Spacecraft cabin thermal field digital twinborn model construction method based on decoupling characterization learning

Abstract

The invention discloses a spacecraft cabin thermal field digital twin model construction method based on decoupling characterization learning, and the method comprises the steps: constructing a training data set, each piece of data in the data set comprising thermal sensing data and thermal field distribution of a spacecraft cabin; constructing an initial neural network model, wherein the network model comprises a public feature extraction network, a working condition feature extraction network and a prediction network; the initial neural network model is trained in an adversarial learning mode, so that the public feature extraction network cannot distinguish data under different working conditions, and the working condition feature extraction network can recognize different working conditions; and based on the model obtained by training, inputting monitoring temperature information of a given working condition of the spacecraft, and outputting corresponding thermal field information. According to the method, the public features and the working condition features are extracted through the public feature extraction network and the working condition feature extraction network respectively, and working condition information influencing thermal field distribution is decoupled, so that the generalization ability of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal field reconstruction, and specifically relates to a method for constructing a digital twin model of the spacecraft cabin thermal field based on decoupled representation learning. Background Art

[0002] During the operation of a spacecraft, it faces complex and extreme external environmental conditions, resulting in extreme temperature changes for the spacecraft. For example, under the intense exposure of solar radiation, its surface temperature may rise to several hundred degrees Celsius, while in the Earth's shadow, the temperature may drop to negative values. Such temperature changes will have varying degrees of impact on the electronic, mechanical components, and instruments of the spacecraft. The digital twin technology of the spacecraft thermal field can, through limited sensing data input and real-time thermal field inversion technology, output the thermal field state of the unmonitored area in real time, thereby simulating and emulating the thermal field environment faced during the operation of the spacecraft in the digital space to ensure the performance and safety of the spacecraft under extreme temperature conditions.

[0003] Deep learning methods have good non-linear fitting capabilities and, through parallel computing acceleration, have achieved good results in fields such as computer vision and have been applied to the calculation process of the spacecraft thermal field twin. Existing deep learning algorithms for constructing a digital twin model of the thermal field inside the spacecraft cabin usually require a large number of training samples to learn the potential laws within the data through the deep learning model, so that the model can effectively perform thermal field reconstruction. However, the environmental working condition information during the operation of the spacecraft varies greatly, making it difficult for the training samples to cover all the spacecraft operation working conditions. Therefore, the generalization ability of the digital twin model of the thermal field inside the spacecraft cabin obtained by existing deep learning algorithms is greatly limited, that is, the thermal field inversion performance outside the training sample distribution is limited and it is difficult to meet the requirements of high-precision twin of the thermal field inside the spacecraft cabin. Summary of the Invention

[0004] 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 digital twin model of the spacecraft cabin thermal field based on decoupled representation learning.

[0005] The technical solution of the present invention is as follows:

[0006] There is provided a method for constructing a digital twin model of the spacecraft cabin thermal field based on decoupled representation learning, the method comprising: constructing a training data set, each piece of data in the data set including the thermal sensing data and the thermal field distribution of the spacecraft cabin;

[0007] Construct an initial neural network model, where the network model includes a common feature extraction network, a working condition feature extraction network, and a prediction network. The common feature extraction network is used to extract common features from thermal sensing data, the working condition feature extraction network is used to extract working condition features from thermal sensor data, and the prediction network is used to generate a predicted thermal field based on the common features and the working condition features;

[0008] Train the initial neural network model through adversarial learning, so that the common feature extraction network cannot distinguish data under different working conditions, and the working condition feature extraction network can identify different working conditions;

[0009] Based on the trained model, input the monitored temperature information of the given working condition of the spacecraft and output the corresponding thermal field information.

[0010] In an embodiment of the present invention, the common feature extraction network and the working condition feature extraction network are two parallel multi-layer perceptron networks.

[0011] In an embodiment of the present invention, when training the initial neural network model through adversarial learning, a discriminant network is introduced. The discriminant network is used to distinguish data under different working conditions, and then judge the working condition category according to the output of the common feature extraction network.

[0012] In an embodiment of the present invention, training the initial neural network model through adversarial learning includes two processes of alternately training the generator network and the discriminant network. The training of the generator network and the discriminant network are alternately carried out until the model converges.

[0013] In an embodiment of the present invention, the training of the generator network includes training the common feature extraction network and the working condition feature extraction network, so that the common feature extraction network cannot distinguish data under different working conditions, while the working condition feature extraction network can identify different working conditions.

[0014] In an embodiment of the present invention, the training of the discriminant network includes: training the discriminant network so that the discriminant network can distinguish data under different working conditions.

[0015] In an embodiment of the present invention, the loss function of the generator network training includes an inversion reconstruction loss function and a generation loss function. The inversion reconstruction loss function is used to measure the difference between the thermal field distribution output by the generator network and the real thermal field distribution, so that the generator network can accurately reconstruct the thermal field distribution corresponding to the input thermal sensing data. The generation loss function encourages the generator network to generate features similar to but indistinguishable from the real thermal field, making it difficult for the discriminant network to distinguish the features output by the generator network and the real features.

[0016] In an embodiment of the present invention, given a training sample (x i ,y i)| m , where x i represents the sample input sensing data, and y i represents the thermal field distribution corresponding to x i , m represents the set of working conditions to which the sample belongs when constructing the sample set, and the common feature extraction function is denoted as f 1 (·), the working condition feature extraction function is denoted as f 2 (·), the discriminant function is denoted as φ(·), then the inverse reconstruction loss function is expressed as:

[0017] L recon =|f 1 (x i )×f 2 (x i ) - y i | 2

[0018] where, |·| 2 represents the L 2 norm;

[0019] The generation loss function is expressed as:

[0020] L gen =-E(φ(x i ), m)

[0021] where, E(·) represents the polynomial logistic loss function;

[0022] The loss function for training the generation network is expressed as:

[0023] L generator =L recon +L gen .

[0024] In an embodiment of the present invention, the loss function for training the discriminant network is the discriminant loss function, which is the opposite of the generation loss function and is used to measure the difference between the working condition category predicted by the network and the true working condition so that the discriminant network can effectively distinguish the common features and the working condition features

[0025] The main advantages of the technical solution of the present invention are as follows:

[0026] The method for constructing a digital twin model of the spacecraft cabin thermal field based on decoupled representation learning of the present invention constructs a digital twin model based on a deep learning model, and utilizes the powerful nonlinear fitting ability of the deep learning model and the parallel acceleration computing ability of the deep learning to greatly improve the accuracy and the inversion calculation efficiency of the spacecraft thermal field inversion. By respectively extracting the common features and the working condition features through the common feature extraction network and the working condition feature extraction network, the working condition information affecting the thermal field distribution is decoupled, thereby improving the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] 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. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0028] Figure 1 It is a flowchart of a method for constructing a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning according to an embodiment of the present invention;

[0029] Figure 2 It is an architecture diagram of a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning according to an embodiment of the present invention;

[0030] Figure 3 It is a structural diagram of the generator network during the adversarial learning training of a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning according to an embodiment of the present invention;

[0031] Figure 4 It is a structural diagram of the discriminator network during the adversarial learning training of a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning provided by an embodiment of the present invention. Detailed implementation manners

[0032] 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 specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0033] The following will detail the technical solutions provided by the embodiments of the present invention in conjunction with the drawings.

[0034] The embodiments of the present invention provide a method for constructing a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning. As shown in the attached Figure 1 figures, it includes:

[0035] S1. Construct a training data set, and each piece of data in the data set includes the thermal sensing data, thermal field distribution, and working condition category of the spacecraft cabin.

[0036] In this embodiment, a training data set is first constructed. Each piece of data in the training data set includes the thermal sensing data and the thermal field distribution of the spacecraft cabin. Specifically, the thermal sensing data is collected by temperature sensors installed inside the spacecraft cabin, including temperature values at different positions and different time points; the thermal field distribution is obtained through numerical simulation or experimental measurement, representing the thermal field state at each position inside the spacecraft cabin.

[0037] The organization form of the data set is {X, Y}, where X represents the set of thermal sensing data, and Y represents the corresponding set of thermal field distributions. In order to cover different working conditions, the samples in the data set are divided into multiple working condition categories, and each category corresponds to a specific spacecraft operating environment (such as solar radiation intensity, earth shadow, etc.).

[0038] S2. Construct an initial neural network model. The network model includes a common feature extraction network, a working condition feature extraction network, and a prediction network. The common feature extraction network is used to extract common features from the thermal sensing data, the working condition feature extraction network is used to extract working condition features from the thermal sensor data, and the prediction network is used to generate a predicted thermal field based on the common features and the working condition features. The neural network architecture is as shown in the appendix Figure 2 as follows.

[0039] In the initial neural network model constructed in this embodiment, the common feature extraction network is used to extract common features from the thermal sensing data. Common features refer to the common thermal field features that exist under different working conditions, such as the heat conduction characteristics and heat capacity characteristics of the materials inside the spacecraft cabin. The working condition feature extraction network is used to extract working condition features from the thermal sensing data. Working condition features refer to the thermal field features related to specific working conditions, such as the influence of external environmental factors such as solar radiation intensity and earth shadow on the thermal field. The prediction network is used to generate a predicted thermal field based on the common features and the working condition features. Specifically, the prediction network fuses the common features and the working condition features and generates the thermal field distribution at each position inside the spacecraft cabin through a non-linear mapping.

[0040] S3. Train the initial neural network model through adversarial learning so that the common feature extraction network cannot distinguish data under different working conditions, and the working condition feature extraction network can identify different working conditions.

[0041] In the model training stage, the initial neural network model is trained through adversarial learning. The purpose of adversarial learning is to make the common feature extraction network unable to distinguish data under different working conditions, and at the same time make the working condition feature extraction network able to identify different working conditions.

[0042] The specific training process is as follows:

[0043] First, input the thermal sensing data into the common feature extraction network and the working condition feature extraction network to extract common features and working condition features respectively. Then, input the common features and working condition features into the prediction network to generate the predicted thermal field. Through adversarial learning, optimize the parameters of the common feature extraction network and the working condition feature extraction network, so that the features extracted by the common feature extraction network are as similar as possible under different working conditions, while the features extracted by the working condition feature extraction network can accurately reflect the differences between different working conditions.

[0044] S4. Based on the trained model, input the monitored temperature information of the given working condition of the spacecraft, and output the corresponding thermal field information.

[0045] After the model training is completed, based on the trained model, input the monitored temperature information of the given working condition of the spacecraft, and output the corresponding thermal field information. Specifically, input the real-time collected thermal sensing data into the model. The model extracts common features and working condition features through the common feature extraction network and the working condition feature extraction network respectively, and generates the thermal field distribution of each position in the spacecraft cabin through the prediction network.

[0046] The method for constructing a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning provided by the embodiments of the present invention constructs a digital twin model based on a deep learning model, and utilizes the powerful non-linear fitting ability of the deep learning model and the parallel acceleration computing ability of deep learning to greatly improve the accuracy and inversion calculation efficiency of the thermal field inversion of the spacecraft. By extracting common features and working condition features through the common feature extraction network and the working condition feature extraction network respectively, the working condition information affecting the thermal field distribution is decoupled, thereby improving the generalization ability of the model.

[0047] The following details each step and the related principles in the method for constructing a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning provided by the embodiments of the present invention.

[0048] The method for constructing a digital twin model of the thermal field of a spacecraft cabin based on decoupled representation learning provided by the embodiments of the present invention mainly includes multiple steps such as training dataset construction, feature extraction network architecture, decoupled representation adversarial learning, and digital twin tasks of the thermal field inside the spacecraft cabin.

[0049] (1). Training dataset construction

[0050] Given the training dataset of the thermal sensing data and the thermal field distribution inside the spacecraft cabin, denoted as {X, Y}, where X represents the set of thermal sensing data under different working conditions inside the spacecraft cabin, and Y is the set of corresponding thermal field distributions.

[0051] Divide X into Λ categories, that is, X = ∪X i (i = 1, 2,..., Λ), X i respectively represents the i-th working condition category. Therefore {X i, Y i} represents the sample corresponding to the working condition category i.

[0052] (2). Feature extraction network architecture

[0053] The present invention uses two parallel networks to extract working condition information and common feature information from the input sensing data respectively. That is, the common features are extracted through the common feature extraction network, and the working condition features are extracted through the working condition feature extraction network. Then, the working condition feature information and the common feature information are fused to obtain the inference result of the thermal field inside the spacecraft cabin.

[0054] Among them, the common feature extraction network and the working condition feature extraction network are two parallel multi-layer perceptrons. A multi-layer perceptron generally includes an input layer, multiple hidden layers, and an output layer. Each layer contains a linear operation and a non-linear activation function. According to the universal approximation theorem of neural networks, when the neural network is large enough, the neural network can approximate any function. The commonly used method for model training is the backpropagation algorithm. According to the error of the output layer and the chain rule of derivatives, the adjustment methods of the weights and biases in the network can be deduced, and the error is backpropagated to update the network parameters. In the embodiment of the present invention, a multi-layer perceptron network is selected as the backbone network model for feature extraction, and the powerful non-linear fitting ability of the deep learning model and the parallel acceleration calculation ability of deep learning are used to greatly improve the accuracy and calculation efficiency of the spacecraft thermal field inversion.

[0055] (3). Decoupled representation adversarial learning

[0056] The present invention constructs a decoupled feature training loss function and trains the neural network through adversarial learning to obtain a twin model of the thermal field inside the spacecraft cabin.

[0057] Training the initial neural network model through adversarial learning includes two processes: generating network training and discriminant network training, which are carried out alternately. Generating network training includes training the common feature extraction network and the working condition feature extraction network, so that the common feature extraction network cannot distinguish data under different working conditions, while the working condition feature extraction network can identify different working conditions. Discriminant network training includes training the discriminant network so that the discriminant network can distinguish data under different working conditions.

[0058] The structure of the generating network can be seen in the appendix Figure 3 . The purpose of generating network training is to make the network model unable to distinguish information under different working conditions, that is, to make the common feature network model able to extract the common feature part in the thermal field features as much as possible. Therefore, its training loss function is mainly divided into two parts, namely the inversion reconstruction loss function and the generating loss function.

[0059] Given the training sample (x i , y i ) |m , where x i represents the sample input sensing information, and y i represents the thermal field information corresponding to x i , and m represents the set of working conditions to which the sample belongs when constructing the sample set. Assume that the common feature extraction function is f 1 (·), the working condition feature extraction function is f 2 (·), and the discriminator function is φ(·). Then the inverse reconstruction loss function can be expressed as:

[0060] L recon = |f 1 (x i ) × f 2 (x i ) - y i | 2

[0061] where |·| 2 represents the L 2 norm.

[0062] The generation loss function can be expressed as:

[0063] L gen = -E(φ(x i ), m)

[0064] where E(·) represents the polynomial logistic loss function.

[0065] The total training loss function is:

[0066] L generator = L recon + L gen .

[0067] The discriminant network structure can be referred to in Appendix Figure 4 . The purpose of training the discriminant network is to enable the network model to classify data under different working conditions into different categories. Therefore, its training loss function is mainly the discriminant loss function, and its expression is:

[0068] L discriminator = E(φ(x i ), m)

[0069] where E(·) represents the polynomial logistic loss function.

[0070] It can be seen that through the adversarial learning method, the decoupled representation learning of the thermal field is realized based on the training loss function, and the internal component working state features, external condition features, common features, etc. that affect the thermal field distribution are explicitly extracted to construct a machine learning model for physical feature decoupling.

[0071] (IV). The Digital Twin Task of the Thermal Field inside the Spacecraft Cabin

[0072] After training the neural network model, when performing the digital twin task of the spacecraft cabin thermal field, the thermal sensing data monitored by the sensor is used as the input of the neural network model, and the thermal field information can be output by the trained neural network. Since the present invention constructs a thermal field digital twin model based on decoupled representation, when using this model to predict the spacecraft cabin thermal field, by decoupling the working condition factors that affect the spacecraft thermal field distribution, the relationship between the thermal field influencing factors and the thermal field distribution can be explicitly established, improving the physical interpretability of the deep learning model and the generalization ability under unknown working conditions. The generalization of the thermal field digital twin model to unknown working condition scenarios under limited working conditions is realized.

[0073] 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 actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising 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", "upper" and "lower" are all referenced to the placement state shown in the drawings.

[0074] 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 recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning, characterized in that: include: Constructing a training data set, wherein each piece of data in the data set includes thermal sensing data and thermal field distribution of a spacecraft cabin; Constructing an initial neural network model, the network model includes a common feature extraction network, an operating condition feature extraction network and a prediction network, the common feature extraction network is used to extract common features from thermal sensor data, the operating condition feature extraction network is used to extract operating condition features from thermal sensor data, and the prediction network is used to generate a predicted thermal field according to the common features and the operating condition features; Training the initial neural network model by adversarial learning so that the common feature extraction network cannot distinguish data under different working conditions, and the working condition feature extraction network can identify different working conditions; Based on the trained model, the temperature information of the spacecraft under given working conditions is input and the corresponding thermal field information is output.

2. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 1 is characterized in that: The common feature extraction network and the operating condition feature extraction network are two parallel multi-layer perceptron networks.

3. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 2 is characterized in that: When training the initial neural network model through adversarial learning, a discriminant network is introduced to distinguish data under different working conditions, and then judge the working condition category based on the output of the common feature extraction network.

4. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 3 is characterized in that: Training the initial neural network model by adversarial learning includes two processes: alternating generative network training and discriminative network training. The generative network training and the discriminative network training are performed alternately until the model converges.

5. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 4 is characterized in that: The generation network training includes training a common feature extraction network and an operating condition feature extraction network, so that the common feature extraction network cannot distinguish data under different operating conditions, while the operating condition feature extraction network can identify different operating conditions.

6. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 5 is characterized in that: The discriminant network training includes: training the discriminant network so that the discriminant network can distinguish data under different working conditions.

7. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 6 is characterized in that: The loss function of the generative network training includes an inversion and reconstruction loss function and a generation loss function. The inversion and reconstruction loss function is used to measure the difference between the thermal field distribution output by the generative network and the real thermal field distribution, so that the generative network can accurately reconstruct the thermal field distribution corresponding to the input thermal sensing data. The generation loss function encourages the generative network to generate features that are similar to but indistinguishable from the real thermal field so that it is difficult for the discriminant network to distinguish between the features output by the generative network and the real features.

8. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 7, characterized in that: Given a training sample (x i ,y i )| m , where x i represents the sample input sensor data, y i Represents x i Corresponding to the thermal field distribution, m represents the working condition set to which the sample belongs when constructing the sample set, the common feature extraction function is expressed as f1(·), the working condition feature extraction function is expressed as f2(·), and the discriminant function is expressed as φ(·). Then the inversion and reconstruction loss function is expressed as: L recon =|f1(x i )×f2(x i )-y i |2 Among them, |·|2 represents the L2 norm; The generation loss function is expressed as: L gen -E(ϕ(x i ),m) Where E(·) represents the multinomial logistic loss function; The loss function of the generated network training is expressed as: L generator =L recon +L gen .

9. The method for constructing a digital twin model of a spacecraft cabin thermal field based on decoupled representation learning according to claim 4, characterized in that: The loss function of the discriminant network training is the discriminant loss function, which is the opposite of the generation loss function and is used to measure the difference between the working condition category predicted by the network and the actual working condition so that the discriminant network can effectively distinguish between common features and working condition features.

Images