Thermal field digital twin model construction method based on one shot neural network architecture search

Abstract

The invention discloses a thermal field digital twin model construction method based on one shot neural network architecture search, which comprises the following steps: designing a search space: defining the search space of a network structure based on a MobileNetV2 model, the search space comprising an expansion rate, a convolution kernel size and a path number; constructing and training a super network: constructing the super network containing all possible paths in the search space, and training the super network based on the heat source layout data set; searching for an optimal model architecture: searching for an approximately optimal model architecture in the trained super network by using an evolutionary algorithm; and constructing a physical field prediction agent model: constructing a lightweight thermal field digital twin agent model based on the searched optimal model architecture for predicting the thermal field distribution in the spacecraft in real time. According to the method, through combination of search space design and super network training, the lightweight model meeting the precision requirement can be quickly generated. The agent model is low in parameter quantity, high in calculation efficiency and adaptive to an embedded environment on a spacecraft end.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twins, and particularly to a method for constructing a thermal field digital twin model based on one-shot neural network architecture search. Background Art

[0002] During the process of a spacecraft traveling in space, it needs to withstand the test of extreme temperature differences. For example, when facing the sun directly, it may encounter extremely high temperatures, reaching hundreds of degrees Celsius; while when facing away from the sun and in the Earth's shadow area, it will encounter extremely low temperatures, even dropping below the freezing point. Such drastic temperature fluctuations pose a severe challenge to the spacecraft. Whether it is sensitive electronic equipment, precision mechanical equipment, or high-precision scientific instruments, they may be damaged or their performance may decline due to thermal stress.

[0003] To address this problem, the spacecraft thermal field digital twin technology has emerged. This technology cleverly utilizes existing sensor data and combines advanced real-time thermal field reverse analysis algorithms to be able to reconstruct the entire thermal distribution of the spacecraft in a virtual environment, even for areas that are difficult to directly monitor. Through this digital replication and simulation method, scientific researchers can accurately predict the performance of the spacecraft in various complex thermal environments on the ground, ensuring its structural integrity, normal function, and safe operation, and being able to respond robustly even in the face of the harsh temperature differences in space. In this way, the spacecraft thermal field digital twin has become a key technical guarantee for ensuring the success of long-term space exploration missions.

[0004] Normally, the physical field-level digital twin transmits sensing data to the cloud and processes large-scale and complex calculations based on the powerful computing resources of the cloud. However, this process highly depends on network connections and has problems such as transmission delays and data security risks, and it is difficult to adapt to scenarios with high data uncertainty. Therefore, it is necessary to build a physical field-level digital twin system on the edge, and by directly performing calculations on the edge, improve the real-time response and autonomous decision-making capabilities of the system. However, in order to ensure the accuracy and real-time performance of the thermal field twin calculation, the current physical field digital twin calculation model usually uses a neural network model to perform inverse calculation on the thermal field inside the spacecraft cabin based on the thermal field information obtained by the spacecraft sensors. However, the traditional neural network model has a large number of parameters and is difficult to apply to the edge computing environment with limited resources on the edge. Summary of the Invention

[0005] To solve some or all of the above-mentioned technical problems existing in the prior art, the present invention provides a method for constructing a thermal field digital twin model based on one-shot neural network architecture search.

[0006] The technical solution of the present invention is as follows:

[0007] A method for constructing a thermal field digital twin model based on one-shot neural network architecture search is provided, and the method includes:

[0008] Design the search space: Define the search space of the network structure based on the MobileNetV2 model, and the search space includes expansion rate, convolution kernel size, and number of paths;

[0009] Construct and train the super network: Construct a super network containing all possible paths within the search space, and train the super network based on the heat source layout data set;

[0010] Search for the optimal model architecture: Use an evolutionary algorithm to search for an approximately optimal model architecture in the trained super network;

[0011] Construct a physical field prediction proxy model: Based on the searched optimal model architecture, construct a lightweight thermal field digital twin proxy model for real-time prediction of the internal thermal field distribution of the spacecraft.

[0012] In an embodiment of the present invention, in the design of the search space, an FPN feature pyramid network is used for multi-scale feature fusion, combined with MobileNetV2 to achieve lightweight design.

[0013] In an embodiment of the present invention, in the search space, the convolution kernel size is selected from 3, 5, 7, 9, and the expansion rate is set to 3 or 6.

[0014] In an embodiment of the present invention, in the search space, different combinations of convolution kernel sizes can be selected in each layer, and the maximum number of selected paths in each layer is a random integer from 1 to 4.

[0015] In an embodiment of the present invention, when training the super network, paths are randomly selected and sub-models are trained, and the combination of expansion rate and convolution kernel size is randomly selected during training.

[0016] In an embodiment of the present invention, the evolutionary algorithm is a multi-objective genetic algorithm.

[0017] In an embodiment of the present invention, during the search process, the NSGA-II algorithm is adopted, the model parameter quantity and prediction accuracy are modeled as a double-objective optimization problem, and new individuals are generated through crossover operation and mutation operation.

[0018] In an embodiment of the present invention, when searching for an approximately optimal model architecture, the performance of the searched model is evaluated through accuracy, parameter quantity, time complexity, and inference time.

[0019] In an embodiment of the present invention, before training the super network, a heat source layout data set is constructed, and the heat source layout data set includes the thermal field data collected by the sensors inside the spacecraft cabin and the corresponding thermal field distribution maps.

[0020] In one embodiment of the present invention, the constructed physical field prediction proxy model is deployed in an embedded device on a spacecraft for real-time response and prediction of the thermal field distribution.

[0021] The main advantages of the technical solution of the present invention are as follows:

[0022] The method for constructing a thermal field digital twin model based on one-shot neural network architecture search of the present invention can quickly generate a lightweight model that meets the accuracy requirements by combining search space design and hypernetwork training. The physical field prediction proxy model obtained by the search has a low number of parameters and high computational efficiency, and is suitable for the embedded environment on a spacecraft. The model can predict the thermal field distribution in real time and meet the rapid response requirements of the spacecraft in a complex thermal environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0024] Figure 1 It is a flowchart of the method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to an embodiment of the present invention;

[0025] Figure 2 It is a structural diagram of the search space in the method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only some, rather than all, embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] The following will detail the technical solutions provided in the embodiments of the present invention with reference to the drawings.

[0028] An embodiment of the present invention provides a method for constructing a thermal field digital twin model based on one-shot neural network architecture search. As shown in the attached Figure 1 figures, it includes:

[0029] S1, Design the search space: Based on the MobileNetV2 model, define the search space of the network structure, which includes the expansion rate, the convolution kernel size, and the number of paths.

[0030] In this embodiment, the design of the search space is based on the MobileNetV2 model, defining the possible architecture combinations in the network structure. The search space includes the following:

[0031] Expansion rate: Used to control the width range of the convolutional layers in the network to adjust the model capacity and feature extraction ability.

[0032] Convolution kernel size: Defined as multiple optional values for extracting feature information of different scales.

[0033] Number of paths: The number of paths that can be selected in each layer, which determines the diversity and flexibility of the model.

[0034] Through the design of the above search space, a basic framework covering various architecture possibilities is provided, laying a foundation for subsequent network training and architecture optimization.

[0035] S2, Construct and train the supernetwork: Construct a supernetwork containing all possible paths within the search space and train the supernetwork based on the heat source layout dataset.

[0036] Construct a supernetwork containing all possible path combinations, which covers all architecture possibilities in the search space. During the training process, the supernetwork gradually updates its parameters by randomly selecting path combinations.

[0037] The training data of the supernetwork comes from the heat source layout dataset, which includes the thermal field data collected inside the spacecraft cabin and the corresponding thermal field distribution maps. In each training, different sub-models of paths are randomly activated to ensure that each possible architecture in the search space can be effectively learned.

[0038] S3, Search for the optimal model architecture: Use an evolutionary algorithm to search for an approximately optimal model architecture in the trained supernetwork.

[0039] Use an evolutionary algorithm to search for the optimal model architecture in the trained supernetwork. The search process includes:

[0040] Taking the number of model parameters and prediction accuracy as the goals, screen the optimal path combination in the supernetwork.

[0041] Through iterative optimization, select the model architecture that best meets the performance requirements, providing a basis for subsequent physical field prediction tasks.

[0042] S4. Construct a physical field prediction surrogate model: Based on the optimal model architecture found, construct a lightweight digital twin surrogate model for the thermal field to predict the internal thermal field distribution of the spacecraft in real time.

[0043] Based on the optimal model architecture found, construct a lightweight physical field prediction surrogate model. This model is used to predict the thermal field distribution inside the spacecraft cabin in real time. On the one hand, it can provide accurate thermal field distribution prediction results to support the thermal environment optimization of the spacecraft; on the other hand, it runs with low power consumption on embedded devices to adapt to the computing resource limitations on the spacecraft side.

[0044] In summary, the method for constructing a digital twin model of the thermal field based on one-shot neural network architecture search in the embodiments of the present invention can quickly generate a lightweight model that meets the accuracy requirements through the combination of search space design and hypernetwork training. The physical field prediction surrogate model found has a low number of parameters and high computational efficiency, and is adapted to the embedded environment on the spacecraft side. The model can predict the thermal field distribution in real time and meet the rapid response requirements of the spacecraft in a complex thermal environment.

[0045] In some optional embodiments of the present invention, in the designed search space, the Feature Pyramid Network (FPN) is used for multi-scale feature fusion, and combined with MobileNetV2 to achieve lightweight design.

[0046] FPN is a multi-scale feature fusion network that can effectively process spatial or temporal variation features at different scales. In the physical field-level digital twin task, the thermal field distribution often involves complex multi-scale phenomena (such as local hotspots and global temperature distribution). FPN can capture both high-level semantic information and low-level spatial details by constructing a feature pyramid, thereby improving the prediction accuracy of the model for the thermal field distribution.

[0047] MobileNetV2 is a lightweight convolutional neural network that significantly reduces the number of model parameters and computational complexity through an inverted residual structure and depthwise separable convolutions. Combining FPN and MobileNetV2 can further reduce the computational burden of the model while ensuring the multi-scale feature fusion ability, making it more suitable for deployment in resource-constrained spacecraft-side devices.

[0048] Multi-scale feature fusion is achieved through FPN to improve the model's prediction ability for complex thermal field distributions; lightweight design is achieved through MobileNetV2 to reduce the number of model parameters and computational complexity. With such a design, on the premise of ensuring prediction accuracy, the computational resource requirements of the model are significantly reduced, making it more suitable for running in the edge computing environment on the spacecraft side.

[0049] In some optional embodiments of the present invention, in the search space, the convolutional kernel size is selected from 3, 5, 7, 9, and the expansion rate is set to 3 or 6.

[0050] The size of the convolutional kernel directly affects the receptive field of the model for the input features. Smaller convolutional kernels (such as 3x3) are suitable for capturing local details, while larger convolutional kernels (such as 9x9) are suitable for capturing global features. By setting multiple convolutional kernel sizes, the model can adaptively select the most suitable feature extraction method, thereby improving the prediction accuracy of the thermal field distribution.

[0051] The expansion rate determines the multiple of the number of channels of the feature map in the convolutional layer. When the expansion rate is set to 3 or 6, it can ensure the feature extraction ability while avoiding excessive parameter redundancy and further reducing the computational complexity of the model.

[0052] By setting multiple convolutional kernel sizes and expansion rates, the feature extraction ability of the model is improved while avoiding excessive parameter redundancy. With such a design, on the premise of ensuring the model accuracy, the number of parameters and the computational complexity of the model are further reduced, making it more suitable for running on resource-constrained spacecraft terminal devices.

[0053] In some optional embodiments of the present invention, in the search space, different combinations of convolutional kernel sizes can be selected in each layer, and the maximum number of selected paths in each layer is a random integer from 1 to 4.

[0054] Different combinations of convolutional kernel sizes (such as 3x3 and 5x5) are allowed to be selected in each layer. This multi-path design can enhance the expressive ability of the model, enabling it to capture feature information of different scales simultaneously. The maximum number of selected paths in each layer is a random integer from 1 to 4. This design can avoid excessive computational burden while ensuring the diversity of the model. Through the multi-path convolutional kernel combination and the random path number design, the ability of the model to capture features of different scales is enhanced, and the expressive ability and diversity of the model are improved.

[0055] In some optional embodiments of the present invention, when training the supernetwork, paths are randomly selected and the sub-models are trained, and the combination of the expansion rate and the convolutional kernel size is randomly selected during training.

[0056] When training the supernetwork, paths are randomly selected and the sub-models are trained. This design can ensure that the supernetwork fully explores all possible paths in the search space during the training process, thereby improving the generalization ability of the model. During the training process, the combination of the expansion rate and the convolutional kernel size is also randomly selected. This design can further improve the diversity of the model and avoid falling into local optimal solutions. By randomly selecting paths and the combination of the expansion rate and the convolutional kernel size, it is ensured that the supernetwork fully explores the search space, and the generalization ability and diversity of the model are improved.

[0057] In some alternative embodiments of the present invention, the evolutionary algorithm is a multi-objective genetic algorithm. Exemplarily, the NSGA-II algorithm is adopted during the search process, and the model parameter quantity and prediction accuracy are modeled as a two-objective optimization problem, and new individuals are generated through crossover operations and mutation operations.

[0058] NSGA-II is a classic multi-objective genetic algorithm that can effectively handle two-objective optimization problems. During the search process, NSGA-II generates new individuals through crossover operations and mutation operations, thereby continuously optimizing the model architecture. Modeling the model parameter quantity and prediction accuracy as a two-objective optimization problem can simultaneously consider the accuracy and computational complexity of the model during the search process, so as to find the optimal model architecture.

[0059] In some alternative embodiments of the present invention, when searching for an approximately optimal model architecture, the performance of the searched model is evaluated by accuracy, parameter quantity, time complexity, and inference time. During the search process, the performance of the model is evaluated by four indicators: accuracy (MAE), parameter quantity, time complexity (FLOPs), and inference time. These indicators can comprehensively reflect the prediction ability, computational complexity, and real-time performance of the model. Through multi-dimensional evaluation, it can be ensured that the searched model not only performs excellently in terms of accuracy but also meets the requirements of spacecraft on-board equipment in terms of computational complexity and real-time performance.

[0060] In some alternative embodiments of the present invention, before training the super network, a heat source layout data set is constructed. The heat source layout data set includes the thermal field data collected by sensors in the spacecraft cabin and the corresponding thermal field distribution maps. The heat source layout data set is the basis for training the super network, including the thermal field data collected by sensors in the spacecraft cabin and the corresponding thermal field distribution maps. These data can reflect the thermal field distribution of the spacecraft under different thermal environments. By constructing a high-quality heat source layout data set, it can be ensured that the super network learns the true thermal field distribution law during the training process, thereby improving the prediction accuracy of the model.

[0061] The following gives a specific embodiment of the method for constructing a thermal field digital twin model based on one-shot neural network architecture search provided by the present invention.

[0062] The present invention proposes a one-shot neural network architecture search method, including search space design, search algorithm construction, and performance evaluation, specifically as follows:

[0063] I. Search Space

[0064] The FPN (Feature Pyramid Network) model constructs a multi-scale feature pyramid and is suitable for tasks that require cross-scale information fusion, especially showing good performance in physical field-level twin computing applications. In physical field-level twin computing, it is usually necessary to model and predict complex physical phenomena, which often involve spatial or temporal variation characteristics at different scales. The FPN model builds a feature pyramid structure on top of a base network (such as ResNet, VGG, etc.), achieving efficient fusion from deep high-semantic features to shallow high-resolution features. Such a design not only retains the semantic information of high-level features but also utilizes the spatial details of low-level features, providing comprehensive representation capabilities for the subtle changes and global patterns of the physical field. This section takes FPN as a representative to study neural network architecture search methods.

[0065] The ResNet50 model in the original FPN is too large. Therefore, defining a suitable search space is the first step in neural network architecture search (NAS). The present invention uses MobileNetV2 as a benchmark model to define the search space. MobileNetV2 is a clean, small, and efficient deep learning model with a stack of multiple inverted blocks. Due to the design of the inverted blocks, the entire network is small but has a strong ability to extract features of the input image. The inverted blocks are shown on the Figure 2 left side of. Given a predetermined number of channels in each layer, the actual number of channels during the convolution operation is multiplied by the expansion rate. The convolutional kernel size of each layer is selected from conv3×3, conv5×5, conv7×7, and conv9×9. They have different abilities to extract feature maps. Then, through the use of the conv1×1 operation, the number of channels of the feature map is reduced to the predefined number of channels in each layer, and this operation is used to change the number of channels of the feature map.

[0066] The present invention searches for the combination of the optimal convolutional kernel size and expansion rate in a fixed super network structure. As shown in Figure 2 , the expansion rate is selected from [3, 6], and the convolutional kernel size is selected from [3, 6, 7, 9]. In addition, the combination of convolutional kernel sizes is allowed to be multi-path, that is, multiple sizes of convolutional kernels are allowed to be set in each layer. m represents the maximum number of selected paths in each layer of the neural network and is set to a random integer between [1, 4]. The input layout image is first input into a fixed preprocessing layer. The total number of layers of the entire network architecture is set to 12, and the backbone model is evenly divided into four parts. Then the backbone structure is combined with the FPN framework to form the final Mixpath_FPN model.

[0067] II. Search Algorithm

[0068] The main idea of the above steps is to first train a supernetwork consisting of all possible paths. During training, each path will be randomly activated. When the training terminates, even if the neural network cannot achieve the accuracy directly used for predicting the temperature field, it has a certain model ranking ability, which is equivalent to providing a surrogate model that can evaluate the accuracy of different models. Therefore, based on the trained supernetwork, evolutionary algorithms can be combined to search for an approximately optimal model architecture. The model found by the evolutionary algorithm will be retrained for a sufficient number of epochs to meet the requirement of high prediction accuracy for the temperature field.

[0069] In the optimization solution process, it is first necessary to encode each model, that is, an individual. Here, one of the sub-paths, that is, a sub-model, is used as an example, and its encoding is as follows:

[0070]

[0071] In the formula: rate: 0 and 1 for each layer represent expansion rates of 3 and 6 respectively; 0, 1, 2, 3 represent the sizes of convolutional kernels conv3×3, conv5×5, conv7×7, and conv9×9.

[0072] Specifically, the training process of the supernetwork can be seen in Algorithm 1. First, a supernetwork composed of all selected paths is constructed. Then, it is trained using the defined heat source layout dataset. Different from the previous training of the entire network, the sub-model is trained after randomly selecting a path. The combination of the expansion rate and the convolutional kernel size is randomly selected during the training of the supernetwork. After training the supernetwork for N epochs, considering two objectives of the model parameter quantity and accuracy, it is modeled as a multi-objective optimization problem, and a multi-objective genetic algorithm is used to search for an approximately optimal sub-model. The specific algorithm process is shown in Algorithm 2. In the i crossover process, after randomly selecting two individuals in i P, the combinations in each layer are randomly exchanged according to the corresponding crossover rate c P. In the mutation process, after randomly selecting an individual in i P, the combinations of the expansion rate and the convolutional kernel size are copied in each layer according to the mutation rate m P. Then all new individuals are saved to Q. After merging the populations i P i and Q, the non-dominated solutions F are calculated. The individuals in F are considered to be approximately optimal model architectures that can meet the requirements of fast deployment or high prediction accuracy.

[0073]

[0074]

[0075] III. Performance Evaluation

[0076] When evaluating the performance of the searched model, it is compared with other manually designed models from the following four aspects:

[0077] · Accuracy: MAE is directly used to evaluate the accuracy of the training model performance. The lower the MAE, the higher the prediction accuracy of the deep learning model.

[0078] · Number of parameters: The total number of parameters of the neural network model. The fewer the number of parameters, the lower the training cost of the neural network model.

[0079] · Time complexity FLOPs: The number of floating-point operations. This metric can evaluate the time complexity of the deep neural network.

[0080] · Inference time: The average time for performing the forward calculation of the neural network on the input layout image.

[0081] IV. Physical field prediction

[0082] Construct a physical field prediction surrogate model through the above steps, and then input the spacecraft cabin sensor data into the trained physical field prediction surrogate model. The physical field prediction surrogate model outputs the thermal field distribution to complete the real-time prediction task.

[0083] It should be noted that in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. In addition, in this article, "front", "rear", "left", "right", "up" and "down" are all referenced to the placement state shown in the drawings.

[0084] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a thermal field digital twin model based on one shot neural network architecture search, characterized in that Including: Design search space: Define the search space of the network structure based on the MobileNetV2 model. The search space includes expansion rate, convolution kernel size, and number of paths; Construct and train the super network: Construct a super network containing all possible paths within the search space, and train the super network based on the heat source layout dataset; Search for the optimal model architecture: Use an evolutionary algorithm to search for an approximately optimal model architecture in the trained super network; Construct a physical field prediction proxy model: Based on the searched optimal model architecture, construct a lightweight digital twin proxy model of the thermal field for real-time prediction of the internal thermal field distribution of the spacecraft.

2. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 1, wherein In the design search space, use the FPN feature pyramid network for multi-scale feature fusion and combine with MobileNetV2 to achieve lightweight design.

3. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 2, wherein In the search space, the convolution kernel size is selected from 3, 5, 7, 9, and the expansion rate is set to 3 or 6.

4. The method for constructing a thermal field digital twin model based on one shot neural network architecture search according to claim 3, wherein In the search space, different combinations of convolution kernel sizes can be selected in each layer, and the maximum number of selected paths in each layer is a random integer from 1 to 4.

5. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 4, wherein When training the super network, randomly select paths and train the sub-models. The combination of expansion rate and convolution kernel size is randomly selected during training.

6. The method for constructing a thermal field digital twin model based on one shot neural network architecture search according to claim 1, characterized in that, The evolutionary algorithm is a multi-objective genetic algorithm.

7. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 6, wherein During the search process, use the NSGA-II algorithm, model the model parameter quantity and prediction accuracy as a bi-objective optimization problem, and generate new individuals through crossover operation and mutation operation.

8. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 1, characterized in that When searching for an approximately optimal model architecture, evaluate the performance of the searched model through accuracy, parameter quantity, time complexity, and inference time.

9. The method for constructing a thermal field digital twin model based on one-shot neural network architecture search according to claim 1, wherein Before training the super network, construct a heat source layout dataset, which includes the thermal field data collected by the sensors in the spacecraft cabin and the corresponding thermal field distribution maps.

10. The method for constructing a thermal field digital twin model based on the search of one shot neural network architecture according to claim 1, characterized in that, The constructed physical field prediction proxy model is deployed in the embedded device on the spacecraft side for real-time response and prediction of the thermal field distribution.

Images