Spacecraft physical field prediction model construction method based on multi-source domain rapid migration

Abstract

The invention provides a spacecraft physical field prediction model construction method based on multi-source domain fast migration, and belongs to the field of physical field prediction.The method comprises the steps that a pre-training model is decomposed into pre-training sub-models, the pre-training sub-models are set according to initial model parameters, fitness calculation is carried out through a sample data set, and the fitness of the pre-training sub-models is calculated; obtaining an initial fitness parameter; setting the pre-trained sub-model according to a second model parameter in the second model parameter population, and performing fitness calculation through the sample data set to obtain a second fitness parameter; setting the pre-training sub-model according to an updating model parameter in the updating model parameter population, and training and optimizing based on the sample data set to obtain an optimal model parameter; and setting the pre-trained sub-models according to the optimal model parameters, and splicing to obtain a multi-source integrated model for predicting physical field data in the spacecraft cabin. According to the method, the problem of limited training samples is solved, and the accuracy of predicting the physical field data in the spacecraft cabin is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of physical field prediction, and in particular to a method for constructing a physical field prediction model of a spacecraft based on fast migration of multiple source domains. Background Art

[0002] The physical field data inside the spacecraft cabin is related to the safety and working efficiency of space missions, as well as the long-term reliable operation of the spacecraft. By predicting the physical field data, abnormal trends in the physical field inside the cabin can be detected in advance, and situations where the cabin environment deviates from the normal range can be quickly identified, so as to promptly initiate adjustment measures, provide a reliable basis for fault prevention and emergency response measures, reduce equipment failures or performance degradation caused by environmental factors, and ensure the stable operation of key systems of the spacecraft.

[0003] Deep learning plays an important role in complex physical field prediction tasks of spacecrafts with its powerful representation ability. Deep learning algorithms are good at learning complex non-linear relationships from high-dimensional and large-scale data sets, and are extremely effective in dealing with common multi-variable and non-steady state phenomena in physical fields. It can quickly process and analyze real-time data from sensors and provide near-instantaneous physical field predictions. However, the training and solution of deep learning models rely on a large number of training samples, and due to the difficulty of obtaining training samples for the physical field of spacecrafts, usually only a limited number of training samples can be obtained, which is difficult to meet the requirements for constructing a physical field prediction model.

[0004] Therefore, a method for constructing a physical field prediction model of a spacecraft based on fast migration of multiple source domains is proposed. Summary of the Invention

[0005] To solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a method for constructing a physical field prediction model of a spacecraft based on fast migration of multiple source domains, so as to solve the problem that it is difficult to obtain training samples for the physical field of spacecrafts in the traditional technology and it is difficult to meet the requirements of model training.

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

[0007] A method for constructing a physical field prediction model of a spacecraft based on fast migration of multiple source domains is provided, including:

[0008] Decompose a pre-trained model associated with the task into a number of mutually independent pre-trained sub-models;

[0009] Obtain an initial model parameter population;

[0010] Set the parameters of the pre-trained sub-models according to the initial model parameters in the initial model parameter population, and calculate the fitness through a sample data set to obtain initial fitness parameters;

[0011] Update the initial model parameter population to obtain a second model parameter population;

[0012] Set the parameters of the pre-trained sub-model according to the second model parameters in the second model parameter population, and calculate the fitness through the sample data set to obtain the second fitness parameter;

[0013] Based on the initial fitness parameter and the second fitness parameter, obtain an updated model parameter population based on the initial model parameter population and the second model parameter population;

[0014] Set the pre-trained sub-model according to the updated model parameters in the updated model parameter population, and train and optimize based on the sample data set to obtain the optimal model parameters;

[0015] Set the pre-trained sub-model according to the optimal model parameters, and splice the pre-trained sub-model to obtain a multi-source integrated model;

[0016] Use the multi-source integrated model for predicting the physical field data in the spacecraft cabin.

[0017] Furthermore, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the step: setting the parameters of the pre-trained sub-model according to the initial model parameters in the initial model parameter population, and calculating the fitness through the sample data set to obtain the initial fitness parameter; includes:

[0018] Obtain a set of pre-trained sub-models \(M = \{M 1 , M 2 , \cdots, M m \}\), assuming there are \(m\) pre-trained sub-models in total, \(M 1 \) is the first pre-trained sub-model, \(M 2 \) is the second pre-trained sub-model, \(M m \) is the \(m\)th pre-trained sub-model;

[0019] Obtain the number of modules in the pre-trained sub-model, and construct a module set for the pre-trained sub-model Assume that the first pre-trained sub-model has a total of \(n 1 \) modules, \) is the first module in the first pre-trained sub-model, \) is the second module in the first pre-trained sub-model, \) is the \(n 1 \)th module in the first pre-trained sub-model. Assume that the \(m\)th pre-trained sub-model has a total of \(n m \) modules, \) is the first module in the \(m\)th pre-trained sub-model, \) is the second module in the \(m\)th pre-trained sub-model, For the nth module in the mth pre-trained sub-model, the total number of modules in the module set is m ;

[0020] Let the initial model parameter population be P = {X 1 , X 2 , …, X p}, and the initial model parameter X i ∈ R n+m is an n + m-dimensional real vector;

[0021]

[0022] Among them, is the learning rate index value of the ith initial model parameter X i in the first module, is the learning rate index value of the ith initial model parameter X i in the second module, is the learning rate index value of the ith initial model parameter X i in the nth module, is the fusion weight for setting the first pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the first pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the second pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the mth pre-trained sub-model according to the ith initial model parameter;

[0023] Initialize the learning rate index value;

[0024]

[0025] Among them, is the lower bound of the learning rate of the ith learning rate index value in the nth module, is the upper bound of the learning rate of the ith learning rate index value in the nth module, and δ is a random variable;

[0026] Obtain the initial model parameter in the initial model parameter population and set the parameters for the pre-trained sub-model. The pre-trained sub-model is trained on the training set based on the initial model parameter X i and calculate the initial fitness parameter f i on the validation set;

[0027]

[0028] Among them, the validation set is D val = {(x k ​,y j )}, w i is the influence weight of the i-th pre-trained sub-model, is the i-th pre-trained sub-model that sets the sub-model parameters based on the initial model parameters.

[0029] Furthermore, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the step: updating the initial model parameter population to obtain a second model parameter population; includes:

[0030] Randomly select model parameter values from the initial model parameter population based on the mutation operator, update the initial model parameters, and obtain a first set of model parameters;

[0031] X′ i = X a + γ · (X b - X c )

[0032] where X′ i is the first model parameter after updating the initial model parameter X i , X a , X b , X c are the model parameter values randomly selected from the initial model parameter population, and γ is a parameter random variable;

[0033] Based on the crossover operator, perform crossover mutation according to the model parameters in the first set of model parameters and the initial set of model parameters to obtain a second set of model parameters;

[0034]

[0035] where, is the learning rate index update value of the i-th second model parameter X i ″ in the n-th module, is the learning rate index update value of the i-th first model parameter X′ i in the n-th module, δ′ is the first random variable, σ CR is the crossover probability factor, and n rand is the random component.

[0036] Furthermore, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the step: randomly select model parameter values from the initial model parameter population based on the mutation operator, update the initial model parameters, and obtain a first set of model parameters; then includes:

[0037] Perform a constraint check on the first model parameters in the first set of model parameters. When the value of an element in the first model parameters violates the coding specification, randomly generate a new element value again.

[0038] Further, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the step: obtaining an updated model parameter population based on the initial model parameter population and the second model parameter population according to the initial fitness parameter and the second fitness parameter; includes:

[0039] According to the initial fitness parameter and the second fitness parameter, construct an updated model parameter population based on the second set of model parameters and the initial model parameter population through a selection operator;

[0040]

[0041] Wherein, X i ″′ is the i-th updated model parameter in the updated model parameter population, and f i ″ The pre-trained sub-model is trained on the training set based on the updated model parameter X i ″ and calculate the second fitness parameter obtained on the validation set.

[0042] Further, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the sample data set includes a training set and a validation set.

[0043] Further, in the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by the present invention, the step: setting the pre-trained sub-model according to the updated model parameters in the updated model parameter population, and performing training optimization based on the sample data set to obtain the optimal model parameters; includes:

[0044] Set the parameters of the pre-trained sub-model according to the updated model parameters, update and optimize the parameters of the pre-trained sub-model through a loss function, and perform training based on the sample data set until the maximum number of iterations is reached to obtain the optimal model parameters;

[0045]

[0046] Wherein, is the parameter after t updates of the i-th parameter θ i of the pre-trained sub-model, is the parameter after t + 1 updates of the parameter θ i , α is the learning rate, is the gradient parameter of the loss function J(θ) with respect to the parameter θ i obtained by using chain differentiation.

[0047] The main advantages of the technical solution of the present invention are as follows:

[0048] In the method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration of the present invention, the training of the pre-trained sub-model is processed based on the sample data set through the initial model parameter population, the second model parameter population, and the updated model parameter population, realizing the reuse of the sample data set and solving the problem of limited model training samples in the traditional technology. The second model parameter population is obtained by cross-variation of the initial model parameter population, and the updated model parameter population is obtained based on the selection operator according to the initial model parameter population and the second model parameter population. By expanding the model parameter population and obtaining the optimal model parameters under the condition of small samples, the constructed multi-source integration model has a high prediction accuracy, reduces the demand for a large number of samples in model training and solution, and further realizes the prediction of the physical field data inside the spacecraft cabin. Brief Description of the Drawings

[0049] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of the present invention, and the schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0050] Figure 1 It is a schematic flow chart of a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration provided by an embodiment of the present invention. Detailed Embodiments

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0052] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the drawings.

[0053] Embodiment 1:

[0054] The embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration, referring to Figure 1 , including:

[0055] Decompose the pre-trained model associated with the task into a number of independent pre-trained sub-models;

[0056] Obtain the initial model parameter population;

[0057] Set the parameters of the pre-trained sub-model according to the initial model parameters in the initial model parameter population, and calculate the fitness through the sample data set to obtain the initial fitness parameters;

[0058] Update the initial model parameter population to obtain the second model parameter population;

[0059] Set the parameters of the pre-trained sub-model according to the second model parameters in the second model parameter population, and calculate the fitness through the sample data set to obtain the second fitness parameters;

[0060] Based on the initial fitness parameters and the second fitness parameters, and based on the initial model parameter population and the second model parameter population, obtain the updated model parameter population;

[0061] Set the pre-trained sub-model according to the updated model parameters in the updated model parameter population, and train and optimize based on the sample data set to obtain the optimal model parameters;

[0062] Set the pre-trained sub-model according to the optimal model parameters, and splice the pre-trained sub-models to obtain a multi-source integrated model;

[0063] Use the multi-source integrated model for the prediction of the physical field data inside the spacecraft cabin.

[0064] In the above embodiments, the pre-trained model is decomposed into several independent pre-trained sub-models. Set the parameters of the pre-trained sub-model according to the initial model parameters in the initial model parameter population, and calculate the fitness through the sample data set to obtain the initial fitness parameters;

[0065] Update the initial model parameter population to obtain the second model parameter population; Set the parameters of the pre-trained sub-model according to the second model parameters in the second model parameter population, and calculate the fitness through the sample data set to obtain the second fitness parameters;

[0066] Based on the initial fitness parameters and the second fitness parameters, and based on the initial model parameter population and the second model parameter population, obtain the updated model parameter population.

[0067] In the above embodiments, based on the updated model parameters in the updated model parameter population, set and optimize the pre-trained sub-model based on the sample data set to obtain the optimal model parameters; Set the pre-trained sub-model according to the optimal model parameters, and splice the pre-trained sub-models to obtain a multi-source integrated model, and use the multi-source integrated model for the prediction of the physical field data inside the spacecraft cabin.

[0068] In the above embodiments, a multi-source transfer learning search space is constructed according to the initial model parameter population, the second model parameter population, and the updated model parameter population, and the accuracy of obtaining the optimal model parameters is improved by increasing the granularity of model module division and expanding the selectable range of the learning rate.

[0069] In the above embodiments, the differential evolution method is used to search for the optimal transfer learning strategy. First, the multi-source transfer learning strategy is encoded, and then the differential evolution algorithm module is used to efficiently optimize the strategy in the encoded strategy space.

[0070] The beneficial effects of the above technology are as follows: The pre-trained sub-model configures parameters according to the initial model parameter population and the second model parameter population, and is trained based on the sample data set to obtain the initial fitness parameter and the second fitness parameter, and then obtains the updated model parameter population. The pre-trained sub-model configures parameters according to the updated model parameter population and is trained based on the sample data set, realizing the acquisition of the optimal model parameters; setting and splicing the pre-trained sub-model according to the optimal model parameters realizes the acquisition of the multi-source integrated model, and the multi-source integrated model is used for the prediction of the physical field data in the spacecraft cabin; in the above method, the training of the pre-trained sub-model based on the initial model parameter population, the second model parameter population, and the updated model parameter population is processed based on the sample data set, realizing the reuse of the sample data set, solving the problem of limited model training samples in the traditional technology. The second model parameter population is obtained by crossover and mutation of the initial model parameter population, and the updated model parameter population is obtained based on the selection operator according to the initial model parameter population and the second model parameter population. By expanding the model parameter population, the optimal model parameters are obtained under the condition of small samples, so that the constructed multi-source integrated model has a high prediction accuracy, reducing the demand for a large number of samples in model training and solution, and thus realizing the prediction of the physical field data in the spacecraft cabin.

[0071] Embodiment 2:

[0072] The embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast transfer, steps: setting parameters for the pre-trained sub-model according to the initial model parameters in the initial model parameter population, and calculating the fitness through the sample data set to obtain the initial fitness parameter; including:

[0073] Obtain a set of pre-trained sub-models M = {M 1 , M 2 , …, M m}, assuming there are m pre-trained sub-models in total, M 1 is the first pre-trained sub-model, M 2 is the second pre-trained sub-model, M m is the m-th pre-trained sub-model;

[0074] Obtain the number of modules in the pre-trained sub-model and construct a set of modules for the pre-trained sub-model Suppose the first pre-trained sub-model has a total of n 1 modules, is the first module in the first pre-trained sub-model, is the second module in the first pre-trained sub-model, is the nth module in the first pre-trained sub-model, 1 Suppose the mth pre-trained sub-model has a total of n m modules, is the first module in the mth pre-trained sub-model, is the second module in the mth pre-trained sub-model, is the nth module in the mth pre-trained sub-model, m The module set has a total of modules;

[0075] Suppose the initial model parameter population is P = {X 1 , X 2 , …, X p}, and the initial model parameter X i ∈R n+m is an n + m-dimensional real vector;

[0076]

[0077] Among them, is the learning rate index value of the ith initial model parameter X i in the first module, is the learning rate index value of the ith initial model parameter X i in the second module, is the learning rate index value of the ith initial model parameter X i in the nth module, is the fusion weight for setting the first pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the first pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the second pre-trained sub-model according to the ith initial model parameter, is the fusion weight for setting the mth pre-trained sub-model according to the ith initial model parameter;

[0078] Initialize the learning rate index value;

[0079]

[0080] Among them, is the lower bound of the learning rate for the i-th learning rate index value in the n-th module, is the upper bound of the learning rate for the i-th learning rate index value in the n-th module, and δ is a random variable, where δ ∈ [0, 1];

[0081] Obtain the initial model parameters in the initial model parameter population and set the parameters of the pre-trained sub-model. The pre-trained sub-model is based on the initial model parameter X i Train on the training set and calculate the initial fitness parameter f on the validation set i ;

[0082]

[0083] where the validation set is D val ={(x j , y j )}, and w i is the influence weight of the i-th pre-trained sub-model, is to set the sub-model parameter θ based on the initial model parameter * of the i-th pre-trained sub-model

[0084] In the above embodiments, obtain the set of pre-trained sub-models and the number of modules in each pre-trained sub-model, and construct a module set; construct an initial model parameter population and perform an initial setting on the learning rate index value

[0085] In the above embodiments, obtain the initial model parameters in the initial model parameter population and set the parameters of the pre-trained sub-model. The pre-trained sub-model is trained on the training set based on the model parameters and calculates the initial fitness parameter on the validation set

[0086] In the above embodiments, perform a decoding operation on the initial model parameters in the initial model parameter population, obtain the corresponding multi-source transfer learning strategy, and calculate the initial fitness parameter of the pre-trained sub-model on the training set and the validation set

[0087] In the above embodiments, the learning rate index value is a continuous real number, and the value range is [1, k + 1], where k is the number of optional learning rates for the module

[0088] In the above embodiments, the fusion weight is a real number in the interval [0, 1]

[0089] Embodiment 3:

[0090] The embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast transfer. The steps are as follows: update the initial model parameter population to obtain a second model parameter population; including:

[0091] Randomly select model parameter values from the initial model parameter population based on the mutation operator, update the initial model parameters, and obtain the first set of model parameters;

[0092] X′ i = X a + γ · (X b - X c )

[0093] where X′ i is the first set of model parameters obtained by updating the initial model parameters X i , X a , X b , X c are model parameter values randomly selected from the initial model parameter population, and γ is a parameter random variable;

[0094] Based on the crossover operator, perform crossover mutation according to the model parameters in the first set of model parameters and the initial set of model parameters to obtain the second set of model parameters;

[0095]

[0096] where is the update value of the learning rate index of the i-th second model parameter X i "" in the n-th module, is the update value of the learning rate index of the i-th first model parameter X′ i in the first set of model parameters in the n-th module, δ′ is the first random variable, and σ CR is the crossover probability factor, and n rand is the random component.

[0097] In the above embodiments, the initial model parameters are updated by the mutation operator to obtain the first set of model parameters, and the crossover operator is used to perform crossover mutation according to the first set of model parameters and the initial set of model parameters to obtain the second set of model parameters, and the second set of model parameters is constructed.

[0098] The beneficial effect of the above technology is that the initial model parameters are updated by the mutation operator and the crossover operator to obtain the second set of model parameters, increasing the diversity of the population.

[0099] Embodiment 4:

[0100] The embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration. The steps are as follows: Randomly select model parameter values from the initial model parameter population based on the mutation operator, update the model parameters, and obtain the first set of model parameters; then include:

[0101] Perform a constraint check on the first model parameters in the first set of model parameters. When the value of an element in the first model parameters violates the coding specification, randomly generate a new element value again.

[0102] In the above embodiments, a constraint check is performed on the value of an element in the first model parameters, and when the coding specification is violated, a new element value is randomly generated to prevent the first model parameters obtained from exceeding the value range.

[0103] Embodiment 5:

[0104] An embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration. The steps are as follows: Based on the initial fitness parameters and the second fitness parameters, obtain an updated model parameter population based on the initial model parameter population and the second model parameter population, including:

[0105] Based on the initial fitness parameters and the second fitness parameters, construct an updated model parameter population through a selection operator based on the second set of model parameters and the initial model parameter population.

[0106]

[0107] Wherein, X″′i is the i-th updated model parameter in the updated model parameter population, and f″ i The pre-trained sub-model is based on the updated model parameter X″ i Train on the training set and calculate the second fitness parameters obtained on the validation set.

[0108] In the above embodiments, based on the initial fitness parameters and the second fitness parameters, an updated model parameter population is obtained through a selection operator based on the second set of model parameters and the initial model parameter population.

[0109] The beneficial effect of the above technology is that: by using the selection operator to combine the second set of model parameters and the initial model parameter population, and adopting a greedy strategy according to the newly generated second fitness parameters, select better model parameters as the updated model parameters.

[0110] Embodiment 6:

[0111] An embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration, a sample data set, including a training set and a validation set.

[0112] In the above embodiments, the sample data set is randomly divided into a training set and a validation set.

[0113] Embodiment 7:

[0114] An embodiment of the present invention provides a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration. The steps are as follows: Set the pre-trained sub-model according to the updated model parameters in the updated model parameter population, and perform training optimization based on the sample data set to obtain the optimal model parameters. The steps include:

[0115] Set the parameters of the pre-trained sub-model according to the updated model parameters, update and optimize the parameters of the pre-trained sub-model through the loss function, and perform training based on the sample data set until the maximum number of iterations is reached to obtain the optimal model parameters.

[0116]

[0117] Wherein, is the parameter after t updates of the i-th parameter θ of the pre-trained sub-model i , is the parameter after t + 1 updates of the parameter θ i , α is the learning rate, is the gradient parameter of the loss function J(θ) with respect to the parameter θ i obtained by using chain derivation.

[0118] In the above embodiment, the parameters of the pre-trained sub-model are updated and optimized, and the single update step size of the parameters is controlled by the learning rate until the maximum number of iterations is reached to obtain the optimal model parameters.

[0119] In the above embodiment, by setting a relatively low value for the learning rate, the original parameter information can be retained to a large extent.

[0120] In the above embodiment, by setting the learning rate to 0, the parameters of the pre-trained sub-model are frozen, and the parameter update and optimization operation is stopped.

[0121] The beneficial effect of the above technology is that: the parameters of the pre-trained sub-model are set through the updated model parameter population, and the parameters are updated and optimized based on the loss function. When it is judged that the maximum number of iterations is reached, the optimal model parameters are obtained.

[0122] In the above embodiment, a method for constructing a spacecraft physical field prediction model based on multi-source domain fast migration proposed by the present invention utilizes the prior knowledge contained in the surrogate models constructed by predicting the physical fields in the cabins of multiple other spacecrafts to assist in the training and solution of the surrogate model for predicting the physical field in the cabin of the current spacecraft, so as to reduce the demand for a large number of training samples in deep learning.

[0123] It should be noted that in this text, 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 terms "comprising", "including" or any other variant thereof are 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.

[0124] 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 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 spacecraft physical field prediction model based on rapid migration of multi-source domains, characterized in that: include: Decompose the pre-trained model associated with the task into several independent pre-trained sub-models; Get the initial model parameter population; The parameters of the pre-trained sub-model are set according to the initial model parameters in the initial model parameter population, and the fitness is calculated through the sample data set to obtain the initial fitness parameters; Updating the initial model parameter population to obtain a second model parameter population; Setting parameters of the pre-trained sub-model according to the second model parameters in the second model parameter population, and performing fitness calculation through the sample data set to obtain the second fitness parameters; According to the initial fitness parameter and the second fitness parameter, based on the initial model parameter population and the second model parameter population, obtaining an updated model parameter population; The pre-trained sub-model is set according to the updated model parameters in the updated model parameter population, and the optimal model parameters are obtained by training and optimizing based on the sample data set; The pre-trained sub-models are set according to the optimal model parameters, and the pre-trained sub-models are spliced ​​to obtain a multi-source integrated model; The multi-source integrated model is used to predict the physical field data in the spacecraft cabin.

2. The method for constructing a spacecraft physical field prediction model based on rapid migration of multi-source domains according to claim 1 is characterized in that: The step of setting parameters of the pre-trained sub-model according to the initial model parameters in the initial model parameter population, and performing fitness calculation through the sample data set to obtain the initial fitness parameters includes: Get the pre-trained sub-model set M = {M1, M2, ..., M m }, suppose there are m pre-trained sub-models, M1 is the first pre-trained sub-model, M2 is the second pre-trained sub-model, M m is the mth pre-trained sub-model; Get the number of modules in the pre-trained sub-model and build a module set about the pre-trained sub-model Assume that the first pre-trained sub-model has a total of n1 modules. is the first module in the first pre-trained sub-model, is the second module in the first pre-trained sub-model. is the n1th module in the first pre-trained sub-model, and the mth pre-trained sub-model has n m modules, is the first module in the mth pre-trained sub-model, is the second module in the mth pre-trained sub-model, is the nth in the mth pre-trained sub-model m modules, the module collection has modules; Assume that the initial model parameter population is P = {X1, X2, ..., X p }, initial model parameters X i ∈R n+m is a real vector of dimension n+m; in, is the i-th initial model parameter X i The learning rate index value in the first module, is the i-th initial model parameter X i The learning rate index value in the second module, is the i-th initial model parameter X i The learning rate index value in the nth module, To set the fusion weight of the first pre-trained sub-model according to the i-th initial model parameters, To set the fusion weight of the first pre-trained sub-model according to the i-th initial model parameters, To set the fusion weight of the second pre-trained sub-model according to the i-th initial model parameters, To set the fusion weight of the mth pre-trained sub-model according to the i-th initial model parameters; Initializing the learning rate index value; in, is the lower bound of the learning rate of the i-th learning rate index value in the n-th module, is the upper bound of the learning rate of the i-th learning rate index value in the n-th module, and δ is a random variable; Get the initial model parameters in the initial model parameter population to set the parameters of the pre-trained sub-model. The pre-trained sub-model is based on the initial model parameters X i Train on the training set and calculate the initial fitness parameter f on the validation set i ; The validation set is D val ={(x j ,y j )},w i is the influence weight of the i-th pre-trained sub-model, Set the i-th pre-trained sub-model parameters based on the initial model parameters.

3. The method for constructing a spacecraft physical field prediction model based on rapid migration of multiple source domains according to claim 1, characterized in that: The step of updating the initial model parameter population to obtain a second model parameter population comprises: randomly selecting model parameter values ​​from the initial model parameter population based on a mutation operator, updating the initial model parameters, and obtaining a first model parameter set; X′ i =X a +γ·(X b -X c ) Among them, X′ i is the initial model parameter X i The first model parameter after updating, X a , X b , X c The model parameter value randomly selected from the initial model parameter population, γ is the parameter random variable; Based on the crossover operator, crossover mutation is performed according to the model parameters in the first model parameter set and the initial model parameter set to obtain a second model parameter set; in, is the i-th second model parameter X in the second model parameter set i ″The learning rate index update value in the nth module, is the i-th first model parameter X in the first model parameter set i ′ The learning rate index update value in the nth module, δ ′ is the first random variable, σ CR is the crossover probability factor, n rand is a random component.

4. The method for constructing a spacecraft physical field prediction model based on rapid migration of multiple source domains according to claim 3 is characterized in that: The step of randomly selecting model parameter values ​​from the initial model parameter population based on a mutation operator, updating the initial model parameters, and obtaining a first model parameter set; then comprising: A constraint check is performed on the first model parameter in the first model parameter set, and when a value of an element in the first model parameter violates a coding specification, the element value is randomly regenerated.

5. The method for constructing a spacecraft physical field prediction model based on rapid migration of multiple source domains according to claim 3 is characterized in that: The step of obtaining an updated model parameter population based on the initial model parameter population and the second model parameter population according to the initial fitness parameter and the second fitness parameter comprises: According to the initial fitness parameter and the second fitness parameter, constructing an updated model parameter population based on the second model parameter set and the initial model parameter population by selecting an operator; Among them, X i ″′ is the i-th updated model parameter in the updated model parameter population, f i The pre-trained sub-model is based on updating the model parameters X i "Training is performed on the training set, and the second fitness parameter obtained is calculated on the validation set.

6. The method for constructing a spacecraft physical field prediction model based on rapid migration of multiple source domains according to claim 2, characterized in that: The sample data set includes a training set and a validation set.

7. The method for constructing a spacecraft physical field prediction model based on rapid migration of multiple source domains according to claim 1, characterized in that: The step of setting the pre-trained sub-model according to the updated model parameters in the updated model parameter population, training and optimizing based on the sample data set, and obtaining the optimal model parameters comprises: The parameters of the pre-trained sub-model are set according to the updated model parameters, the parameters of the pre-trained sub-model are updated and optimized through the loss function, and training is performed based on the sample data set until the maximum number of iterations is reached to obtain the optimal model parameters; in, is the i-th parameter θ of the pre-trained sub-model i The parameters after t updates, For the parameter θ i The parameters after t+1 updates, α is the learning rate, To use chain derivation to obtain the loss function J(θ) with respect to the parameter θ i The gradient parameter.

Images