A temperature field reconstruction method based on thermal load inversion guidance

Abstract

The present application relates to a kind of temperature field reconstruction methods based on heat load inversion guide, belong to digital twin field, solve the problem that existing temperature field reconstruction method does not invert heat load information, cannot real-time, safely, reliably predict temperature field data.The method comprises: obtaining the real-time temperature value of sampling point continuous time;The real-time temperature value is input into the heat load inverter of temperature field reconstruction model trained, and heat load information code is obtained;The heat load information code and random noise matrix are input into the generator of temperature field reconstruction model trained, and temperature field reconstruction data are obtained.Heat load information can be inverted, and temperature field data can be safely, reliably and efficiently predicted, and temperature field can be monitored in real time.

Description

Technical Field

[0001] This invention relates to the field of digital twins, and more particularly to a temperature field reconstruction method based on thermal load inversion guidance. Background Technology

[0002] With the development of modern defense science and technology and the application requirements of target infrared radiation characteristics, the analysis of aircraft surface temperature fields has become a key research focus in the defense and military field. Aircraft may endure different loads under different missions and environments, such as atmospheric reentry, high-speed flight, and atmospheric re-entry. Under each load condition, the operating status and performance of the thermal protection system will differ.

[0003] Taking the thermal protection characteristic analysis of aircraft compartments as an example, this analysis mainly relies on finite element model derivation based on physical mechanisms. However, the finite element model is calculated strictly according to physical properties, resulting in a long calculation cycle. Furthermore, different load models require recalculation, leading to very low efficiency and failing to meet the needs of practical engineering problems. Therefore, the research on its surrogate reduced-order model has significant practical significance and application value. This study aims to develop a twin surrogate model for the physical model of aircraft compartment thermal protection, breaking through the limitations of designing and developing reduced-order models for the evolution of the temperature field in aircraft compartment thermal protection, and applying it to predict the thermal protection characteristics of compartments under different loads. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a temperature field reconstruction method based on thermal load inversion to solve the problems of low reliability, security and efficiency of existing temperature field reconstruction methods in generating temperature field data, and the inability to predict the temperature field in real time.

[0005] The objective of this invention is mainly achieved through the following technical solutions:

[0006] This invention provides a temperature field reconstruction method based on thermal load inversion guidance, comprising:

[0007] Step 1: Obtain the real-time temperature values ​​of the sampling points at continuous intervals;

[0008] Step 2: Input the real-time temperature value into the thermal load inversion unit of the trained temperature field reconstruction model to obtain the thermal load information encoding; wherein, the temperature field reconstruction model includes a thermal load inversion unit and a generator;

[0009] Step 3: Input the encoded thermal load information and random noise matrix into the generator of the trained temperature field reconstruction model to obtain temperature field reconstruction data.

[0010] Furthermore, the step of inputting the real-time temperature value into the thermal load inversion unit of the trained temperature field reconstruction model to obtain the thermal load information encoding includes:

[0011] The real-time temperature value is encoded into an encoding matrix representing the current heat load information by the Encoder module of the Transformer of the heat load inverter, through multiple self-attention layers and feedforward fully connected layers.

[0012] Furthermore, the step of inputting the encoded thermal load information and the random noise matrix into the generator of the trained temperature field reconstruction model to obtain temperature field reconstruction data includes:

[0013] The thermal load information encoding matrix and the random noise matrix are fused together by the encoder module of the generator to extract features and obtain the encoder module output value.

[0014] The encoder module output value is passed through the generator's dilated convolution module, and the receptive field is increased to obtain the dilated convolution module output value.

[0015] The output value of the dilated convolution module is used by the decoder module of the generator to restore the original resolution and generate temperature field reconstruction data.

[0016] Furthermore, the thermal load information encoding matrix and the random noise matrix are passed through the encoder module of the generator, including:

[0017] The thermal load information encoding matrix and the random noise matrix are fused by summing the elements at their corresponding positions in the first downsampling module of the encoder module, and then the result is output after convolution operation in the convolution layer with a stride of 1.

[0018] The output value of the first downsampling module is then output after being convolved by a convolutional layer with a stride of 2 and a convolutional layer with a stride of 1 in the second downsampling module of the encoder module.

[0019] The output value of the second downsampling module is sequentially convolved through a convolutional layer with a stride of 2 and two convolutional layers with a stride of 1 in the third downsampling module of the encoder module before being output as the encoder module output value.

[0020] Furthermore, the encoder module output value is passed through the generator's dilated convolution module to increase the receptive field, resulting in the dilated convolution module output value, including:

[0021] The output value of the encoder module is sequentially input into four dilated convolutional layers of the dilated convolutional module with a stride of 1 and dilation rates of 2, 4, 6 and 8, respectively, to obtain the output value of the dilated convolutional module after expanding the receptive field.

[0022] Furthermore, the output value of the dilated convolution module is passed through the decoder module of the generator to restore the original resolution and generate temperature field reconstruction data, including:

[0023] The output value of the dilated convolution module is passed through a convolutional layer with a stride of 1 in the first upsampling module of the decoder, and then concatenated with the output value of the third downsampling module of the encoder module in the feature concatenation module. Finally, it is passed through a convolutional layer with a stride of 1 for convolution operation before being output.

[0024] The output value of the first upsampling module is passed through a deconvolution layer with a stride of 1 / 2 in the second upsampling module of the decoder, and then concatenated with the output value of the second downsampling module of the encoder module in the feature concatenation module. Finally, it is convolved through a convolution layer with a stride of 1 before being output.

[0025] The output value of the second upsampling module is passed through a deconvolution layer with a stride of 1 / 2 in the third upsampling module of the decoder, and then concatenated with the output value of the first downsampling module of the encoder module in the feature concatenation module. Finally, it is passed through a convolution layer with a stride of 1 to perform a convolution operation and output the device temperature field reconstruction data.

[0026] Furthermore, the temperature field reconstruction model is trained using the following method:

[0027] Step 101: Construct a training dataset for the temperature field reconstruction model; wherein, the training dataset includes real-time temperature values ​​at continuous sampling points, a random noise matrix, and the corresponding real temperature field data;

[0028] Step 102: Based on the training set, initially train the temperature field reconstruction model, update the parameters of the temperature field reconstruction model, and obtain the initially trained temperature field reconstruction model; wherein, the temperature field reconstruction model includes a thermal load inverter and a generator;

[0029] Step 103: Train the discriminator based on the training set and the temperature field reconstruction model trained in Step 102, update the discriminator parameters, and obtain the initially trained discriminator;

[0030] Step 104: Based on the training set, simultaneously train the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103 to obtain the trained temperature field reconstruction model.

[0031] Furthermore, the training temperature field reconstruction model includes:

[0032] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model to generate thermal load information data;

[0033] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model, and generate temperature field reconstruction data with the thermal load information data;

[0034] Load the corresponding real temperature field data of the training set as labels;

[0035] The generator for the temperature field reconstruction model is trained using the mean squared error loss function. Through backpropagation and gradient descent optimization algorithms, the generator can continuously adjust its parameters to minimize the loss function. After training, the parameters of the temperature field reconstruction model are saved.

[0036] Furthermore, the trained discriminator includes:

[0037] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data;

[0038] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is used as a false sample input discriminator;

[0039] Load the corresponding real temperature field data of the training set as real samples into the discriminator;

[0040] The parameters of the temperature field reconstruction model trained in step 102 are fixed, and the discriminator is trained using the binary cross-entropy loss function as the loss function. By minimizing the binary cross-entropy loss, the discriminator parameters are updated so that the discriminator can accurately classify the input data into real samples or fake samples. The discriminator parameters are saved after training.

[0041] Furthermore, the simultaneous training of the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103 includes:

[0042] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data;

[0043] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is then used as a fake sample input into the discriminator trained in step 103.

[0044] Load the corresponding real temperature field data of the training set as real samples into the discriminator trained in step 103;

[0045] The temperature field reconstruction model and the discriminator are trained simultaneously using the mean squared error loss function and the binary cross-entropy loss function. The parameters of the temperature field reconstruction model and the discriminator are updated. After training, the parameters of the temperature field reconstruction model are saved to obtain the final temperature field reconstruction model.

[0046] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0047] 1. The technical solution of this invention, when predicting temperature field data, uses a thermal load inversion guide and generator to predict the temperature field of the equipment based on the real-time temperature values ​​sampled by the sensor. This enables real-time monitoring and prediction of the temperature distribution of the equipment, allowing for the early detection of potential temperature anomalies.

[0048] 2. Since changes in thermal load cannot be accurately obtained in real time, the technical solution of this invention can obtain real-time thermal load information of the equipment by inverting the temperature values ​​sampled by the sensor. This provides a key input for temperature field reconstruction, enabling more accurate prediction and analysis of the equipment's temperature distribution.

[0049] 3. The real-time temperature field prediction and thermal load information acquisition of the technical solution of this invention can provide reliable data support for the temperature management and thermal control of spacecraft. By timely monitoring the temperature changes of equipment, corresponding measures can be taken to optimize thermal control strategies, improve equipment efficiency and lifespan, and reduce potential failure risks.

[0050] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0051] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0052] Figure 1 This is a schematic diagram of a temperature field reconstruction method based on thermal load inversion in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the temperature field reconstruction model in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the temperature field reconstruction model generator in an embodiment of the present invention;

[0055] Figure 4This is a schematic diagram of a method for constructing a temperature field reconstruction model according to an embodiment of the present invention;

[0056] Figure 5 This is a schematic diagram of the discriminator in an embodiment of the present invention. Detailed Implementation

[0057] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0058] A specific embodiment of the present invention discloses a temperature field reconstruction method based on thermal load inversion guidance, such as... Figure 1 As shown, it includes the following steps:

[0059] Step 1: Obtain the real-time temperature values ​​of the sampling points at continuous intervals;

[0060] Step 2: Input the real-time temperature value into the thermal load inversion unit of the trained temperature field reconstruction model to obtain the thermal load information encoding; wherein, the temperature field reconstruction model includes a thermal load inversion unit and a generator;

[0061] Step 3: Input the encoded thermal load information and random noise matrix into the generator of the trained temperature field reconstruction model to obtain temperature field reconstruction data.

[0062] Furthermore, sampling points refer to several sensors pre-installed in the spacecraft to collect real-time temperature values ​​inside the spacecraft.

[0063] Specifically, such as Figure 2 As shown, the temperature field reconstruction model includes a thermal load inverter and a generator.

[0064] Furthermore, the step of inputting the real-time temperature value into the thermal load inversion unit of the trained temperature field reconstruction model to obtain the thermal load information encoding includes:

[0065] The real-time temperature value is encoded into an encoding matrix representing the current heat load information by the Encoder module of the Transformer of the heat load inverter, through multiple self-attention layers and feedforward fully connected layers.

[0066] Specifically, the heat load inverter is a model used to infer heat load information. It can learn and extract heat load information from input temperature data, reduce the number of times the generator is trained, and load heat load information can be added during application to predict temperature field data more quickly and accurately.

[0067] It should be noted that the thermal load inversion generator uses the Transformer's Encoder module, whose multiple self-attention layers can extract different features at a more granular level; the feedforward fully connected layers are used to enhance the network's capabilities. This effectively captures the temporal dependencies and thermal load variation trends in the input temperature data. Through encoding the input data, the thermal load inversion generator generates an encoding matrix containing thermal load information at the current moment as output.

[0068] Furthermore, the step of inputting the encoded thermal load information and the random noise matrix into the generator of the trained temperature field reconstruction model to obtain temperature field reconstruction data includes the following steps:

[0069] Specifically, such as Figure 3 As shown, the generator includes an encoder module, a dilated convolution module, and a decoder module; the encoder module is a downsampling structure, including three downsampling modules; the decoder module is an upsampling structure, including three upsampling modules.

[0070] Step S1: The thermal load information encoding matrix and the random noise matrix are fused together by the encoder module of the generator to extract features and obtain the encoder module output value, including:

[0071] The thermal load information encoding matrix and the random noise matrix are fused by summing at corresponding positions in the first downsampling module of the encoder module, and then input into a convolutional layer with a kernel of 5*5 and a stride of 1 for convolution operation before output.

[0072] It should be noted that the random noise matrix is ​​composed of random noise data that conforms to a normal Gaussian distribution arranged into a matrix structure.

[0073] The output value of the first downsampling module is sequentially passed through a convolutional layer with a kernel size of 3*3 and a stride size of 2 and a convolutional layer with a kernel size of 3*3 and a stride size of 1 in the second downsampling module of the encoder module before being output.

[0074] It should be noted that the second downsampling module uses a 3*3 convolutional layer with a stride of 2 to reduce the feature map by 1 / 2.

[0075] The output value of the second downsampling module is sequentially passed through a convolutional layer with a kernel size of 3*3 and a stride of 2 and two convolutional layers with kernel size of 3*3 and a stride of 1 in the third downsampling module of the encoder module, and then output as the encoder module output value.

[0076] It should be noted that the third downsampling module uses a 3*3 convolutional layer with a stride of 2 to further reduce the feature map by 1 / 2.

[0077] Step S2: The encoder module output value is passed through the generator's dilated convolution module to increase the receptive field, resulting in the dilated convolution module output value, including:

[0078] The output value of the encoder module is sequentially input into four dilated convolutional layers of the dilated convolutional module with 3*3 kernels, a stride of 1, and dilation rates of 2, 4, 6, and 8, respectively, to obtain the output value of the dilated convolutional module after expanding the receptive field.

[0079] Step S3: The output value of the dilated convolution module is passed through the decoder module of the generator to restore the original resolution and generate temperature field reconstruction data, including:

[0080] The output value of the dilated convolution module is passed through a convolutional layer with a kernel of 3*3 and a stride of 1 in the first upsampling module of the decoder, and then concatenated with the output value of the third downsampling module of the encoder module in the feature concatenation module. Finally, it is passed through a convolutional layer with a kernel of 3*3 and a stride of 1 for convolution operation before being output.

[0081] The output value of the first upsampling module is passed through a deconvolution layer with a kernel of 4*4 and a stride of 1 / 2 in the second upsampling module of the decoder, and then concatenated with the output value of the second downsampling module of the encoder module in the feature concatenation module. Finally, it is convolved through a convolution layer with a kernel of 3*3 and a stride of 1 before being output.

[0082] It should be noted that the deconvolution layer of the second upsampling module is used to magnify the image by 1 time.

[0083] The output value of the second upsampling module is passed through a deconvolution layer with a kernel of 4*4 and a stride of 1 / 2 in the third upsampling module of the decoder. After feature concatenation with the output value of the first downsampling module of the encoder module in the feature concatenation module, it is then passed through a convolution layer with a kernel of 3*3 and a stride of 1 to perform a convolution operation and output the device temperature field reconstruction data.

[0084] It should be noted that the deconvolution layer of the third upsampling module is used to enlarge the image by a factor of 1 to the original size.

[0085] Specifically, feature concatenation is used in the three upsampling modules to retain more dimensionality / positional information. This allows subsequent network layers to freely choose between shallow and deep features, which is more advantageous for semantic segmentation tasks, enabling edge feature retrieval. Since the feature maps are fused additively during feature concatenation, it is necessary to ensure that the size of the concatenated feature maps is the same.

[0086] Furthermore, during the generator's reconstruction of the temperature field, dilated convolutions are used in intermediate layers to increase the receptive field, aiming to capture as much image information as possible without losing additional detail. By learning the relationship between random noise and thermal load inversion coding, the generator can transform these input data into a high-quality reconstruction of the device temperature field.

[0087] Furthermore, the temperature field reconstruction model is achieved through, as follows: Figure 4 The method shown in the figure was used to train and obtain:

[0088] Step 101: Construct a training dataset for the temperature field reconstruction model; wherein, the training dataset includes real-time temperature values ​​at continuous sampling points, a random noise matrix, and the corresponding real temperature field data;

[0089] Step 102: Based on the training set, initially train the temperature field reconstruction model, update the parameters of the temperature field reconstruction model, and obtain the initially trained temperature field reconstruction model; wherein, the temperature field reconstruction model includes a thermal load inverter and a generator;

[0090] Step 103: Train the discriminator based on the training set and the temperature field reconstruction model trained in Step 102, update the discriminator parameters, and obtain the initially trained discriminator;

[0091] Step 104: Based on the training set, simultaneously train the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103 to obtain the trained temperature field reconstruction model.

[0092] In this embodiment, the sampling points in the training dataset refer to several sensors pre-installed in the spacecraft to collect temperature values ​​inside the spacecraft. The random noise matrix is ​​composed of random noise data that conforms to a normal Gaussian distribution arranged into a matrix structure. The corresponding real temperature field data is the temperature field data of the spacecraft calculated using finite element simulation software. This simulation software uses computer-aided design (CAD) software to create a geometric model, determine the material properties of each region in the model, divide the geometric model into finite element meshes, define the model boundary conditions, select a finite element solver according to the size and complexity of the model, assemble the system equations using appropriate boundary conditions, and then solve the system equations using numerical methods (such as the conjugate gradient method or the direct solution method) to obtain the numerical solution of the temperature field.

[0093] Furthermore, the first phase of training involves training the temperature field reconstruction model, including:

[0094] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model to generate thermal load information data;

[0095] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model, and generate temperature field reconstruction data with the thermal load information data;

[0096] Load the corresponding real temperature field data of the training set as labels;

[0097] The generator for the temperature field reconstruction model is trained using the mean squared error loss function. Through backpropagation and gradient descent optimization algorithms, the generator can continuously adjust its parameters to minimize the loss function. After training, the parameters of the temperature field reconstruction model are saved.

[0098] Specifically, the loss function measures the difference between the generator's output temperature field and the real temperature field, enabling the generator to progressively optimize the generated results to make them closer to the real temperature field distribution.

[0099] It should be noted that because the generator often performs poorly in the initial stage, the generated predicted structures are easily identified as fake samples by the discriminator. Therefore, the discriminator's loss is very small, close to zero, and it cannot function as a discriminator. To address this issue, in the first stage of training, the generator is trained first. Through multiple iterations of independent training of the generator and the thermal load inversion unit, its generation capability is gradually improved, generating more realistic temperature field data. In this way, by the time the second stage of training begins, the generator already possesses a certain level of generation capability.

[0100] Furthermore, the second phase of training involves training the discriminator, including:

[0101] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data;

[0102] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is used as a false sample input discriminator;

[0103] Load the corresponding real temperature field data of the training set as real samples into the discriminator;

[0104] The parameters of the temperature field reconstruction model trained in step 102 are fixed, and the discriminator is trained using the binary cross-entropy loss function as the loss function. By minimizing the binary cross-entropy loss, the discriminator parameters are updated so that the discriminator can accurately classify the input data into real samples or fake samples. The discriminator parameters are saved after training.

[0105] Specifically, such as Figure 5 As shown, the discriminator includes a fully convolutional neural network and a fully connected network, used to determine whether the device temperature field reconstruction data generated by the temperature field reconstruction model is a real sample or a fake sample;

[0106] The real temperature field data and the device temperature field reconstruction data generated by the generator are transformed into a 1024-dimensional feature vector through multiple convolution operations in the fully convolutional neural network layer of the discriminator.

[0107] The feature vector is passed through a fully connected layer to determine whether the generated device temperature field reconstruction data belongs to a real sample or a fake sample.

[0108] It should be noted that in the second stage of training, the discriminator is trained separately. By comparing the discriminator with the temperature field data predicted by the generator, it gradually learns how to accurately distinguish between real and generated samples. The training process of the discriminator helps improve its discrimination ability, enabling it to better evaluate the authenticity of the temperature field data generated by the generator.

[0109] Furthermore, the third stage of training: simultaneously training the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103, includes:

[0110] Load the real-time temperature values ​​of the training set into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data;

[0111] Load the random noise matrix of the training set into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is then used as a fake sample input into the discriminator trained in step 103.

[0112] Load the corresponding real temperature field data of the training set as real samples into the discriminator trained in step 103;

[0113] The temperature field reconstruction model and the discriminator are trained simultaneously using the mean squared error loss function and the binary cross-entropy loss function. The parameters of the temperature field reconstruction model and the discriminator are updated. After training, the parameters of the temperature field reconstruction model are saved to obtain the final temperature field reconstruction model.

[0114] It should be noted that in the third stage of training, through adversarial training between the generator and the discriminator, the discriminator continuously provides feedback on the generator's output, helping the generator to produce more realistic temperature field data. This game-like process prompts the generator to gradually improve its generation capabilities, making the generated temperature field data closer to the real temperature field distribution, thereby further improving the overall performance of thermal load inversion and temperature field reconstruction.

[0115] In summary, compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0116] 1. The technical solution of this invention, when predicting temperature field data, uses a thermal load inversion guide and generator to predict the temperature field of the equipment based on the real-time temperature values ​​sampled by the sensor. This enables real-time monitoring and prediction of the temperature distribution of the equipment, allowing for the early detection of potential temperature anomalies.

[0117] 2. Since changes in thermal load cannot be accurately obtained in real time, the technical solution of this invention can obtain real-time thermal load information of the equipment by inverting the temperature values ​​sampled by the sensor. This provides a key input for temperature field reconstruction, enabling more accurate prediction and analysis of the equipment's temperature distribution.

[0118] 3. The real-time temperature field prediction and thermal load information acquisition of the technical solution of this invention can provide reliable data support for the temperature management and thermal control of spacecraft. By timely monitoring the temperature changes of equipment, corresponding measures can be taken to optimize thermal control strategies, improve equipment efficiency and lifespan, and reduce potential failure risks.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A temperature field reconstruction method based on thermal load inversion guidance, characterized in that, include: Step 1: Obtain the real-time temperature values ​​of the sampling points at continuous intervals; Step 2: Input the real-time temperature value into the thermal load inversion unit of the trained temperature field reconstruction model to obtain the thermal load information encoding; wherein, the temperature field reconstruction model includes a thermal load inversion unit and a generator; Step 3: Input the encoded thermal load information and random noise matrix into the generator of the trained temperature field reconstruction model to obtain temperature field reconstruction data, including: The thermal load information encoding matrix and the random noise matrix are fused together by the encoder module of the generator to extract features and obtain the output value of the encoder module. The encoder module output value is passed through the generator's dilated convolution module, and the receptive field is increased to obtain the dilated convolution module output value. The output value of the dilated convolution module is used by the decoder module of the generator to restore the original resolution and generate temperature field reconstruction data.

2. The method according to claim 1, characterized in that, The step of inputting the real-time temperature value into the heat load inversion unit of the trained temperature field reconstruction model to obtain the heat load information encoding includes: The real-time temperature value is encoded into an encoding matrix representing the current heat load information by the Encoder module of the Transformer of the heat load inverter, through multiple self-attention layers and feedforward fully connected layers.

3. The method according to claim 2, characterized in that, The heat load information encoding matrix and the random noise matrix are passed through the encoder module of the generator, including: The thermal load information encoding matrix and the random noise matrix are fused by summing the elements at their corresponding positions in the first downsampling module of the encoder module, and then the result is output after convolution operation in the convolution layer with a stride of 1. The output value of the first downsampling module is then output after being convolved by a convolutional layer with a stride of 2 and a convolutional layer with a stride of 1 in the second downsampling module of the encoder module. The output value of the second downsampling module is sequentially convolved through a convolutional layer with a stride of 2 and two convolutional layers with a stride of 1 in the third downsampling module of the encoder module before being output as the encoder module output value.

4. The method according to claim 3, characterized in that, The encoder module output value is passed through the generator's dilated convolution module, and the receptive field is increased to obtain the dilated convolution module output value, including: The output value of the encoder module is sequentially input into four dilated convolutional layers of the dilated convolutional module with a stride of 1 and dilation rates of 2, 4, 6 and 8, respectively, to obtain the output value of the dilated convolutional module after expanding the receptive field.

5. The method according to claim 4, characterized in that, The output value of the dilated convolution module is used by the decoder module of the generator to restore the original resolution and generate temperature field reconstruction data, including: The output value of the dilated convolution module is passed through a convolutional layer with a stride of 1 in the first upsampling module of the decoder, and then concatenated with the output value of the third downsampling module of the encoder module in the feature concatenation module. Finally, it is passed through a convolutional layer with a stride of 1 for convolution operation before being output. The output value of the first upsampling module is passed through a deconvolution layer with a stride of 1 / 2 in the second upsampling module of the decoder, and then concatenated with the output value of the second downsampling module of the encoder module in the feature concatenation module. Finally, it is convolved through a convolution layer with a stride of 1 before being output. The output value of the second upsampling module is passed through a deconvolution layer with a stride of 1 / 2 in the third upsampling module of the decoder, and then concatenated with the output value of the first downsampling module of the encoder module in the feature concatenation module. Finally, it is passed through a convolution layer with a stride of 1 to perform a convolution operation and output the device temperature field reconstruction data.

6. The method according to claim 1, characterized in that, The temperature field reconstruction model was trained using the following method: Step 101: Construct a training dataset for the temperature field reconstruction model; wherein, the training dataset includes real-time temperature values ​​at continuous sampling points, a random noise matrix, and the corresponding real temperature field data; Step 102: Based on the training dataset, initially train the temperature field reconstruction model, update the parameters of the temperature field reconstruction model, and obtain the initially trained temperature field reconstruction model; wherein, the temperature field reconstruction model includes a thermal load inverter and a generator; Step 103: Train the discriminator based on the training dataset and the temperature field reconstruction model trained in Step 102, update the discriminator parameters, and obtain the initially trained discriminator; Step 104: Based on the training dataset, simultaneously train the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103 to obtain the trained temperature field reconstruction model.

7. The method according to claim 6, characterized in that, The training temperature field reconstruction model includes: Load the real-time temperature values ​​of the training dataset into the thermal load inversion unit of the temperature field reconstruction model to generate thermal load information data; Load the random noise matrix of the training dataset into the generator of the temperature field reconstruction model, and generate temperature field reconstruction data with the thermal load information data; Load the corresponding real temperature field data of the training dataset as labels; The generator for the temperature field reconstruction model is trained using the mean squared error loss function. Through backpropagation and gradient descent optimization algorithms, the generator can continuously adjust its parameters to minimize the loss function. After training, the parameters of the temperature field reconstruction model are saved.

8. The method according to claim 7, characterized in that, The trained discriminator includes: Load the real-time temperature values ​​of the training dataset into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data; Load the random noise matrix of the training dataset into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is used as a false sample input discriminator; Load the corresponding real temperature field data of the training dataset as real sample input to the discriminator; The parameters of the temperature field reconstruction model trained in step 102 are fixed, and the discriminator is trained using the binary cross-entropy loss function as the loss function. By minimizing the binary cross-entropy loss, the discriminator parameters are updated so that the discriminator can accurately classify the input data into real samples or fake samples. The discriminator parameters are saved after training.

9. The method according to claim 8, characterized in that, The simultaneous training of the temperature field reconstruction model trained in step 102 and the discriminator trained in step 103 includes: Load the real-time temperature values ​​of the training dataset into the thermal load inversion unit of the temperature field reconstruction model trained in step 102 to generate thermal load information data; Load the random noise matrix of the training dataset into the generator of the temperature field reconstruction model trained in step 102, and generate temperature field reconstruction data with the heat load information data, which is then used as a fake sample input into the discriminator trained in step 103. Load the corresponding real temperature field data of the training dataset as real samples into the discriminator trained in step 103; The temperature field reconstruction model and the discriminator are trained simultaneously using the mean squared error loss function and the binary cross-entropy loss function. The parameters of the temperature field reconstruction model and the discriminator are updated. After training, the parameters of the temperature field reconstruction model are saved to obtain the final temperature field reconstruction model.

Images