across time frames, rather than directly changing the time step order.

[0032] Therefore, the features extracted during the downsampling stage are aggregated and fed into the Conv-LSTM module. Since all internal calculations involve convolution operations, historical temporal dependencies are captured while maintaining spatial topological information. Based on this, the SE attention module uses global average pooling to compress the feature map and generate channel-level statistics. A multilayer perceptron dynamically adjusts the weights of each channel to achieve adaptive feature fusion and redundancy suppression. Finally, deconvolution upsampling is used to gradually restore the resolution, outputting the heat flux prediction result.

[0033] For example, for Figure 2 The input shown is [ H , W , 3]× t The feature sequence, in the encoder stage, undergoes successive convolution (CONV) and downsampling operations, with its channel number expanding progressively with network depth; after two convolutional layers, the dimension becomes [ H , W ,32]× t Then adjust downwards in sequence to [ H , W , 64]× t 、[ H , W , 128]× t And the lowest level [ H , W , 256]× t At the lowest bottleneck level of the network, the feature sequence is fed into the LSTM module (i.e., Conv-LSTM) for long-range temporal feature extraction. In the decoder stage, features are upsampled layer by layer, and before each feature concatenation (CONCAT), the corresponding level skip features extracted by the encoder (such as u1, u2, u3, u4) are processed. k ,u n The data is fed into the SE module for adaptive weight calibration along the channel dimension. The concatenated features are then subjected to two convolutional layers to gradually restore spatial resolution, and finally mapped to a dimension of […]. H , W ,1]×t Output of heat flux prediction results.

[0034] Furthermore, such as Figure 3 As shown, the specific operating mechanism of the Convolutional Long Short-Term Memory (Conv-LSTM) network module of this invention is as follows: Receive the high-dimensional spatial feature set output by the encoder; The matrix multiplication operation inside the traditional long short-term memory unit is replaced by two-dimensional convolution operation, which preserves the spatial topological structure information of the input feature set during the transmission of hidden state information in the time series.

[0035] At each time step, the convolutional long short-term memory network module simultaneously receives the spatial feature input of the current time. x t And the hidden state of the previous moment. h t-1 and cell state c t-1 The module contains three control gates: input gates (which control the input of new information). i t Forget Gate (Regulating the Forgetting of Historical Information) f t ) and output gate ( o t Both methods employ two-dimensional convolution operations to process input information, followed by feature mapping and state selection using sigmoid and tanh nonlinear activation functions. This operational mechanism enables the updating of the cell state at the current time step. c t and hidden state h t In this way, not only the temporal dependencies of the input features are preserved, but also their spatial two-dimensional topological structure is preserved.

[0036] Furthermore, such as Figure 4 As shown, the compression and excitation channel attention module (SE Block) is used for adaptive calibration of cross-frame feature weights. Its operation mechanism includes three sub-steps: compression, excitation, and recalibration. Compression steps: Global average pooling is used to reduce the dimensionality of the cross-time frame feature maps spliced ​​at the skip connections along the spatial dimension, aggregate global spatial information, and generate channel-level statistical feature vectors. Activation step: The channel-level statistical feature vector is input into a bottleneck structure multilayer perceptron containing a dimensionality reduction layer and an dimensionality increase layer. After mapping by a nonlinear activation function, a normalized channel weight vector with values ​​between 0 and 1 is output. Recalibration step: The channel weight vector is multiplied element-wise with the original input cross-time frame feature map, which is used to adjust the cross-time frame feature map according to the channel weight.

[0037] At the skip connection, feature maps from different time frames in the encoder stage (e.g., u1, u2) k ,u n First, features are concatenated along the channel dimension. The concatenated high-dimensional features then enter the compression and activation module (SE Block). In the compression stage, the module compresses the spatial information into a one-dimensional channel description vector through global average pooling. In the activation stage, this vector is processed by a multilayer perceptron network consisting of dimensionality reduction and expansion layers, and then processed by an activation function to generate a dynamic channel weight distribution (i.e., the small square weight array in the figure). In the recalibration stage, the learned channel weight distribution is multiplied channel-by-channel with the original concatenated feature map, thereby assigning corresponding response weights to the channel features with higher weights. Finally, a 1×1 convolution is used to adjust the channel dimension to adapt to the input requirements of the subsequent decoder.

[0038] Step 3: Train the network model based on historical operating data.

[0039] A sweating cooling aerodynamic thermal sample dataset covering various flight profile parameters (including Mach number, flight altitude, and angle of attack) was established and divided into training, validation, and test sets, covering flight altitudes from 40 km to 60 km, Mach numbers from 5 to 20, and angles of attack from 0° to 20°. Based on the Adam optimizer, the network model constructed in step 2 was iteratively trained using the training set to optimize the network parameters until the loss function converged, thus obtaining the trained prediction model.

[0040] Step 4: Online identification and multi-step prediction.

[0041] The four-dimensional temperature features collected online under actual flight conditions are input into the trained prediction model. The model simultaneously outputs the online identified heat flux of the outer surface of the sweating cooling system at the current moment, as well as the evolution prediction curve of time and heat flux within a preset time window in the future.

[0042] In one specific embodiment of the present invention, multi-step prediction and identification performance verification was performed. A sweating cooling process was implemented in a predetermined area of ​​the high-speed aircraft, establishing a three-dimensional cuboid computational domain under the assumption of a local plane. The simulated temperature was... T cThe cooling medium is injected into the porous material from the bottom liquid reservoir. The convective heat transfer between the internal fluid and solid phases is modeled using a local thermal non-equilibrium model, with initial values ​​set so that the porous material and cooling medium maintain the same initial temperature and pressure. During system operation, temperature array data (each layer with a resolution of 8×8) is collected along the thickness direction. 200 flight profile data points from altitudes of 40km to 60km, Mach numbers of 5 to 20, and angles of attack of 0° to 20° are used as training data, with the training, validation, and test sets configured in a 6:2:2 ratio. The trained CATU model is then applied to the test set for analysis to evaluate its feedforward predictive capability. Figure 5 and Figure 6 As shown, for a 100-second ascent or descent profile, the model identifies and extrapolates the heat flux trend over a future time window from 80 to 100 seconds based on the historical inner wall temperature sequence from 0 to 80 seconds. Test results show that, under varying flight conditions, the average error within a 20-second prediction window is approximately 1.8%, with a maximum prediction error of 2.5%.

[0043] In summary, the method of this invention can reconstruct the current heat flux and predict the heat flux evolution trend within future time windows, providing a response time margin and feedforward data basis for the optimized design and online control of the cooling fluid mass flow rate supply, and has engineering application value.

[0044] Based on this, the present invention also provides a corresponding sweating cooling surface heat flux identification and prediction system, including: a data acquisition module, a preprocessing module, a core calculation module, and an output prediction module. The data acquisition module includes a temperature sensor array arranged inside the porous material, used to acquire transient temperature sequence data inside the porous medium material of the sweating cooling system online. The preprocessing module is a software processing unit that can be integrated into the temperature sensor, a separate module, or integrated into the processor, used to reconstruct the acquired transient temperature sequence data into a four-dimensional temperature feature input tensor format with the spatiotemporal dimension described in this invention. The core calculation module is a unit that deploys a trained channel attention temporal U-Net model, used to extract the nonlinear spatiotemporal mapping relationship between the internal temperature and the external surface heat flux. It is the main calculation unit of the present invention and can be a separate processor or integrated with the preprocessing module and the output prediction module. The output prediction module is used to simultaneously output the online identified heat flux value of the external surface of the sweating cooling system and the predicted heat flux curve within a future preset time window.

Claims

1. A method for identifying and predicting the heat flux of a sweating cooling surface, characterized in that, Includes the following steps: Step 1: Obtain the spatiotemporal temperature evolution characteristics inside the sweating cooling system: A multi-layer temperature sensor array is arranged along the thickness direction inside the porous medium material of the sweating cooling system; discrete internal temperature data for multiple consecutive time steps are collected at a preset sampling frequency, and these data are constructed into a four-dimensional temperature feature input tensor with temporal and spatial topological structure, thereby obtaining the multi-dimensional spatiotemporal temperature feature sequence of the sweating cooling system; wherein, the physical duration covered by the multiple consecutive time steps is not less than the inherent thermal response hysteresis time of the sweating cooling system; Step 2: Construct a spatiotemporal feature extraction and prediction network model for mapping the relationship between internal temperature and external surface heat flux. The network model includes a spatial feature extraction module for extracting the spatial temperature distribution of a single frame, and a temporal feature extraction module for learning the heat conduction time delay law. A channel attention mechanism is introduced into the network structure to adaptively calibrate the channel weights in the spliced ​​features across time frames in order to focus on historical temperature features that are highly correlated with changes in heat flux. Step 3, training the network model based on historical operating condition data: acquire historical aerodynamic thermal sample data covering the sweating cooling system under various flight conditions, and use the sample data to train the prediction network model so that it learns the physical mapping law between the internal temperature hysteresis response of the porous medium and the heat flux on the outer surface. Step 4: Use the trained network model for online identification and multi-step prediction: Input the multi-dimensional spatiotemporal temperature feature sequence collected online during actual flight into the trained network model. After model processing, the online identification heat flux of the outer surface of the sweating cooling system at the current moment is output synchronously, and the evolution trend of the outer surface heat flux within the preset time window is predicted.

2. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 1, characterized in that, The multidimensional spatiotemporal temperature feature sequence mentioned in step 1 is specifically constructed as a four-dimensional temperature feature input tensor, whose dimensions are represented as follows: t × W × H × n ,in, t This represents the number of time steps for continuous data collection. W and H These represent the horizontal and vertical spatial sampling node dimensions of each layer of the temperature sensor array, respectively. n The number of layers in the temperature sensor array arranged along the thickness direction.

3. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 1, characterized in that, The spatiotemporal feature extraction and prediction network model described in step 2 adopts a symmetric encoder-decoder architecture, taking the four-dimensional temperature feature input tensor as input and outputting the heat flux prediction result. Its temporal feature extraction module is specifically a convolutional long short-term memory network module embedded in the network bottleneck layer, and its operating mechanism is as follows: Receive the high-dimensional spatial feature set output by the encoder; The matrix multiplication operation inside the traditional long short-term memory unit is replaced by two-dimensional convolution operation, which preserves the spatial topological structure information of the input feature set during the transmission of hidden state information in the time series.

4. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 3, characterized in that, The compression and activation channel attention module described in step 2 is configured in series at the skip connections corresponding to the encoder and decoder. Before each feature concatenation in the decoder, the corresponding level skip features extracted by the encoder are first fed into the compression and activation channel attention module for adaptive calibration of channel dimension weights. The concatenated features are then subjected to two layers of convolution to gradually restore spatial resolution, and finally mapped to a dimension of [ H , W , 1]× t Output of heat flux prediction results.

5. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 4, characterized in that, The compression and activation channel attention module described in step 2 operates through three sub-steps: compression, activation, and recalibration. The compression employs a global average pooling operation to reduce the dimensionality of the cross-time frame feature map spliced ​​at the jump connection along the spatial dimension, aggregates global spatial information, and generates a channel-level statistical feature vector. The stimulus inputs the channel-level statistical feature vector into a bottleneck structure multilayer perceptron containing a dimensionality reduction layer and an dimensionality increase layer. After being mapped by a nonlinear activation function, a normalized channel weight vector with values ​​between 0 and 1 is output. The recalibration involves performing a channel-by-channel element-wise multiplication operation between the channel weight vector and the original input cross-time frame feature map, which is used to adjust the cross-time frame feature map according to the channel weights.

6. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 1, characterized in that, Step 3 involves establishing a sweating cooling aerodynamic thermal sample dataset covering various flight profile parameters, dividing it into a training set, a validation set, and a test set. The network model constructed in Step 2 is iteratively trained using the training set, optimizing network parameters until the loss function converges, thus obtaining a trained prediction model. Each historical aerodynamic thermal sample in the dataset generates a corresponding four-dimensional temperature feature input tensor using the same acquisition and construction method as in Step 1, with time-series data of external surface heat flux under the same flight conditions used as a supervision label. During the training phase, the four-dimensional temperature feature input tensor is used to learn the mapping relationship between internal temperature hysteresis response and external surface heat flux. During the online phase, the four-dimensional temperature feature input tensor constructed in real-time according to Step 1 serves as the input to the trained network model.

7. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 6, characterized in that, The flight profile parameters include Mach number, flight altitude, and angle of attack. The sample dataset covers a range of flight altitudes from 40 km to 60 km, Mach numbers from 5 to 20, and angles of attack from 0° to 20°. The Adam optimizer is used during the training of the network model.

8. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 1, characterized in that, In step 4, the four-dimensional temperature feature input tensor collected online under actual flight conditions is input into the trained network model. The model synchronously outputs the online identified heat flux of the outer surface of the sweating cooling system at the current moment, as well as the evolution prediction curve of time and heat flux within a preset time window in the future.

9. The method for identifying and predicting the heat flux of a sweating cooling surface according to claim 1, characterized in that, The temperature sensor array is arranged in an 8×8 dot matrix on each plane; in the convolutional layer of the deep neural network during the encoding stage, a LeakyReLU activation function with a negative slope of 0.1 is used in conjunction with instance normalization to maintain the transfer of spatiotemporal features and improve the stability of model training.

10. A system for identifying and predicting heat flux on a sweating cooling surface, used to implement the method for identifying and predicting heat flux on a sweating cooling surface according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire transient temperature sequence data of the porous media material inside the sweating cooling system online. The preprocessing module is used to reconstruct the transient temperature sequence data into a four-dimensional temperature feature input tensor format with spatiotemporal dimensions; The core computing module is equipped with a trained channel attention temporal U-Net model to extract the nonlinear spatiotemporal mapping relationship between internal temperature and external surface heat flux. The output prediction module is used to simultaneously output the online identified heat flux value of the outer surface of the sweating cooling system and the predicted heat flux curve within a preset time window.

Images