A method for reconstructing the temperature field inside a satellite cabin based on model uncertainty and semi-supervised learning

Abstract

This invention discloses a method for reconstructing the temperature field inside a satellite cabin based on model uncertainty and semi-supervised learning, comprising: acquiring first training data and second training data; preprocessing the training data; constructing a deep learning neural network model; randomly initializing the deep learning neural network model, training the model using the first training data, repeating the initialization and training process multiple times to obtain multiple trained models; determining the model to be updated and an auxiliary model; inputting the temperature of a preset temperature measurement point in the second training data into the auxiliary model, calculating the predicted temperature field and prediction uncertainty corresponding to the temperature of the preset temperature measurement point; training the model to be updated using the first training data, the second training data, and their corresponding predicted temperature field and prediction uncertainty; and reconstructing the temperature field inside the satellite cabin. This invention can achieve rapid reconstruction of the temperature field inside a satellite cabin using the temperatures of a limited number of temperature measurement points, and requires low computational cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite thermal control, and particularly relates to a satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning. BACKGROUND

[0002] Small-sized electronic devices and their applications have become one of the fastest growing fields at present, and the normal operation of electronic devices depends greatly on stable operating environment temperature. For a satellite in working state, the electronic devices in the satellite cabin inevitably generate heat dissipation when working, causing the temperature of the electronic devices to rise. Since the vacuum environment in space makes the electronic devices unable to dissipate heat outward through thermal convection, heat is more likely to accumulate, causing the temperature of the devices to rise rapidly, affecting their service life, safety and reliability, and even causing device failure or damage. In order to ensure the normal operation of the electronic devices in the satellite cabin, it is necessary to monitor and control the environmental temperature in the satellite cabin.

[0003] In order to realize the monitoring of the environmental temperature in the satellite cabin, conventionally, a number of temperature sensors are arranged at selected position points in a specific area in the satellite cabin, and the temperature data of the limited temperature measurement points collected by the temperature sensors are used to reconstruct the overall temperature field of the specific area by using interpolation methods including bilinear interpolation and Kriging method. However, in actual application, the calculation efficiency of the traditional interpolation method is low, numerical calculation needs a large number of iterations, real-time prediction is difficult to achieve, and high-dimensional modeling problems cannot be handled, the reconstruction accuracy is poor, and the actual engineering problem requirements cannot be met.

[0004] In order to overcome the problems existing in the traditional method, some researches have proposed to use a deep learning agent model for temperature field reconstruction. This method constructs a deep learning agent model, trains the agent model using a large amount of temperature measurement point temperature data with overall temperature field labels, and then uses the trained deep learning agent model for temperature field reconstruction.

[0005] However, the existing method of using a deep learning agent model to reconstruct the temperature field of a specific area in a satellite cabin needs to use a large amount of temperature measurement point temperature data with temperature field labels to train the agent model, and the real data of the overall temperature field corresponding to the specific temperature measurement point temperature data is difficult to obtain, still needs to be obtained through simulation experiment and interpolation calculation, which also needs to consume more computing resources and computing time, still has the problems of high calculation cost and low temperature field reconstruction efficiency. Moreover, for specific temperature measurement point temperature data, a single deep learning agent model can only give a positive prediction result, and cannot give a quantity evaluating the goodness of the current prediction result, and the prediction performance of the agent model cannot be evaluated. SUMMARY

[0006] To solve the above problems in the prior art, the present application provides a satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning.

[0007] The technical solutions of the present application are as follows:

[0008] A satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning is provided, and the method comprises:

[0009] A plurality of first training data and a plurality of second training data are obtained, the first training data comprising the temperature of a preset temperature measuring point in the satellite cabin and the corresponding temperature field in the satellite cabin, and the second training data comprising the temperature of the preset temperature measuring point in the satellite cabin;

[0010] The training data is preprocessed;

[0011] A deep learning neural network model is constructed;

[0012] The deep learning neural network model is randomly initialized, the first training data is used to train the deep learning neural network model, a trained deep learning neural network model is obtained, the random initialization and training process are repeated for multiple times, and a plurality of trained deep learning neural network models are obtained;

[0013] A to-be-updated model and an auxiliary model are determined, the to-be-updated model being a deep learning neural network model with the best prediction performance in the plurality of trained deep learning neural network models, and the auxiliary model being other models in the plurality of trained deep learning neural network models except the to-be-updated model;

[0014] The temperature of the preset temperature measuring point in the second training data is input into the auxiliary model, and the predicted temperature field and the prediction uncertainty corresponding to the input temperature of the preset temperature measuring point are calculated;

[0015] The to-be-updated model is trained by using the first training data, the second training data, and the corresponding predicted temperature field and prediction uncertainty;

[0016] The trained to-be-updated model is used to reconstruct the temperature field in the satellite cabin.

[0017] In some possible implementation manners, obtaining the first training data comprises:

[0018] The position of the preset temperature measuring point in the satellite cabin is determined;

[0019] For each electronic device in the satellite cabin, a power value is randomly sampled from the working power distribution to which each electronic device is subject as the current power of the corresponding electronic device, to obtain an electronic device layout in the satellite cabin under a specific working condition, to determine the temperature of each preset temperature measuring point and the temperature field in the satellite cabin under the current specific working condition, to obtain first training data including the temperature of the preset temperature measuring point in the satellite cabin and the corresponding temperature field in the satellite cabin, and the process is repeated for multiple times until a preset number of first training data is obtained.

[0020] In some possible implementation manners, obtaining the second training data includes:

[0021] Determining the position of the preset temperature measuring point in the satellite cabin;

[0022] For each electronic device in the satellite cabin, a power value is randomly sampled from the working power distribution to which each electronic device is subject as the current power of the corresponding electronic device, to obtain an electronic device layout in the satellite cabin under a specific working condition, to determine the temperature of each preset temperature measuring point under the current specific working condition, to obtain second training data including the temperature of the preset temperature measuring point in the satellite cabin, and the process is repeated for multiple times until a preset number of second training data is obtained.

[0023] In some possible implementation manners, the pre-processing of the training data includes:

[0024] Grid division is performed on the electronic device layout area in the satellite cabin, a matrix is constructed according to the number of divided grids, and the temperature data of the preset temperature measuring point in the satellite cabin and the temperature field data in the satellite cabin are represented by the matrix, respectively.

[0025] In some possible implementation manners, the training of the deep learning neural network model by using the first training data includes:

[0026] The temperature of the preset temperature measuring point in the first training data is taken as input, and the temperature field corresponding to the input temperature of the preset temperature measuring point is taken as output, and the deep learning neural network model is trained by using the gradient descent method.

[0027] In some possible implementation manners, the predicted temperature field corresponding to the input temperature of the preset temperature measuring point is calculated and determined by using the following formula:

[0028]

[0029] Wherein, T represents the predicted temperature field corresponding to the input temperature of the preset temperature measuring point, Q represents the number of auxiliary models, T q represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point output by the qth auxiliary model.

[0030] In some possible implementations, the prediction uncertainty corresponding to the temperature of the preset temperature measuring point input is determined by using the following formula:

[0031]

[0032] wherein, T std represents the prediction uncertainty corresponding to the temperature of the preset temperature measuring point input.

[0033] In some possible implementations, the to-be-updated model is trained by using the first training data, the second training data, and the corresponding predicted temperature field and prediction uncertainty, comprising:

[0034] The temperature of the preset temperature measuring point in the first training data and the temperature of the preset temperature measuring point in the second training data are taken as inputs, the temperature field corresponding to the temperature of the preset temperature measuring point in the input first training data and the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the input second training data are taken as outputs, and the to-be-updated model is trained by using a preset loss function through a gradient descent method.

[0035] In some possible implementations, when the to-be-updated model is trained, the loss function is set as:

[0036]

[0037] wherein, a and b represent preset weight coefficients, represents the number of elements of the temperature field matrix, N represents the number of the first training data input to the to-be-updated model, sum(*) represents the sum of all elements in *, and and represents the temperature of the preset temperature measuring point in the i-th first training data and the temperature field in the satellite cabin, represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the i-th first training data output by the to-be-updated model, and N' represents the number of the second training data input to the to-be-updated model, represents the temperature of the preset temperature measuring point in the j-th second training data, represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the j-th second training data output by the to-be-updated model, and W j represents the uncertainty weight matrix corresponding to the temperature of the preset temperature measuring point in the j-th second training data, represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the j-th second training data.

[0038] In some possible implementations, the uncertainty weight matrix W j is determined by using the following formula:

[0039]

[0040] wherein, T std_j represents the prediction uncertainty corresponding to the temperature of the preset temperature measuring point in the jth second training data, and γ represents an uncertainty threshold.

[0041] The main advantages of the technical scheme of the present application are as follows:

[0042] The satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning can obtain a deep learning agent model with extremely high prediction accuracy by using a small amount of labeled training data, significantly reduces the demand for labeled training data for model training, effectively reduces the calculation time and calculation resources required for agent model training, can quickly predict the overall temperature field in the satellite cabin by using the temperature data of a limited number of temperature measuring points, and realizes real-time, rapid and high-precision reconstruction of the temperature field in the satellite cabin. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 The flowchart of the satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning of an embodiment of the present application;

[0045] Figure 2 The training process schematic diagram of the model to be updated of an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the present application more clear, the technical scheme of the present application will be described in detail below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] The technical scheme provided by the embodiments of the present application will be described in detail below in combination with the drawings.

[0048] Reference Figure 1 An embodiment of the present application provides a satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning, which comprises the following steps S1-S8:

[0049] Step S1, obtaining a plurality of first training data and a plurality of second training data, the first training data comprising temperatures of preset temperature measuring points in the satellite cabin and corresponding temperature fields in the satellite cabin, and the second training data comprising temperatures of preset temperature measuring points in the satellite cabin.

[0050] The layout area of the electronic equipment in the satellite cabin can be regarded as a square layout area of a two-dimensional plane. Since the satellite is in a vacuum environment, there is no convection heat exchange, and the radiation heat exchange can be ignored. The heat generated by the electronic equipment in the layout area of the satellite cabin is in the form of heat conduction. Therefore, the layout area of the electronic equipment in the satellite cabin can be regarded as a two-dimensional plane heat conduction application scene.

[0051] Further, it is provided that: the layout area in the satellite cabin is arranged with Λ electronic equipment, each electronic equipment is regarded as a heat source, the heat source represents a component capable of generating energy dissipation when working, the power distribution of the i-th heat source is φ i (x,y), the layout area is arranged with M temperature measuring points with temperature sensors, the position of the m-th temperature measuring point is

[0052] Since the heat generated by the electronic equipment in the layout area in the satellite cabin is in the form of heat conduction, the temperature field of the layout area satisfies the following heat conduction differential equation:

[0053]

[0054] Wherein, T represents the temperature field of the layout area in the satellite cabin, k represents the heat conduction coefficient, and (x,y) represents the position coordinates of a certain point in the temperature field.

[0055] For different boundary heat dissipation modes of the layout area in the satellite cabin, the temperature field of the layout area in the satellite cabin can satisfy the following boundary conditions:

[0056] T=T0,

[0057] Wherein, T0 represents the temperature value at the isothermal boundary, n represents the normal line perpendicular to the, h represents the heat convection coefficient, T=T0 represents the Dirichlet boundary condition, i.e. the first type of boundary condition, represents the Neumann boundary condition, i.e. the second type of boundary condition, represents the Robin boundary condition, i.e. the third type of boundary condition.

[0058] In an embodiment of the present application, the satellite cabin temperature field reconstruction represents obtaining the temperature field of the layout area in the satellite cabin by using the temperature values of the temperature measuring points in the layout area in the satellite cabin. To this end, based on the above setting, the satellite cabin temperature field reconstruction problem can be modeled as:

[0059]

[0060] wherein, denotes the temperature of the location O in the reconstructed temperature field O denotes the temperature of the location O in the reconstructed temperature field m denotes the temperature value of the mth temperature measuring point.

[0061] Since the positions of the electronic devices in the layout area in the satellite cabin are fixed, but the power changes with the actual working condition, in an embodiment of the present application, the first training data is obtained in the following manner:

[0062] The positions of the preset temperature measuring points in the satellite cabin are determined.

[0063] For each electronic device in the satellite cabin, a power value is randomly sampled from the working power distribution to which each electronic device is subjected as the current power of the corresponding electronic device, to obtain an electronic device layout in the satellite cabin in a specific working condition, the temperatures of the preset temperature measuring points in the current specific working condition are determined, to obtain a first training data including the temperatures of the preset temperature measuring points in the satellite cabin and the corresponding temperature field in the satellite cabin, and the process is repeated for multiple times until a preset number of first training data is obtained.

[0064] In the process of obtaining the first training data, the temperature field in the satellite cabin in the specific working condition can be calculated and analyzed by using a finite element analysis method or an interpolation method; the temperature of the preset temperature measuring point can be extracted from the obtained temperature field in the satellite cabin, or the power of the electronic device in the actual layout area in the satellite cabin can be adjusted according to the sampled power, and the temperature of the temperature measuring point is directly collected by using the temperature sensor arranged on the preset temperature measuring point after the power adjustment is completed.

[0065] Further, in an embodiment of the present application, the second training data is obtained in the following manner:

[0066] The positions of the preset temperature measuring points in the satellite cabin are determined.

[0067] For each electronic device in the satellite cabin, a power value is randomly sampled from the working power distribution to which each electronic device is subjected as the current power of the corresponding electronic device, to obtain an electronic device layout in the satellite cabin in a specific working condition, the temperatures of the preset temperature measuring points in the current specific working condition are determined, to obtain a first training data including the temperatures of the preset temperature measuring points in the satellite cabin and the corresponding temperature field in the satellite cabin, and the process is repeated for multiple times until a preset number of first training data is obtained.

[0068] In an embodiment of the present application, when the second training data is acquired, the temperature field in the satellite cabin under a specific working condition can be calculated and analyzed by using a finite element analysis method or an interpolation method to acquire the temperature field in the satellite cabin under the specific working condition, and the temperature of the preset temperature measuring point is extracted from the acquired temperature field in the satellite cabin; or the power of the electronic equipment in the actual layout area in the satellite cabin is adjusted according to the sampled power, and the temperature of the temperature measuring point is directly collected by using the temperature sensor arranged on the preset temperature measuring point after the power adjustment is completed.

[0069] Further, in an embodiment of the present application, the temperature measuring point is determined in the following manner:

[0070] The center of a plurality of electronic equipment is selected as a temperature measuring point, a plurality of position points are selected as temperature measuring points at each boundary of the electronic equipment layout area in the satellite cabin, and a plurality of position points are randomly selected as temperature measuring points in other areas in the electronic equipment layout area in the satellite cabin except the area where the electronic equipment is located.

[0071] The number of the selected temperature measuring points can be set according to actual conditions.

[0072] Further, in an embodiment of the present application, the specific number of the first training data and the specific number of the second training data can be set according to the training efficiency and the accuracy requirement, for example, the number of the first training data is 1000-5000 or other values, and the number of the second training data is 20000-50000 or other values. In order to improve the prediction accuracy of the finally obtained model and reduce the calculation cost, the number of the second training data is greater than the number of the first training data.

[0073] In an embodiment of the present application, when the temperature field in the satellite cabin is calculated and analyzed by using the finite element analysis method, the electronic equipment layout area in the satellite cabin is divided into N1×N2 finite element calculation grids for simulation calculation to acquire the corresponding temperature field. The specific value of the finite element calculation grid can be determined according to the required resolution and accuracy, when the required resolution and accuracy are high, the more the finite element calculation grids divided for simulation calculation, but correspondingly, the more the time and resources required for simulation calculation.

[0074] In step S2, the training data is preprocessed.

[0075] In an embodiment of the present application, in order to facilitate the training of the subsequently constructed deep learning neural network model by using the training data, the preprocessing of the training data includes:

[0076] The electronic equipment layout area in the satellite cabin is divided into grids, a matrix is constructed according to the number of the divided grids, and the temperature data of the preset temperature measuring point in the satellite cabin and the temperature field data in the satellite cabin are represented by using the matrix respectively.

[0077] Specifically, taking the satellite cabin as an example, the satellite cabin is divided into N1* N2 grids, and the temperature data of the preset temperature measuring point in the satellite cabin and the temperature field data in the satellite cabin are represented by a temperature measuring point matrix and a temperature field matrix, and the dimensions of the temperature measuring point matrix and the temperature field matrix are N1* N2.

[0078] Further, in an embodiment of the present application, the temperature data of the preset temperature measuring point in the satellite cabin and the temperature field data in the satellite cabin are represented by a matrix, including:

[0079] For the matrix representing the temperature data of the preset temperature measuring point in the satellite cabin, the matrix elements corresponding to the grid positions where the temperature measuring points exist are set to the temperatures of the corresponding temperature measuring points, and the matrix elements corresponding to the grid positions where the temperature measuring points do not exist are set to 0.

[0080] For the matrix representing the temperature field in the satellite cabin, the matrix elements corresponding to each grid position are set to the temperature of the temperature field position point corresponding to the center point of the grid position, and the matrix elements of the temperature field matrix are normalized to ensure that the matrix elements belong to [0, 1].

[0081] Step S3, constructing a deep learning neural network model.

[0082] In an embodiment of the present application, a U-Net convolutional neural network is used as a deep learning neural network model to fit the mapping relationship between the temperature of the preset temperature measuring point and the temperature field in the satellite cabin. By training the deep learning neural network model using the training data obtained above, the deep learning neural network model can learn the physical laws in the training data, thereby obtaining a neural network with strong generalization ability, and realizing fast prediction of the corresponding temperature field in the satellite cabin under the temperature of the specific preset temperature measuring point. In essence, the deep learning neural network model is a proxy model.

[0083] Step S4, randomly initializing the deep learning neural network model, training the deep learning neural network model using the first training data, obtaining a trained deep learning neural network model, and repeating the random initialization and training process multiple times to obtain multiple trained deep learning neural network models.

[0084] In order to obtain a proxy model with high prediction accuracy using a small amount of training data with temperature field labels, in an embodiment of the present application, a plurality of deep learning proxy models, i.e. a plurality of trained deep learning neural network models, are obtained by randomly initializing and independently training the first training data including the temperature field in the satellite cabin.

[0085] Specifically, in the process of obtaining a deep learning agent model, the initial parameters of the model are determined by randomly initializing the deep learning neural network model, and then the initialized deep learning neural network model is trained by using the first training data, so as to obtain a trained deep learning neural network model.

[0086] In an embodiment of the present application, the training of the deep learning neural network model by using the first training data comprises:

[0087] The temperature of the preset temperature measuring point in the first training data is taken as the input, and the temperature field corresponding to the input temperature of the preset temperature measuring point is taken as the output, and the deep learning neural network model is trained by using the gradient descent method.

[0088] Further, in an embodiment of the present application, the temperature of the preset temperature measuring point in the first training data is taken as the input, and the temperature field corresponding to the input temperature of the preset temperature measuring point is taken as the output, and the deep learning neural network model is trained by using the gradient descent method, comprising the following steps S401-S403:

[0089] In step S401, the temperature of the preset temperature measuring point in the plurality of first training data is sequentially input into the deep learning neural network model, and the predicted temperature field corresponding to each first training data output by the deep learning neural network model is obtained.

[0090] In an embodiment of the present application, the temperature of the preset temperature measuring point in the first training data is input from the input end of the deep learning neural network model, sequentially processed by the parameters of each layer in the deep learning neural network model, and output from the output end of the deep learning neural network model, and the information output from the output end of the deep learning neural network model is the predicted temperature field corresponding to the input temperature of the preset temperature measuring point.

[0091] In step S402, the predicted temperature field output by the deep learning neural network model is compared with the temperature field in the first training data, and the prediction accuracy of the deep learning neural network model is calculated.

[0092] In an embodiment of the present application, the average value of all ratios is taken as the prediction accuracy by calculating the difference between the predicted temperature field corresponding to each first training data and the temperature field in the first training data and the ratio of the temperature field in the first training data.

[0093] In step S403, it is judged whether the prediction accuracy is greater than a preset accuracy threshold, if yes, the current deep learning neural network model is taken as the trained deep learning neural network model, and if not, the parameters of the deep learning neural network model are adjusted by using a preset loss function, and the step S401 is returned.

[0094] In an embodiment of the present application, the preset accuracy threshold is set according to the prediction accuracy of the actual required deep learning neural network model.

[0095] Further, in an embodiment of the present application, when training the deep learning neural network model, the corresponding loss function can be set according to the actual situation, for example, using the Mean Absolute Error (MAE) as the loss function.

[0096] Further, in an embodiment of the present application, the following formula is used to update the parameters of the deep learning neural network model:

[0097]

[0098] wherein, Θ t+1 represents the parameters of the deep learning neural network model at the t+1th iteration, Θ t represents the parameters of the deep learning neural network model at the tth iteration, Δ[·] represents the optimizer, η represents the learning rate, L(Θ) represents the loss function, and Θ represents the parameters of the deep learning neural network model. The optimizer is, for example, Adam, SGD, etc., and the learning rate needs to be set in advance to control the speed of parameter updating.

[0099] Since the more deep learning neural network models are obtained, the prediction accuracy is higher when using the integrated multiple deep learning neural network models for temperature field prediction, but the training time and training period are longer. In an embodiment of the present application, the number of deep learning neural network models is set according to the actual required prediction accuracy and training time. For example, it is set to 4-8.

[0100] Step S5, determine the to-be-updated model and the auxiliary model, the to-be-updated model is the deep learning neural network model with the best prediction performance in the plurality of trained deep learning neural network models, and the auxiliary model is the other model in the plurality of trained deep learning neural network models except the to-be-updated model.

[0101] In an embodiment of the present application, the prediction performance of the deep learning neural network model is determined by calculating the Mean Absolute Error (MAE) corresponding to the model, and the lower the MAE, the better the prediction performance of the model.

[0102] Specifically, the deep learning neural network model with the best prediction performance in the plurality of trained deep learning neural network models is determined in the following manner:

[0103] A certain number of first training data is selected, the temperature of the preset temperature measuring point in the selected first training data is input into the trained deep learning neural network model, a predicted temperature field corresponding to the first training data output by the deep learning neural network model is obtained, the mean absolute error is calculated according to the temperature field in the first training data and the predicted temperature field corresponding to the first training data, and the deep learning neural network model with the minimum mean absolute error is determined as the deep learning neural network model with the best prediction performance.

[0104] The number of the selected first training data is specifically set according to actual requirements.

[0105] In step S6, the temperature of the preset temperature measuring point in the second training data is input into the auxiliary model, and a predicted temperature field and a prediction uncertainty corresponding to the input temperature of the preset temperature measuring point are calculated.

[0106] Referring to Figure 2 , the accompanying Figure 2 In this embodiment, the deep learning neural network model is 5, including Unet1, Unet2, Unet3, Unet4 and Unet5, and the number of auxiliary models is 4, including Unet2, Unet3, Unet4 and Unet5. In an embodiment of the present application, the temperature of the preset temperature measuring point in the second training data is input into each auxiliary model, a predicted temperature field output by each auxiliary model is obtained, and a predicted temperature field and a prediction uncertainty corresponding to the input temperature of the preset temperature measuring point are calculated according to the predicted temperature fields output by all auxiliary models.

[0107] In an embodiment of the present application, the predicted temperature field corresponding to the input temperature of the preset temperature measuring point is calculated and determined by using the following formula:

[0108]

[0109] In this embodiment, The predicted temperature field corresponding to the input temperature of the preset temperature measuring point is represented by T, Q represents the number of auxiliary models, T q The predicted temperature field corresponding to the temperature of the preset temperature measuring point output by the qth auxiliary model is represented by Tq.

[0110] In an embodiment of the present application, the prediction uncertainty corresponding to the input temperature of the preset temperature measuring point is calculated and determined by using the following formula:

[0111]

[0112] In this embodiment, std The prediction uncertainty corresponding to the input temperature of the preset temperature measuring point is represented by U.

[0113] According to the calculation formula, by using multiple auxiliary models to predict the temperature field corresponding to the temperature of the preset temperature measuring point in each second training data, the predicted temperature field and the prediction uncertainty corresponding to each second training data can be determined.

[0114] In an embodiment of the present application, by using multiple auxiliary models to perform integrated prediction of the temperature field in the satellite cabin corresponding to the second training data, the prediction accuracy of the temperature field can be further improved. Moreover, the prediction accuracy of the predicted temperature field can be measured according to the obtained prediction uncertainty, and the smaller the prediction uncertainty, the higher the reliability of the predicted temperature field.

[0115] In step S7, the first training data, the second training data, the predicted temperature field corresponding thereto, and the prediction uncertainty are used to train the to-be-updated model.

[0116] In an embodiment of the present application, by using the first training data, the second training data, the predicted temperature field corresponding thereto, and the prediction uncertainty to train the to-be-updated model, the problem of the small number of first training data can be compensated, and the prediction accuracy of the to-be-updated model can be further improved.

[0117] Referring to Figure 2 In an embodiment of the present application, the first training data, the second training data, the predicted temperature field corresponding thereto, and the prediction uncertainty are used to train the to-be-updated model, including:

[0118] The temperature of the preset temperature measuring point in the first training data and the temperature of the preset temperature measuring point in the second training data are taken as input, the temperature field corresponding to the temperature of the preset temperature measuring point in the input first training data and the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the input second training data are taken as output, and the to-be-updated model is trained by using a preset loss function and a gradient descent method.

[0119] Further, in an embodiment of the present application, the temperature of the preset temperature measuring point in the first training data and the temperature of the preset temperature measuring point in the second training data are taken as input, the temperature field corresponding to the temperature of the preset temperature measuring point in the input first training data and the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the input second training data are taken as output, and the to-be-updated model is trained by using a preset loss function and a gradient descent method, including the following steps S701-S703:

[0120] In step S701, the temperatures of the preset temperature measuring points in the multiple first training data are sequentially input into the to-be-updated model, and the temperatures of the preset temperature measuring points in the multiple second training data are sequentially input into the to-be-updated model, to obtain the predicted temperature field corresponding to each first training data and the predicted temperature field corresponding to each second training data output by the to-be-updated model.

[0121] Step S702, comparing the predicted temperature field output by the to-be-updated model with the temperature field in the first training data and the predicted temperature field corresponding to the second training data, and calculating the prediction accuracy of the to-be-updated model;

[0122] Step S703, judging whether the prediction accuracy is greater than a preset accuracy threshold, if yes, taking the current to-be-updated model as the to-be-updated model after training, if not, adjusting the parameters of the to-be-updated model by using a preset loss function, and returning to step S701.

[0123] In an embodiment of the present application, the difference between the predicted temperature field corresponding to each first training data output by the to-be-updated model and the temperature field in the first training data and the ratio of the temperature field in the first training data, and the difference between the predicted temperature field corresponding to each second training data output by the to-be-updated model and the predicted temperature field corresponding to the second training data obtained in step S6 and the ratio of the predicted temperature field corresponding to the second training data obtained in step S6 are calculated, and the average of all ratios is taken as the prediction accuracy.

[0124] In an embodiment of the present application, the preset accuracy threshold is set according to the prediction accuracy of the deep learning neural network model required in practice.

[0125] In an embodiment of the present application, when training the to-be-updated model, the loss function is set as:

[0126]

[0127] Wherein, α and β represent preset weight coefficients, is the loss function part corresponding to the first training data, Ω represents the number of elements of the temperature field matrix, N represents the number of first training data input into the to-be-updated model, sum(*) represents summing all elements in *, and and represents the temperature of the preset temperature measuring point in the i-th first training data and the temperature field in the satellite cabin, represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the i-th first training data output by the to-be-updated model, is the loss function part corresponding to the second training data, N' represents the number of second training data input into the to-be-updated model, represents the temperature of the preset temperature measuring point in the j-th second training data, represents the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the j-th second training data output by the to-be-updated model, W j represents the uncertainty weight matrix corresponding to the temperature of the preset temperature measuring point in the j-th second training data, a predicted temperature field corresponding to the temperature of the preset temperature measuring point in the jth second training data.

[0128] In an embodiment of the present application, the uncertainty weight matrix W j is determined by using the following formula:

[0129]

[0130] wherein T std_j represents a predicted uncertainty corresponding to the temperature of the preset temperature measuring point in the jth second training data, and γ represents an uncertainty threshold.

[0131] wherein the predicted temperature field corresponding to the temperature of the preset temperature measuring point in the jth second training data and the predicted uncertainty T std_j The predicted temperature field and the predicted uncertainty calculated in the above step S6, and the uncertainty threshold γ are set according to actual conditions.

[0132] Further, in an embodiment of the present application, the following formula is used to update the parameters of the to-be-updated model:

[0133]

[0134] wherein θ t+1 represents the parameters of the to-be-updated model at the t+1th iteration, θ t represents the parameters of the to-be-updated model at the tth iteration, Δ[·] represents an optimizer, η represents a learning rate, and θ represents the parameters of the to-be-updated model. The optimizer is, for example, Adam, SGD, etc., and the learning rate needs to be set in advance and is used to control the speed of parameter updating.

[0135] In step S8, the trained to-be-updated model is used to reconstruct the temperature field in the satellite cabin.

[0136] Specifically, after the training of the to-be-updated model is completed, the temperature of the preset temperature measuring point of the to-be-solved temperature field in the satellite cabin is input into the to-be-updated model, so that the temperature field in the satellite cabin corresponding to the temperature of the preset temperature measuring point output by the to-be-updated model can be obtained.

[0137] The satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning provided in an embodiment of the present application can obtain a deep learning agent model with extremely high prediction accuracy by using a small amount of labeled training data, significantly reduces the demand for labeled training data for model training, effectively reduces the calculation time and calculation resources required for agent model training, can quickly predict the overall temperature field in the satellite cabin by using the temperature data of limited temperature measuring points, and realizes real-time, rapid and high-precision reconstruction of the temperature field in the satellite cabin.

[0138] It should be noted that, in this article, relational terms such as "first" and "second", and the like, are used solely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, "front", "rear", "left", "right", "upper", "lower", and the like, are all referred to the placement state shown in the drawings.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; 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 application.

Claims

1. A satellite cabin temperature field reconstruction method based on model uncertainty and semi-supervised learning, characterized in that, The method comprises the following steps: obtaining a plurality of first training data and a plurality of second training data, the first training data comprising the temperature of a preset temperature measuring point in a satellite cabin and the corresponding temperature field in the satellite cabin, and the second training data comprising the temperature of the preset temperature measuring point in the satellite cabin, wherein the temperature of the preset temperature measuring point is collected by a temperature sensor arranged on the preset temperature measuring point; preprocessing the training data, wherein preprocessing the training data comprises: dividing the layout area of the electronic equipment in the satellite cabin into a grid, constructing a matrix according to the number of the divided grid, and respectively using the matrix to represent the temperature data of the preset temperature measuring point in the satellite cabin and the temperature field data in the satellite cabin; constructing a deep learning neural network model; randomly initializing the deep learning neural network model, training the deep learning neural network model by using the first training data to obtain a trained deep learning neural network model, and repeatedly performing the random initialization and training process to obtain a plurality of trained deep learning neural network models; determining an updated model and an auxiliary model, the updated model being a deep learning neural network model with the best prediction performance among the plurality of trained deep learning neural network models, and the auxiliary model being other models among the plurality of trained deep learning neural network models except the updated model; inputting the temperature of the preset temperature measuring point in the second training data into the auxiliary model, and calculating the predicted temperature field and the prediction uncertainty corresponding to the input temperature of the preset temperature measuring point; training the updated model by using the first training data, the second training data, and the corresponding predicted temperature field and prediction uncertainty; reconstructing the temperature field in the satellite cabin by using the trained updated model; the temperature field reconstruction in the satellite cabin represents obtaining the temperature field of the layout area in the satellite cabin by using the temperature values of the temperature measuring points in the layout area in the satellite cabin; obtaining the first training data comprises: determining the positions of the preset temperature measuring points in the satellite cabin; for each electronic equipment in the satellite cabin, randomly sampling a power value from the working power distribution of each electronic equipment as the current power of the corresponding electronic equipment to obtain a specific working condition of the layout of the electronic equipment in the satellite cabin, determining the temperature of each preset temperature measuring point and the temperature field in the satellite cabin under the current specific working condition, obtaining a first training data comprising the temperature of the preset temperature measuring point in the satellite cabin and the corresponding temperature field in the satellite cabin, and repeating the process for a plurality of times until a preset number of first training data is obtained; obtaining the second training data comprises: determining the positions of the preset temperature measuring points in the satellite cabin; for each electronic equipment in the satellite cabin, randomly sampling a power value from the working power distribution of each electronic equipment as the current power of the corresponding electronic equipment to obtain a specific working condition of the layout of the electronic equipment in the satellite cabin, determining the temperature of each preset temperature measuring point, obtaining a second training data comprising the temperature of the preset temperature measuring point in the satellite cabin, and repeating the process for a plurality of times until a preset number of second training data is obtained.

2. The method of claim 1, wherein, training the deep learning neural network model by using the first training data comprises: The temperature of the preset temperature measuring point in the first training data is taken as input, and the temperature field corresponding to the input temperature of the preset temperature measuring point is taken as output, and the deep learning neural network model is trained by the gradient descent method.

3. The method of claim 1, wherein, The predicted temperature field corresponding to the input temperature of the preset temperature measuring point is calculated and determined by using the following formula: ； wherein, represents a predicted temperature field corresponding to a temperature of a preset temperature measuring point input by the temperature field prediction model, represents a number of auxiliary models, represents a predicted temperature field corresponding to a temperature of a preset temperature measuring point output by the i-th auxiliary model. represents a predicted temperature field corresponding to a temperature of a preset temperature measuring point output by the i-th auxiliary model.

4. The method of claim 3, wherein, The predicted uncertainty corresponding to the input temperature of the preset temperature measuring point is calculated and determined by using the following formula: ； wherein, represents the predicted uncertainty corresponding to the temperature of the input preset temperature measuring point.

5. The method of claim 1, wherein, The first training data, the second training data, and the corresponding predicted temperature field and predicted uncertainty are used to train the to-be-updated model, including: The temperature of the preset temperature measuring point in the first training data and the temperature of the preset temperature measuring point in the second training data are taken as input, and the temperature field corresponding to the input temperature of the preset temperature measuring point in the first training data and the predicted temperature field corresponding to the input temperature of the preset temperature measuring point in the second training data are taken as output, and the to-be-updated model is trained by using the preset loss function and the gradient descent method.

6. The method of claim 5, wherein, When training the to-be-updated model, the loss function is set as: ； in, and This indicates the preset weighting coefficient. This indicates the number of elements in the temperature field matrix. This indicates the number of training data points input into the model to be updated. Indicates to Summing all elements in the array and Indicates the first The temperature at the preset temperature measurement points and the temperature field inside the satellite cabin in the first training data. Indicates the output of the model to be updated. The predicted temperature field corresponding to the temperature of the preset temperature measurement points in the first training data. This indicates the number of second training data points input into the model to be updated. Indicates the first The temperature of the preset temperature measurement points in the second training data. Indicates the output of the model to be updated. The predicted temperature field corresponding to the temperature of the preset temperature measurement points in the second training data. Indicates the first The uncertainty weight matrix corresponding to the temperature of the preset temperature measurement points in the second training data. Indicates the first The predicted temperature field corresponding to the temperature of the preset temperature measurement point in the second training data.

7. The method of claim 6, wherein, An uncertainty weight matrix is determined using the following equation: ； wherein, represents the prediction uncertainty corresponding to the temperature of the preset temperature measuring point in the i-th second training data, represents the prediction uncertainty corresponding to the temperature of the preset temperature measuring point in the i-th second training data, represents the uncertainty threshold.

Images