CN115730509A

CN115730509A - Spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning

Info

Publication number: CN115730509A
Application number: CN202211143594.2A
Authority: CN
Inventors: 姚雯; 龚智强; 周炜恩; 李星辰; 陈小前
Original assignee: National Defense Technology Innovation Institute PLA Academy of Military Science
Current assignee: National Defense Technology Innovation Institute PLA Academy of Military Science
Priority date: 2022-09-20
Filing date: 2022-09-20
Publication date: 2023-03-03
Anticipated expiration: 2042-09-20
Also published as: CN115730509B

Abstract

The invention discloses a spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning, which comprises the following steps of: constructing a task optimization problem for reconstructing a temperature field in a spacecraft cabin; determining boundary conditions, power distribution and structure of the heat source assembly; selecting temperature monitoring points for arranging temperature sensors in the spacecraft cabin; acquiring at least one of first training data, second training data, third training data and fourth training data; determining the mapping relation between any point position in the layout area and the temperature of the point by utilizing training data through an interpolation method, or training a traditional machine learning model or a neural network, or training a multilayer perceptron, or training a deep neural network model, or training a convolutional neural network model of a graph; and reconstructing the temperature field according to the mapping relation or the trained model. The invention can realize real-time, rapid and high-precision prediction and reconstruction of the temperature of other positions and/or the temperature field of the whole area by using the limited temperature monitoring point data.

Description

Spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning

Technical Field

The invention relates to the technical field of spacecraft thermal control, in particular to a benchmark method for researching a spacecraft cabin internal temperature field reconstruction task based on machine learning.

Background

When the electronic equipment (or components) in the spacecraft cabin works, heat is inevitably dissipated, so that the temperature of the electronic equipment (or components) is increased, and the electronic equipment (or components) cannot dissipate heat outwards through heat convection due to the vacuum environment in space, so that heat is easily accumulated, the temperature of the equipment (or components) is rapidly increased, the service life, the safety and the reliability of the equipment (or components) are influenced, and the equipment is seriously even failed or damaged. Research has shown that the failure rate of some electronic devices (or components) increases exponentially with temperature, typically doubling the failure rate for every 10 ℃ increase in temperature. In order for the electronic equipment (or components) in the spacecraft cabin to operate within the allowable temperature range, reasonable thermal control design and temperature monitoring and control must be performed on the electronic equipment (or components).

In order to implement thermal control design and temperature monitoring and control, a suitable position point is generally selected on an electronic device (or component) in a spacecraft cabin, a temperature sensor (such as a thermal resistor) is arranged for on-rail temperature monitoring, temperature data is converted into an electric signal and then transmitted to a ground control center, and thermal control design and temperature monitoring and control are performed based on the acquired temperature data.

In order to ensure the accuracy and reliability of thermal control design and temperature monitoring and control, it is currently common practice to arrange as many temperature sensors as possible to monitor as many electronic devices (or components) and cabin temperature conditions as possible. However, the number of the temperature sensors is directly related to the requirements of the spacecraft platform on hardware resources such as temperature monitoring sensing devices, measurement channels, cables and the like, and if the number of the temperature monitoring points is too large, resource consumption, system weight increase, development cost increase of engineering implementation, testing and the like are inevitably caused, so that in the actual application process, the number of the temperature monitoring points for arranging the temperature sensors is limited, and electronic equipment (or components) which cannot be monitored possibly exist. Also, since the temperature sensor is only able to monitor a local small range of temperatures, it is difficult to measure the overall temperature profile of the device (or component). Therefore, how to obtain the temperatures of other positions and attention points and even the overall temperature field of the area where the electronic device (or component) is located according to the temperature data of the limited temperature monitoring points is a critical problem to be solved when performing thermal control on the spacecraft at present.

How to obtain the temperature of other points of interest and even the overall temperature field of the area where the electronic device (or component) is located according to the temperature data of the limited temperature monitoring points can also be called as a temperature field reconstruction problem. Aiming at the problem of temperature field reconstruction, the traditional method accumulates data through a ground test stage, and then estimates the data based on the temperature data of temperature monitoring points by combining an empirical formula, an interpolation algorithm and other continuous modeling methods so as to complete the temperature field reconstruction. However, in practical applications, the conventional method has at least the following problems: the calculation efficiency is too low, a large amount of iteration is needed in numerical calculation, and real-time prediction is difficult to realize; the problem of high-dimensional modeling cannot be solved, the reconstruction precision is poor, and the requirements of actual engineering problems cannot be met; because of the gap between engineering and scientific research, the reconstruction of the temperature field in engineering is usually realized only by using methods such as finite elements, but the method is simple, but consumes a large amount of resources, and cannot effectively and continuously promote the research in the reconstruction of the temperature field.

Disclosure of Invention

In order to solve part or all of technical problems in the prior art, the invention provides a basic method for researching a spacecraft cabin temperature field reconstruction task based on machine learning.

The technical scheme of the invention is as follows:

a benchmark method for researching a spacecraft in-cabin temperature field reconstruction task based on machine learning is provided, and the method comprises the following steps:

constructing a task optimization problem for reconstructing a temperature field in the spacecraft cabin according to a layout structure in the spacecraft cabin;

determining boundary conditions of a layout area in the spacecraft cabin and power distribution conditions and structures of the heat source components;

selecting temperature monitoring points for arranging temperature sensors in the spacecraft cabin by utilizing a cluster analysis algorithm;

acquiring at least one of first training data, second training data, third training data and fourth training data, wherein the first training data comprises the position and the temperature of a temperature monitoring point in a layout area, the second training data comprises the temperature monitoring point in the layout area and the temperature of a set attention point, the third training data comprises the position and the temperature of the temperature monitoring point in the layout area and a temperature field corresponding to the layout area, and the fourth training data comprises the position and the temperature of the temperature monitoring point in the layout area and the position and the temperature of the set attention point in the layout area;

determining the mapping relation from the position of any point in the layout area to the temperature of the point by an interpolation method by utilizing the first training data; or training a constructed traditional machine learning model or a neural network model by utilizing first training data to fit a mapping relation from the position of any point in the layout area to the temperature of the point; or training the constructed multilayer perceptron by utilizing second training data to fit the mapping relation from the temperature of the temperature monitoring points in the layout area to the temperature of the set attention points; or generating a corresponding temperature monitoring matrix according to the position and temperature information of the temperature monitoring points in the layout area in the third training data, and fitting a mapping relation from the temperature monitoring matrix to the temperature field by using a deep neural network model constructed by the temperature monitoring matrix and the temperature field training in the third training data; or a graph model is constructed according to the position relation between the temperature monitoring points and the set attention points, and a graph convolution neural network model constructed by utilizing the graph model, the temperature of the temperature monitoring points in the fourth training data and the temperature training of the set attention points is utilized to fit the mapping relation between the graph model and the temperature of the temperature monitoring points to the temperature of the set attention points;

and reconstructing the temperature field according to the determined mapping relation or the trained model.

In some possible implementations, it is set that: the layout area in the spacecraft cabin is provided with a plurality of heat source components, the power distribution of the ith heat source component is phi _i (x, y) M temperature monitoring points provided with temperature sensors are arranged in the layout area, and the position of the mth temperature monitoring point is

Based on the setting, the task optimization problem of reconstructing the temperature field in the spacecraft cabin is as follows:

wherein T represents the temperature field of the reconstructed layout region,

in the temperature field representing the reconstruction

Temperature of the location, O _m The temperature monitoring value of the mth temperature monitoring point is shown, k represents the heat conduction coefficient, x and y represent the position coordinate of a certain point in the temperature field, T ₀ Denotes the temperature value at the isothermal boundary, n denotes the normal perpendicular to, h denotes the thermal convection coefficient, T = T ₀ Indicating the Dirichlet boundary conditions and,

indicating the conditions of the Neumann boundaries,

representing the Robin boundary conditions.

In some possible implementations, the selecting, by using a cluster analysis algorithm, a temperature monitoring point for arranging a temperature sensor in the spacecraft cabin includes:

step S31, randomly sampling a power from the power distribution of each heat source assembly as the actual power of the corresponding heat source assembly for each heat source assembly to obtain the layout of a specific working condition, and repeating the random sampling process for multiple times to obtain the layouts of multiple different working conditions;

step S32, dividing the layout area into N ₁ ×N ₂ The temperature control method comprises the following steps that grids can only share one constant temperature value in each grid, and a plurality of temperature fields corresponding to the layouts of a plurality of different working conditions are obtained through calculation by using a finite element method based on the layout areas after the grids are divided;

step S33, the number of temperature monitoring points is specified, and N with multiple dimensions is obtained by utilizing a K-Means clustering algorithm ₁ ×N ₂ Clustering the temperature points in the spacecraft cabin to obtain a plurality of categories with the same number as the designated temperature monitoring points, calculating generalized distances based on Euclidean distance measurement, selecting the temperature point closest to the clustering center of each category as a representative of the corresponding category, and taking the selected temperature point and the corresponding position as the temperature monitoring point and the corresponding position.

In some possible implementations, the first training data is obtained by:

determining the position of a temperature monitoring point in a layout area;

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component aiming at each heat source component in the layout area to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point under the layout of the current specific working condition to obtain first training data comprising the position and the temperature of the temperature monitoring point in the layout area, and repeating the process for multiple times until first training data with the preset quantity are obtained;

acquiring the second training data in the following way:

determining the positions of the temperature monitoring points and the set attention points in the layout area;

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component aiming at each heat source component in the layout area to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point and each set attention point under the layout of the current specific working condition to obtain second training data comprising the temperature monitoring points and the temperature of the set attention points in the layout area, and repeating the process for multiple times until second training data with preset quantity are obtained;

obtaining the third training data in the following manner:

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition aiming at each heat source component in the layout area, carrying out simulation analysis on the layout of the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, extracting the temperature of a temperature monitoring point from the obtained temperature field to obtain third training data comprising the position and the temperature of the temperature monitoring point in the layout area and the temperature field corresponding to the layout area, and repeating the process for multiple times until the third training data with the preset number are obtained;

acquiring the fourth training data in the following way:

determining the positions of temperature monitoring points and set attention points in a layout area;

for each heat source assembly in the layout area, randomly sampling a power value from the power distribution of each heat source assembly as the current power of the corresponding heat source assembly to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point and each set attention point under the layout of the current specific working condition, obtaining fourth training data comprising the position and the temperature of the temperature monitoring point in the layout area and the position and the temperature of the set attention point in the layout area, and repeating the process for multiple times until obtaining the preset number of fourth training data.

In some possible implementations, the interpolation method is one of a k-nearest neighbor nonlinear interpolation method and a global gaussian interpolation method;

when the interpolation method is a k-nearest neighbor nonlinear interpolation method, the mapping relationship between the position of any point in the layout area and the temperature of the point can be expressed as:

when the interpolation method is a global gaussian interpolation method, the mapping relationship from the position of any point in the layout area to the temperature of the point can be expressed as:

wherein, T (x) ₀ ,y ₀ ) Indicates the set point of interest (x) ₀ ,y ₀ ) Predicted temperature value of S _k (x ₀ ,y ₀ ) Indicates a distance setting focus (x) ₀ ,y ₀ ) The most recent k temperature monitoring points are,

indicating the ith temperature monitoring point, M indicating the total number of temperature monitoring points in the layout area,

indicating the jth temperature monitoring point at which,

representing the temperature value of the ith temperature monitoring point.

In some possible implementations, the traditional machine learning model is one of a polynomial regression model, a random forest regression model, a gaussian process regression model, and a support vector machine regression model.

In some possible implementations, the neural network model is one of a multi-layer perceptron, a constrained boltzmann machine, and a deep belief network;

the training of the neural network model using the first training data comprises:

and taking the position coordinates of the temperature monitoring points in the first training data as the input of the neural network model, and taking the temperature of the temperature monitoring points in the first training data as the output of the neural network model, thereby training the neural network model.

In some possible implementations, the training the built multi-layered perceptron with the second training data includes:

and taking the temperature vector of the temperature monitoring point in the second training data as the input of the multilayer perceptron, and taking the temperature vector of the set attention point in the second training data as the output of the multilayer perceptron to train the multilayer perceptron.

In some possible implementations, the deep neural network model is one of a full convolution neural network model, a feature pyramid network model, a SegNet network model, and a UNet network model;

according to the position and the temperature information of the temperature monitoring points in the layout area in the third training data, generating a corresponding temperature monitoring matrix by adopting the following mode:

discretely dividing layout area into M ₁ ×M ₂ Each grid can only share one constant temperature value;

according to the layout area after grid division, in the grids with the temperature monitoring points, the temperature of the temperature monitoring points is used as a matrix element, in the grids without the temperature monitoring points, zero is used as a matrix element, and the generated dimension is M ₁ ×M ₂ The temperature monitoring matrix of (a);

the deep neural network model trained and constructed by utilizing the temperature monitoring matrix and the temperature field in the third training data comprises:

and taking the temperature monitoring matrix corresponding to the third training data as the input of the deep neural network model, taking the temperature field in the third training data as the output of the deep neural network model, and training the deep neural network model.

In some possible implementations, the constructing the graph model according to the position relationship between the temperature monitoring points and the set attention points includes:

taking a temperature monitoring point as a node, taking a set concern point as a node, and determining all nodes corresponding to all position points;

determining the actual distance between each node, and if the actual distance between two nodes is smaller than a preset distance threshold, adding an undirected edge between the two nodes;

constructing a corresponding graph model according to all the determined nodes and edges;

the graph convolution neural network model constructed by utilizing the graph model, the temperature of the temperature monitoring point in the fourth training data and the temperature training of the set attention point comprises the following steps:

and taking the temperature of the temperature monitoring point in the fourth training data as the input characteristic of the corresponding node in the graph model, taking the graph model which comprises the node input characteristic and corresponds to the fourth training data as the input of the graph convolution neural network model, and taking the temperature of the set attention point in the fourth training data as the output of the graph convolution neural network model, thereby training the graph convolution neural network model.

The technical scheme of the invention has the following main advantages:

the spacecraft in-cabin temperature field reconstruction task research benchmark method based on machine learning can utilize data of limited temperature monitoring points to quickly predict the temperatures of other position points and/or the whole temperature field of the whole area, realize real-time, quick and high-precision reconstruction of the spacecraft in-cabin temperature field, have fewer required temperature monitoring points, effectively reduce the demand of a spacecraft platform on related hardware resources, reduce resource consumption and equipment cost, fully assist a spacecraft in temperature monitoring and control, and promote the research of spacecraft in-cabin temperature field reconstruction.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

FIG. 1 is a flowchart of a benchmark method for researching a temperature field reconstruction task in a spacecraft cabin based on machine learning according to an embodiment of the invention;

FIG. 2 is a schematic diagram of a network architecture of a multi-tier perceptron according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a network architecture of a constrained Boltzmann machine according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a graph model constructed according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the specific embodiments of the present invention and the accompanying drawings. It is to be understood that the described embodiments are only some of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.

The technical scheme provided by the embodiment of the invention is explained in detail in the following with the accompanying drawings.

Referring to fig. 1, an embodiment of the present invention provides a benchmark method for researching a spacecraft cabin temperature field reconstruction task based on machine learning, where the method includes the following steps:

s1, constructing a task optimization problem for reconstructing a temperature field in a spacecraft cabin according to a layout structure in the spacecraft cabin;

s2, determining boundary conditions of a layout area in the spacecraft cabin and the power distribution condition and structure of the heat source assembly;

s3, selecting temperature monitoring points for arranging temperature sensors in the spacecraft cabin by using a cluster analysis algorithm;

s4, acquiring at least one of first training data, second training data, third training data and fourth training data, wherein the first training data comprise the positions and the temperatures of the temperature monitoring points in the layout area, the second training data comprise the temperatures of the temperature monitoring points in the layout area and the set attention points, the third training data comprise the positions and the temperatures of the temperature monitoring points in the layout area and the temperature fields corresponding to the layout area, and the fourth training data comprise the positions and the temperatures of the temperature monitoring points in the layout area and the positions and the temperatures of the set attention points in the layout area;

s5, determining a mapping relation from the position of any point in the layout area to the temperature of the point by an interpolation method by using the first training data; or training a constructed traditional machine learning model or a neural network model by utilizing first training data to fit a mapping relation from the position of any point in the layout area to the temperature of the point; or training the constructed multilayer perceptron by utilizing second training data to fit the mapping relation from the temperature of the temperature monitoring points in the layout area to the temperature of the set attention point; or generating a corresponding temperature monitoring matrix according to the position and temperature information of the temperature monitoring points in the layout area in the third training data, and utilizing the temperature monitoring matrix and a deep neural network model constructed by the temperature field training in the third training data to fit the mapping relation from the temperature monitoring matrix to the temperature field; or a graph model is constructed according to the position relation between the temperature monitoring points and the set attention points, and a graph convolution neural network model constructed by training is trained by utilizing the temperature of the temperature monitoring points in the graph model and the fourth training data and the temperature of the set attention points so as to fit the mapping relation between the graph model and the temperature of the temperature monitoring points to the temperature of the set attention points;

and S6, reconstructing a temperature field according to the determined mapping relation or the trained model.

The steps and the principle of the benchmark method for researching the temperature field reconstruction task in the spacecraft cabin based on machine learning provided by the embodiment of the invention are specifically explained as follows:

s1, constructing a task optimization problem for reconstructing a temperature field in the spacecraft cabin according to a layout structure in the spacecraft cabin.

Generally, a layout area in an aircraft cabin can be regarded as a square layout area of a two-dimensional plane, since the environment of the aircraft is a vacuum environment, convective heat exchange does not exist, radiation heat exchange can be ignored, and heat generated by components in the layout area in the aircraft cabin transfers heat in a heat conduction mode, the layout area in the aircraft cabin can be regarded as a heat conduction application scene of the two-dimensional plane.

Further, setting: the layout area in the spacecraft cabin is provided with Lambda heat source components, the heat source components represent components capable of generating energy dissipation during work, and the power distribution of the ith heat source component is phi _i (x, y) M temperature monitoring points provided with temperature sensors are arranged in the layout area, and the position of the mth temperature monitoring point is

Since the heat generated by the components in the layout area within the spacecraft cabin is transferred in a heat conduction mode, the temperature field of the layout area satisfies the following heat conduction differential equation:

where T denotes a temperature field of the layout area, k denotes a heat transfer coefficient, and (x, y) denotes a position coordinate of a certain point in the temperature field.

For different boundary heat dissipation modes of the layout area, the temperature field of the layout area may satisfy the following boundary conditions:

T＝T ₀ ,

wherein, T ₀ Denotes the temperature value at the isothermal boundary, n denotes the normal perpendicular to, h denotes the thermal convection coefficient, T = T ₀ Indicates Dirichlet boundary conditions, i.e. boundary conditions of the first type,

indicating that the Neumann boundary conditions, i.e. the second type of boundary conditions,

representing the Robin boundary conditions, i.e. the third class of boundary conditions.

Further, in an embodiment of the present invention, the spacecraft cabin interior temperature field reconstruction task indicates that the temperature of the set attention point or the temperature field of the entire layout area is obtained by using the temperature monitoring value of the temperature monitoring point in the layout area, and therefore, based on the setting, the spacecraft cabin interior temperature field reconstruction task optimization problem can be modeled as follows:

wherein,

in the temperature field representing the reconstruction

Temperature of the location, O _m Indicating the temperature monitoring value of the mth temperature monitoring point.

And S2, determining boundary conditions of a layout area in the spacecraft cabin and the power distribution condition and structure of the heat source assembly.

Aiming at the characteristics of systems in different spacecraft cabins, layout areas in the spacecraft cabins have different boundary conditions, and the common boundary conditions of the layout areas in the spacecraft cabins comprise Dirichlet boundary conditions such as small hole heat dissipation boundary conditions and sine distribution boundary conditions. The boundary condition of the small hole heat dissipation is a typical Dirichlet boundary condition which is an irreplaceable boundary condition in industrial electronics, and the boundary temperature of the small hole heat dissipation is a constant value T ₀ A boundary condition of (1). The boundary condition of a sinusoidal distribution is another typical DiricThe hlet boundary condition, which simulates a boundary condition in which the boundary temperature varies sinusoidally, can be expressed as:

wherein, T _m Denotes the sine wave temperature conversion amplitude, and L denotes the boundary length.

Further, the power distribution of heat source components within an aircraft cabin typically has both a uniform power distribution and a non-uniform power distribution. Uniform power distribution is the most common model of power assumptions, usually assuming the same power consumption at different locations on a component. The non-uniform power distribution is another power distribution form commonly used in engineering, and a Gaussian distribution mode is usually adopted, wherein the power consumption at the center of the component is the maximum, and the power consumption is lower as the distance from the center is farther, the power consumption distribution can be expressed as:

wherein Q is _i Denotes the maximum temperature at the center of the heat source assembly, λ denotes the offset coefficient, r _n Represents the radius of the Gaussian heat source component (x) ₀ ,y ₀ ) Represents the coordinate of the center point of the heat source component, omega _i Representing the region of the ith gaussian heat source element.

Further, the heat source components within the spacecraft capsule may have a variety of different shapes, such as rectangular, circular, capsule-shaped, oval, etc., and different heat source components may have different arrangement angles.

In one embodiment of the invention, the boundary condition of the layout area in the spacecraft cabin and the power distribution condition and structure of the heat source assembly are determined according to the actual condition in the spacecraft cabin, so that the training data and the test data can be acquired subsequently.

And S3, selecting temperature monitoring points for arranging the temperature sensors in the spacecraft cabin by utilizing a cluster analysis algorithm.

In one embodiment of the invention, the method for selecting the temperature monitoring points for arranging the temperature sensors in the spacecraft cabin by utilizing the cluster analysis algorithm comprises the following steps:

because the position of the heat source assembly in the layout area in the spacecraft cabin is fixed, but the power can change along with the change of the actual working condition, the layout under a specific working condition can be obtained by randomly sampling one power from the power distribution obeyed by each heat source assembly as the actual power of the corresponding heat source assembly.

Step S32, dividing the layout area into N ₁ ×N ₂ The temperature control method comprises the following steps that grids can only share one constant temperature value, and a finite element method is utilized to calculate and obtain a plurality of temperature fields corresponding to the layouts of a plurality of different working conditions based on the layout areas after the grids are divided;

in an embodiment of the present invention, a finite element solver FEniCS may be used to obtain a plurality of temperature fields corresponding to a plurality of layouts under different working conditions.

Step S33, the number of the temperature monitoring points is specified, and N with multiple dimensions is subjected to K-Means clustering algorithm ₁ ×N ₂ Clustering the temperature points in the spacecraft cabin to obtain a plurality of categories with the same number as the designated temperature monitoring points, calculating generalized distances based on Euclidean distance measurement, selecting the temperature point closest to the clustering center of each category as a representative of the corresponding category, and taking the selected temperature point and the corresponding position as the temperature monitoring point and the corresponding position.

In one embodiment of the present invention, N in the temperature field ₁ ×N ₂ Each temperature point is used as a characteristic point of cluster analysis, and the temperature values of the temperature points under different working conditions can be used as characteristic dimensions of corresponding characteristic points, for example, the layout of W different working conditions is obtained, and then the temperature points have W dimensions.

By adopting the mode to determine the temperature monitoring points used for arranging the temperature sensors in the spacecraft cabin, the redundancy among the temperature monitoring points can be smaller, the high-precision temperature field in the spacecraft cabin can be reconstructed by fewer temperature monitoring points, and the demand of the spacecraft platform on the temperature monitoring sensing devices, the measurement channels, the cables and other resources is reduced.

And S4, acquiring at least one of first training data, second training data, third training data and fourth training data, wherein the first training data comprises the position and the temperature of the temperature monitoring points in the layout area, the second training data comprises the temperature monitoring points in the layout area and the temperature of the set attention points, the third training data comprises the position and the temperature of the temperature monitoring points in the layout area and the temperature field corresponding to the layout area, and the fourth training data comprises the position and the temperature of the temperature monitoring points in the layout area and the position and the temperature of the set attention points in the layout area.

In an embodiment of the invention, on the basis of determining the boundary condition of the layout area in the spacecraft cabin, the position of the heat source assembly in the layout area, and the power distribution condition and structure of the heat source assembly, first training data, second training data, third training data and fourth training data are obtained.

In an embodiment of the present invention, the first training data may be obtained in the following manner:

determining the position of a temperature monitoring point in a layout area;

for each heat source assembly in the layout area, randomly sampling a power value from the power distribution of each heat source assembly as the current power of the corresponding heat source assembly to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point under the layout of the current specific working condition, obtaining first training data comprising the position and the temperature of the temperature monitoring point in the layout area, and repeating the process for multiple times until first training data with preset quantity is obtained.

The temperature of the temperature monitoring point can be determined in the following way:

carrying out simulation analysis on the layout under the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, and extracting the temperature of a temperature monitoring point from the obtained temperature field; or the power of the components in the actual spacecraft cabin layout area is adjusted according to the sampled power, and after the power adjustment is completed, the temperature of the temperature monitoring point is directly acquired by using the temperature sensor arranged on the temperature monitoring point.

Further, in an embodiment of the present invention, the second training data may be obtained by:

and randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component aiming at each heat source component in the layout area to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point and each set attention point under the layout of the current specific working condition to obtain second training data comprising the temperature monitoring points and the temperature of the set attention points in the layout area, and repeating the process for multiple times until second training data with preset quantity are obtained.

The temperature of the temperature monitoring point and the set attention point can be determined in the following way:

and carrying out simulation analysis on the layout under the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, and extracting temperature monitoring points and setting the temperature of the attention point from the obtained temperature field.

Further, in an embodiment of the present invention, the third training data may be obtained by:

and randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition aiming at each heat source component in the layout area, carrying out simulation analysis on the layout of the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, extracting the temperature of a temperature monitoring point from the obtained temperature field to obtain third training data comprising the position and the temperature of the temperature monitoring point in the layout area and the temperature field corresponding to the layout area, and repeating the process for multiple times until third training data of a preset number are obtained.

Further, in an embodiment of the present invention, the fourth training data may be obtained in the following manner:

The temperature of the temperature monitoring point and the set attention point can be determined in the following mode:

In an embodiment of the present invention, the set focus may be selected and determined according to actual requirements. The temperature monitoring point can be selected and determined by the method given in the step S2, or can be determined in the following manner:

the center of each heat source component is taken as a temperature monitoring point, three position points are selected as temperature monitoring points at each boundary of the layout area, and a plurality of position points are randomly selected as temperature monitoring points in other areas except the area where the heat source component is located in the layout area, for example, 10 position points can be randomly selected as temperature monitoring points in other areas.

In an embodiment of the present invention, the position coordinates of the temperature monitoring point and the set attention point may be determined by presetting a two-dimensional plane coordinate system in the layout area.

Further, in the case of comprehensively considering the data acquisition cost and the model performance, 10000 pieces of generated first training data, second training data, third training data and fourth training data may be generated, wherein, in the generated training data, 80% of the data may be used for training, and the remaining 20% of the data may be used for verification.

Further, in order to comprehensively evaluate the performance of different temperature field reconstruction modes, in an embodiment of the present invention, a random sampling strategy and a special sample strategy are used to generate a plurality of different types of test data for subsequent model tests.

In one embodiment of the invention, one type of test data is generated through a random sampling strategy, and a plurality of different types of test data are generated through a special sample strategy. Each test data includes a layout under a specific working condition and a temperature field corresponding to the layout.

Specifically, in an embodiment of the present invention, generating test data by using a random sampling strategy includes the following steps:

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition aiming at each heat source component in the layout area, carrying out simulation analysis on the layout of the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, obtaining test data comprising the layout under the specific working condition and the temperature field corresponding to the layout, and repeating the process for multiple times until obtaining the preset number of test data.

A special sample policy means that diversified test data is generated according to some special rules. There are two main types of special sample strategies, including a power consistency sample strategy and a zero power sample strategy. The power consistency sample strategy indicates that all heat sources in the layout area have the same power intensity. The zero power sample strategy indicates that a portion of the heat sources in the layout area have zero power intensity. One type of test data is generated through a power consistency sample strategy, and a plurality of different types of test data are generated through a zero power sample strategy.

Specifically, in an embodiment of the present invention, generating test data through a power consistency sample policy includes the following steps:

according to the power distribution of each heat source component in the layout area, the same power range of all the heat source components is determined, one power is randomly sampled from the determined power range to serve as the current power of all the heat source components, the layout of a specific working condition is obtained, the finite element method is utilized to carry out simulation analysis on the layout of the specific working condition to obtain a temperature field corresponding to the layout, test data including the layout under the specific working condition and the temperature field corresponding to the layout are obtained, and the process is repeated for multiple times until a preset number of test data are obtained.

Specifically, in an embodiment of the present invention, generating test data by a zero-power sample strategy includes the following steps:

selecting part of heat source components in the layout area, and setting the power of the selected heat source components to be zero;

and randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition aiming at the rest heat source components, carrying out simulation analysis on the layout of the specific working condition by using a finite element method to obtain a temperature field corresponding to the layout, obtaining test data comprising the layout and the temperature field corresponding to the layout under the specific working condition, and repeating the process for multiple times until obtaining the preset number of test data.

Wherein a plurality of different types of test data can be obtained by adjusting the number of heat source assemblies having a zero power intensity. For example, 1/4 of the heat source components, or 1/2 of the heat source components, or 3/4 of the heat source components, or all but one of the heat source components in the layout area may be set to zero power.

Further, under the condition of comprehensively considering data acquisition cost and model performance test, 10000 test data can be generated when test data are generated through a random sampling strategy, and 2000 test data can be generated for each type when various different types of test data are generated through a special sample strategy.

Furthermore, because the boundary conditions of the spacecraft indoor layout are diversified, for the convenience of subsequent reconstruction of the temperature field, for different spacecraft indoor layout characteristics, a corresponding group of first training data, a group of second training data, a group of third training data, a group of fourth training data and corresponding test data can be generated by adopting the above mode.

In particular, corresponding training data and test data may be generated for three common spacecraft in-cabin systems. The first spacecraft cabin system is a small-hole radiating spacecraft cabin system, radiating holes are formed in the boundary of the layout area of the spacecraft cabin system, other boundaries except the radiating holes are thermal insulation boundaries, and the width of the radiating holes and the temperature value of the radiating hole area can be set to be 0.01m and 298K, for example. A second type of spacecraft inboard system is a spacecraft inboard system with different Dirichlet boundary conditions, where the temperature at one boundary of the layout area of the spacecraft inboard system is a sine wave profile and the temperature at the other boundary is a fixed value, such as 298K. The third spacecraft cabin system is a spacecraft cabin system with a sine wave distribution boundary, the temperature of one boundary of a layout area of the spacecraft cabin system is in sine wave distribution, and other boundaries are heat insulation boundaries.

S5, determining a mapping relation from the position of any point in the layout area to the temperature of the point by an interpolation method by using the first training data; or training a constructed traditional machine learning model or a neural network model by utilizing first training data to fit a mapping relation from the position of any point in the layout area to the temperature of the point; or training the constructed multilayer perceptron by utilizing second training data to fit the mapping relation from the temperature of the temperature monitoring points in the layout area to the temperature of the set attention point; or generating a corresponding temperature monitoring matrix according to the position and temperature information of the temperature monitoring points in the layout area in the third training data, and fitting a mapping relation from the temperature monitoring matrix to the temperature field by using a deep neural network model constructed by the temperature monitoring matrix and the temperature field training in the third training data; or a graph model is constructed according to the position relation between the temperature monitoring points and the set attention points, and a graph convolution neural network model constructed by training the temperature of the temperature monitoring points in the graph model and the fourth training data and the temperature of the set attention points is utilized to fit the mapping relation between the graph model and the temperature of the temperature monitoring points to the temperature of the set attention points.

In an embodiment of the present invention, the interpolation method may be a k-nearest neighbor non-linear interpolation method (kinterpation) or a global gaussian interpolation method (ginterpation).

k-nearest neighbor nonlinear interpolation method (Kinterplation) measures the distance between the set attention point and the temperature monitoring point by Euclidean distance measurement, and reconstructs the temperature value of the set attention point through interpolation by considering the relationship between the set attention point and k nearest temperature monitoring points with the distance, wherein the temperature value is at the set attention point (x) ₀ ,y ₀ ) The reconstructed temperature value at (a) may be expressed as:

wherein, T (x) ₀ ,y ₀ ) Indicates the set point of interest (x) ₀ ,y ₀ ) I.e. temperature values reconstructed on the basis of the temperature at the temperature monitoring point, S _k (x ₀ ,y ₀ ) Indicates the distance setting attention point (x) ₀ ,y ₀ ) The most recent k temperature monitoring points are,

indicating the jth temperature monitoring point at which,

representing the temperature value of the ith temperature monitoring point.

The global Gaussian interpolation method (GInterplation) considers setting the attention point and all temperature monitorsThe relationship between the measuring points, the temperature value of the concerned point is reconstructed by interpolation, and the concerned point (x) is set ₀ ,y ₀ ) The reconstructed temperature value at (a) may be expressed as:

wherein, T (x) ₀ ,y ₀ ) Indicates the set point of interest (x) ₀ ,y ₀ ) I.e., temperature values reconstructed based on the temperature at the temperature monitoring points,

indicating the jth temperature monitoring point at which,

and the temperature value of the ith temperature monitoring point is represented.

Further, in an embodiment of the invention, the traditional machine learning model may be one of a polynomial regression model, a random forest regression model, a gaussian process regression model and a support vector machine regression model.

Specifically, the first training data trains the polynomial regression model to fit a mapping relationship from the position of any point in the layout region to the temperature of the point, i.e., a nonlinear mapping relationship from the position of any point in the layout region to the temperature of the point is fitted by using a polynomial function. Polynomial Regression (multinomial Regression) is the Regression in which the Regression function is a Regression variable Polynomial. In statistics, polynomial regression is a form of regression analysis in which the relationship between an independent variable x and a dependent variable y is modeled as an nth order polynomial on the independent variable. Polynomial regression fits the nonlinear relationship between the values of the independent variables and the corresponding conditional mean values of the dependent variables, which can be denoted as E (y | x), and has been used to describe the nonlinear phenomenon. While polynomial regression is a non-linear model to fit data, it is linear as a statistical estimation problem. In a sense, the regression function E (y | x) is linear in the unknown parameters estimated from the data. Thus, polynomial regression is considered a special case of multiple linear regression. Since any function can be approximated by a polynomial, polynomial regression can be used to fit the temperature field to reconstruct a high-dimensional mapping problem. The method has the advantage that the actual measurement point is approximated by increasing the high-order terms of the independent variable until the actual measurement point is satisfied. Polynomial regression can deal with non-linear problems, which plays an important role in regression analysis. Since any function can be approximated by a polynomial in a piecewise manner, in a general practical problem, polynomial regression can be used for analysis regardless of the relationship between the dependent variable and other independent variables.

Random Forest (Random Forest) is a supervised learning method based on a decision tree composition formula. The decision tree is a very simple algorithm and also conforms to the intuitive thinking of human beings. This is a supervised learning algorithm based on if-then-else rules. And the random forest is composed of a plurality of decision trees, and different decision trees have no relation. When a task of reconstructing a system temperature field in an aircraft cabin is faced, a new input sample enters, each decision tree in the forest is subjected to regression respectively, each decision tree can obtain a regression result, and the random forest can take the average value of the results as a final result. In the random forest, a plurality of prediction models are generated simultaneously, and the results of the models are summarized to improve the accuracy of the prediction models. The method has the advantages that the method can output data with high dimension (a plurality of characteristics), does not need dimension reduction and characteristic selection, and can judge the importance degree of the characteristics and the mutual influence among different characteristics. Random forests are not easy to over-fit, the training speed is high, a parallel method is easy to make, and the realization is simple. For unbalanced data sets, it may balance the error. Accuracy can still be maintained if a significant portion of the features are missing.

The Gaussian process regression model obtains the whole temperature field function distribution of the layout area by using the known positions and temperature information of the temperature monitoring points in the layout area, so that the prediction of the temperature of the unknown set attention point is realized.

Data set D formed by information of set temperature monitoring points _obs :(X _obs ,T _obs ) Wherein X is _obs ＝(x _obs ,y _obs ) Position coordinates, T, representing temperature monitoring points _obs Indicating the temperature at the temperature monitoring point. Let g (X) _obs,i )＝T _obs,i ，X _obs,i Position coordinates, T, representing the ith temperature monitoring point _obs,i Represents the temperature of the ith temperature monitoring point, resulting in the vector g = [ g (X) _obs,1 ),g(X _obs,2 ),...,g(X _obs,M )]Defining a set of set points of interest as X ^* Setting the temperature value of the point of interest to g ^* ＝T ^* According to a Bayesian formula, the following can be obtained:

where p (·) represents a probability.

When a Gaussian process regression model is trained by utilizing first training data to fit a mapping relation from the position of any point in a layout area to the temperature of the point, the joint probability density distribution g-N (mu, K) among temperature monitoring point samples in a data set formed by the information of the temperature monitoring points is calculated, and then the g is predicted according to the requirement ^* To calculate g from the prior probability distribution of ^* Wherein μ is g = [ g (X) _obs,1 ),g(X _obs,2 ),...,g(X _obs,M )]K is a covariance matrix. The method specifically comprises the following steps:

in that

Under the condition that p (T) is _obs |X _obs ) At maximum, the objective function is set to logp (T) _obs |X _obs )＝logN(μ,K _T ) According to a gradient descent algorithm, an optimal value is searched by using the following formula;

at the known g (X) _obs )～N(μ,K _T ) G (X × N) to N (μ × k (X, X)), the joint probability distribution is calculated by the following formula;

estimating a temperature value vector g (x) of a set attention point by using a Bayesian formula according to the obtained joint probability distribution;

g (X) ~ mu ', K ') can be obtained, wherein, mu ' = K ^T K ^-1 g(X _obs )，K′＝K* ^T K ^-1 K*+K**；

Wherein g (X) _obs ) Represents X _obs A joint probability distribution, m (-) represents a mean function, m (X) _obs ) Represents X _obs Is corresponding to the average value of the average values is calculated,

represents X _obs And

the radial basis function kernel of (a) is,

alpha and l are radial basis function kernel hyperparameters,

represents X _obs The transpose of (a) is performed,

represents T _obs Transpose of (K) _T Represent

g (X) represents the joint probability distribution at X, μ represents the vector consisting of the mean of g, K = K (X, X), K (X, X)And K, K (X, X) represents the radial basis function kernels of X and X.

The support vector machine regression model uses the location and temperature information of the temperature monitoring points in the known layout area to enable prediction of the temperature of the unknown set point of interest.

Training a regression model of a support vector machine by utilizing first training data to fit a mapping relation from the position of any point in a layout region to the temperature of the point, and specifically comprising the following steps of:

data set D formed by information of set temperature monitoring points _obs :(X _obs ,T _obs ) Wherein X is _obs ＝(x _obs ,y _obs ) Position coordinates, T, representing temperature monitoring points _obs Indicating the temperature at the temperature monitoring point.

Setting the regression model as g, and g (X) = w ^T X + b, w and b represent the model parameters to be determined;

setting that a deviation parameter epsilon exists, and calculating the loss if and only if the absolute error of g (X) and T is larger than the deviation parameter;

and converting the temperature field reconstruction target into a target with a support vector machine method:

reintroducing variable eta _i And

lagrange multiplier mu _i And a hyper-parameter alpha, setting the optimization objective as:

obtaining the corresponding w, b value by using the partial derivative of L to w, b, eta as 0, and obtaining a solution:

where C denotes a regularization constant, l _∈ Representing the insensitive loss function, M representing the total number of temperature monitoring points, eta _i And

denotes the upper and lower bound relaxation variables, μ _i 、

α _i 、

Representing the Lagrange coefficient, g (X) _obs,i ) Shows that the regression model g (-) is at X _obs,i At a predicted temperature value, X _obs,i Position coordinates, T, representing the ith temperature monitoring point _obs,i Representing the actual temperature value of the ith temperature monitoring point;

wherein the loss insensitive function l _∈ Expressed as:

wherein z represents a variable.

Further, in an embodiment of the present invention, the neural network model may be one of a multi-layer perceptron, a constrained boltzmann machine, and a deep belief network.

Referring to fig. 2, a Multi-layer Perceptron (Multi-layer Perceptron) is a typical neural network, which includes an input layer, a plurality of hidden layers and an output layer, each layer is composed of a plurality of neurons, the neurons between layers are connected in a full connection manner, and each neuron includes a weight parameter and a bias parameter. In addition, in order to introduce nonlinear transformation, the nonlinear transformation of element operation is carried out on hidden variables by using an activation function, and then the hidden variables are used as the input of the next full-connection layer. The activation function may be a Sigmoid function, a Relu function, a Tanh function, or the like.

In an embodiment of the present invention, when the neural network model is a multi-layer perceptron, training the neural network model using the first training data includes:

and taking the position coordinates of the temperature monitoring points in the first training data as the input of the multilayer perceptron, and taking the temperature of the temperature monitoring points in the first training data as the output of the multilayer perceptron to train the multilayer perceptron.

Specifically, in an embodiment of the present invention, the training of the multi-layer sensor by using the position coordinates of the temperature monitoring points in the first training data as the input of the multi-layer sensor and using the temperature of the temperature monitoring points in the first training data as the output of the multi-layer sensor further includes the following steps:

step S501, sequentially inputting the position coordinates of the temperature monitoring points in the first training data into the multilayer perceptron to obtain the temperature predicted values of the temperature monitoring points output by the multilayer perceptron;

in an embodiment of the invention, the position coordinates of the temperature monitoring points in the first training data are input from an input layer of the multi-layer perceptron, are sequentially processed by parameters of each layer in the multi-layer perceptron, and are output from an output layer of the multi-layer perceptron, and information output by the output layer is the temperature predicted value of the corresponding temperature monitoring points. The initial multi-layer perceptron can be an untrained neural network or an untrained neural network, each layer of the initial multi-layer perceptron is provided with initialized parameters, and the parameters can be continuously updated and adjusted in the training process of the neural network.

Step S502, comparing the temperature predicted value of the temperature monitoring point with the temperature of the temperature monitoring point in the first training data, and calculating the prediction accuracy of the multilayer perceptron;

in an embodiment of the invention, the average value of all ratios can be used as the prediction accuracy by calculating the ratio of the difference between the temperature predicted value of the temperature monitoring point corresponding to each first training data and the temperature of the temperature monitoring point to the temperature of the temperature monitoring point.

Step S503, judging whether the prediction accuracy obtained at least twice continuously is larger than a preset accuracy threshold, if so, taking the current multilayer perceptron as the trained multilayer perceptron, if not, calculating a loss function, updating parameters of the multilayer perceptron by using the loss function, and returning to the step S501.

In an embodiment of the invention, an average absolute error (MAE) between a temperature predicted value and a temperature actual value of a temperature monitoring point can be used as a loss function of the multilayer perceptron, and a gradient descent method can be used for optimizing and updating parameters of the multilayer perceptron.

Specifically, for a spacecraft in-cabin system, the position coordinates of temperature monitoring points in a spacecraft in-cabin layout area are taken as batch samples

Where M is the batch sample size, i.e. the number of temperature monitoring points, d is the number of inputs, and when the inputs are (x, y) coordinates, d =2.

Setting the hidden layer output to H, the output of the hidden layer can be expressed as:

H＝φ(XW _h +b _h )

wherein, W _h And b _h The weights and bias parameters of the hidden layer are indicated, respectively, phi denotes the activation function.

Taking the output of the hidden layer as the input of the output layer, the output of the corresponding output layer can be expressed as:

T＝HW _T +b _T

wherein, W _T And b _T And respectively representing the weight and the bias parameter of the output layer, and T represents the output of the output layer, namely the temperature corresponding to the temperature monitoring point.

Wherein, the weight and the bias parameter of the hidden layer and the weight and the bias parameter of the output layer are determined by the training process.

The multi-layer perceptron is trained by utilizing the first training data, a point modeling temperature field reconstruction model from any point coordinate in a layout area in the spacecraft cabin to a corresponding temperature value can be constructed, and when the multi-layer perceptron is used, the coordinate of a set focus point in the layout area is input into the multi-layer perceptron, so that the corresponding predicted temperature can be obtained.

Referring to fig. 3, in an embodiment of the present invention, the restricted boltzmann machine is a neural network model formed by two layers of neurons, and includes: a visible layer (visible layer) and a hidden layer (hidden layer). No connection exists between the nerve cells of each layer, and the connection is complete between the layers, so that a bipartite graph is formed. Aiming at a task of reconstructing a temperature field in an aircraft cabin, the restricted Boltzmann machine is mainly used for extracting model characteristics.

The limited boltzmann machine is an unsupervised model, and aims to reconstruct probability distribution of input data, parameters of the model can be obtained by maximizing a log-likelihood function on training data, and the log-likelihood function on the training data can be specifically expressed as:

wherein N represents the total number of training data, v represents a neuron value of a visible layer, h represents a neuron value of a hidden layer, θ represents a parameter of a limited boltzmann machine, the parameter of the limited boltzmann machine includes a bias vector of the visible layer, a bias vector of the hidden layer, and a connection weight between the hidden layer and the visible layer, P (v, h | θ) represents a probability distribution of v and h, and P (v, h | θ) is specifically represented as:

wherein E (v, h | θ) represents an energy function, Z (θ) represents a normalization factor,

in one embodiment of the invention, when the first training data is used for training the limited boltzmann machine, the position coordinates of the temperature monitoring points in the first training data are used as the input of the limited boltzmann machine, the temperature of the temperature monitoring points in the first training data is used as the output of the limited boltzmann machine, and the limited boltzmann machine is trained.

Specifically, the position coordinates of the temperature monitoring points in the first training data are used as the input of the visible layer of the restricted Boltzmann machine, hidden layer variables are obtained by maximizing a log-likelihood function on the first training data, and then the hidden variables are mapped to the temperatures of the corresponding position points through linear regression, so that a point modeling and reconstruction model from the position coordinates to the temperatures is constructed.

Further, in an embodiment of the present invention, a Deep Belief network (Deep Belief Networks) includes a plurality of bounded boltzmann machines stacked in series, which is a typical Deep neural network. Wherein, the hidden layer output of the previous layer of limited Boltzmann machine is the visible layer input of the next layer of limited Boltzmann machine, and the last layer is the output layer. And aiming at the task of reconstructing the temperature field in the spacecraft cabin, the deep confidence network is mainly used for extracting model characteristics.

In one embodiment of the invention, when the deep belief network is trained by using the first training data, the position coordinates of the temperature monitoring points in the first training data are used as the input of the deep belief network, and the temperature of the temperature monitoring points in the first training data is used as the output of the deep belief network, so that the deep belief network is trained.

The training process of the deep belief network comprises an unsupervised pre-training and a supervised reverse fine tuning process based on a limited Boltzmann machine. The unsupervised pre-training mode is the same as the training mode of the limited Boltzmann machine.

Specifically, when the deep confidence network is trained, a first limited Boltzmann machine is fully trained, then the weight and the bias parameters of the first limited Boltzmann machine are fixed, the output of a hidden layer of the first limited Boltzmann machine is used as the visible layer input of the next limited Boltzmann machine and is fully trained, the first limited Boltzmann machine is placed above the first limited Boltzmann machine, and finally the steps are iterated until the unsupervised pre-training of all the limited Boltzmann machines is completed; after the unsupervised pre-training is completed, the parameters of the model are subjected to supervised training adjustment through a back propagation algorithm by utilizing first training data.

When the first training data is used for carrying out supervised training adjustment on the parameters of the deep belief network through a back propagation algorithm, a loss function is defined as:

wherein N represents the total number of training data,

a temperature prediction value T representing the output of the deep confidence network corresponding to the ith first training data _i Representing the temperature values in the ith first training data.

In an embodiment of the invention, the position coordinates of the temperature monitoring points in the first training data are used as the input of the deep confidence network to obtain the output of the last hidden layer in the deep confidence network, and the hidden variables are further mapped to the temperatures of the corresponding position points through linear regression, so that a point modeling reconstruction model from the position coordinates to the temperatures is constructed.

Further, in an embodiment of the present invention, the training of the built multi-layer perceptron by using the second training data includes the following steps:

In an embodiment of the present invention, the multi-layer sensor constructed in this step may have the same structure as the above-mentioned multi-layer sensor. Specifically, the multi-layer perceptron may include an input layer, a plurality of hidden layers, and an output layer, each layer being composed of a plurality of neurons, the neurons between layers being connected in a fully connected manner, each neuron including a weight parameter and a bias parameter.

Further, in an embodiment of the present invention, the training of the multi-layered sensor by using the temperature vector of the temperature monitoring point in the second training data as the input of the multi-layered sensor and using the temperature vector of the set attention point in the second training data as the output of the multi-layered sensor further includes the following steps:

step S511, the temperature vectors of the temperature monitoring points in the plurality of second training data are sequentially input into the multilayer perceptron to obtain the predicted temperature vector of the set concern point output by the multilayer perceptron;

in an embodiment of the present invention, the temperature vector of the temperature monitoring point in the second training data is input from the input layer of the multi-layer sensor, sequentially processed by the parameters of each layer in the multi-layer sensor, and output from the output layer of the multi-layer sensor, where the information output by the output layer is the predicted temperature vector of the corresponding set point of interest. The initial multi-layer perceptron can be an untrained neural network or an untrained neural network, each layer of the initial multi-layer perceptron is provided with initialized parameters, and the parameters can be continuously updated and adjusted in the training process of the neural network. The temperatures of a plurality of temperature monitoring points in one second training datum form a temperature vector, and the temperatures of a plurality of set attention points form a temperature vector.

Step S512, comparing the predicted temperature vector of the set attention point with the temperature vector of the set attention point in the second training data, and calculating the prediction accuracy of the multilayer perceptron;

in an embodiment of the present invention, a ratio between a difference between the predicted temperature vector of the set attention point corresponding to each second training data and the temperature vector of the set attention point may be calculated, and an average value of all ratios is used as the prediction accuracy.

Step S513, judging whether the prediction accuracy obtained at least twice continuously is larger than a preset accuracy threshold, if so, using the current multilayer perceptron as the trained multilayer perceptron, if not, calculating a loss function, updating parameters of the multilayer perceptron by using the loss function, and returning to step S511.

In an embodiment of the present invention, the loss function may be set as:

wherein B represents the second training data number, m represents the number of set points of interest,

the predicted temperature value of the jth set attention point output by the multilayer perceptron corresponding to the (b) th second training data is represented,

and a temperature value representing the jth set point of interest in the jth second training data.

Further, parameters of the multilayer perceptron can be optimized and updated by adopting a gradient descent method according to the set loss function.

In an embodiment of the invention, the multilayer perceptron is trained by utilizing the second training data, a vector mapping proxy model from a vector to a vector can be constructed, and the temperature reconstruction of the set focus in the spacecraft cabin layout area is realized.

Further, in order to utilize the spatial relationship between the temperature monitoring point and the set attention point, an embodiment of the present invention further provides a temperature field reconstruction method based on image-to-image regression, and the temperature field reconstruction task is modeled as an image-to-image regression task. Since the task of temperature field reconstruction is modeled as an image-to-image regression task, it can be divided into an encoder process and a decoder process. Using different encoder and decoder methods, the depth regression framework can be divided into several categories, such as Full Convolution Networks (FCNs), segNet, unet, feature Pyramid Networks (FPNs), and the like.

Therefore, in an embodiment of the present invention, when the deep neural network model constructed by using the temperature monitoring matrix and the temperature field in the third training data for training is used to fit the mapping relationship from the temperature monitoring matrix to the temperature field, the deep neural network model may be one of a full convolution neural network model, a feature pyramid network model, a SegNet network model, and a UNet network model.

Fully convolutional neural networks (FCNs) perform only convolution operations (down-sampling or up-sampling), which is a typical image-to-image regression framework. Typically, a full convolutional neural network is implemented by replacing the parameter-rich fully-connected layer in the standard CNN architecture with a 1 × 1-core convolutional layer. The key to generating dense prediction outputs in a full convolutional neural network is to use deconvolution, which is simply the convolutional layer where the inputs and outputs are swapped. For a fully convolutional neural network, a network architecture with different backbone networks can be constructed, such as AlexNet, VGG16, and ResNet. Wherein the kernel size of the first layer in AlexNet is set to 7 and the padding of the maximum pool layer is set to 1.

A Feature Pyramid Network (FPN) is a special type of full convolution neural network that utilizes multi-scale information through a feature pyramid structure. In order to construct high-level semantic feature maps of different scales, a feature pyramid network uses top-down paths and horizontal connections. The feature pyramid network consists of three parts, including: bottom-up path, top-down path, and cross-connect. The purpose of the top-down path is to convert the low resolution feature map to higher resolution by upsampling. Through transverse connection, feature graphs from bottom-up and top-down paths of the same space size are merged to combine high-level and low-level semantic information to realize multi-scale feature fusion. Therefore, the characteristic pyramid network can fully utilize multi-scale information and provide more accurate prediction.

In an embodiment of the invention, the constructed feature pyramid network model architecture can expand multi-scale feature mapping to the same size through two times of continuous upsampling and 3 × 3 conversion, and then the features are fused through addition operation, and finally prediction is generated by 1 convolution and upsampling. In addition, resNet-18, resNet-50 and ResNet-101 can be used as the backbone network for the bottom-up path, respectively, to achieve spacecraft in-cabin temperature field reconstruction.

The SegNet network is a typical convolutional encoder-decoder architecture, and an encoder mainly performs feature extraction through a down-sampling operation to down-sample an input image to a low-resolution feature map. The decoder stage then upsamples the low resolution picture to a high resolution temperature field prediction, as opposed to the encoder stage. In the SegNet network, a decoder is composed of a group of upsampling layers and convolutional layers, and the structure of the decoder corresponds to that of the encoder. Unlike the feature pyramid network, the SegNet network does not utilize a hopping structure that connects the shallow information, and the SegNet network employs an inverse pooling operation to upsample the feature map.

In an embodiment of the present invention, alexNet, VGG, and ResNet may be used as encoders of the SegNet network, and the convolution operation and the inverse pooling operation in the decoder correspond to the encoders, respectively. When AlexNet is used as a backbone network, the first maximum pool layer in the encoder is deleted, and two-layer deconvolution with the kernel size of 2 and the step length of 2 is adopted, so that 4 times of upsampling is realized at the final stage of the decoder.

UNet network is a coder-decoder depth regression framework, and is composed of a coder for capturing context and performing downsampling operation and a decoder for performing upsampling and temperature field reconstruction. The UNet network adopts a jump connection architecture, combines the down-sampling feature map with the corresponding up-sampling feature map, reduces information loss in the down-sampling process and improves the reconstruction precision of the temperature field.

Further, in an embodiment of the present invention, according to the position and the temperature information of the temperature monitoring point in the layout area in the third training data, the corresponding temperature monitoring matrix is generated in the following manner:

according to the layout area after grid division, in the grids with the temperature monitoring points, the temperature of the temperature monitoring points is used as a matrix element, in the grids without the temperature monitoring points, zero is used as a matrix element, and the generated dimension is M ₁ ×M ₂ The temperature monitoring matrix of (a).

Further, in an embodiment of the present invention, the deep neural network model constructed by training using the temperature monitoring matrix and the temperature field in the third training data includes the following steps:

Further, in an embodiment of the present invention, the training of the deep neural network model by using the temperature monitoring matrix corresponding to the third training data as the input of the deep neural network model and using the temperature field in the third training data as the output of the deep neural network model further includes the following steps:

step S521, sequentially inputting the temperature monitoring matrixes corresponding to the plurality of third training data into a deep neural network model to obtain a predicted temperature field output by the deep neural network model;

in an embodiment of the present invention, the temperature monitoring matrix corresponding to the third training data is input from the input layer of the deep neural network model, sequentially processed by the parameters of each layer in the deep neural network model, and output from the output layer of the deep neural network model, where the information output by the output layer is the corresponding predicted temperature field. The initial deep neural network model can be an untrained deep neural network or an untrained deep neural network, each layer of the initial deep neural network model is provided with initialized parameters, and the parameters can be continuously updated and adjusted in the training process of the deep neural network.

Step S522, comparing the predicted temperature field with the temperature field in the third training data, and calculating the prediction accuracy of the deep neural network model;

in an embodiment of the present invention, the average value of all the ratios may be used as the prediction accuracy by calculating the difference between the predicted temperature field corresponding to each third training data and the temperature field in the third training data and the ratio between the temperature fields in the third training data.

Step S523, determining whether the prediction accuracy obtained at least two consecutive times is greater than a preset accuracy threshold, if yes, using the current deep neural network model as the trained deep neural network model, if not, calculating a loss function, updating parameters of the deep neural network model by using the loss function, and returning to step S521.

In an embodiment of the present invention, the loss function may be an average absolute error (MAE) loss function, and according to the set loss function, a gradient descent method is used to optimize and update parameters of the deep neural network model.

In an embodiment of the invention, the deep neural network model is trained by utilizing the third training data, so that an image mapping proxy model from an image to an image can be constructed, and the reconstruction of the temperature field of the layout area in the spacecraft cabin is realized.

Further, in an embodiment of the present invention, in order to obtain a convolutional neural network model capable of being used for reconstructing a temperature field in an spacecraft cabin, three processes are required to be performed, including: graph model building, graph convolution neural network and graph convolution neural network training.

The graph model is an unstructured data structure, and can store and process irregular data. The key for constructing the graph model is the determination of nodes and edges in the graph model, and a graph G = (V, E) is defined, wherein V represents a node set of the graph, is a set of a series of entity nodes and can store the characteristics of the nodes; and E represents an edge set of the graph, is used for describing the connection relation between the nodes, and can represent different connection relations of the edges through the weights of the edges. In order to better describe the connection relationship between the nodes and the edges of the graph, the connection relationship is generally represented by an adjacency matrix a and a degree matrix D, the adjacency matrix of the graph is a matrix representing the adjacency relationship between the nodes, the degree matrix is a matrix formed by taking the degree of each node as an element, and the degree of the node represents the number of the edges associated with the node.

In one embodiment of the invention, one position point is taken as a node, and a graph model is constructed according to the position relation of all temperature monitoring points and set attention points.

Referring to fig. 4, in an embodiment of the present invention, constructing a graph model according to the position relationship between all temperature monitoring points and a set attention point includes the following steps:

and constructing a corresponding graph model according to all the determined nodes and edges.

The preset distance threshold value can be specifically set according to the influence ranges of the component structure type, the component layout mode, the heat dissipation mode and the component temperature in the spacecraft cabin.

The graph convolution neural network introduces deep learning into graph theory for processing the problem of irregular regions. The graph convolution neural network is a neural network introduced into graph processing by the neural network, and can extract neighborhood information of nodes and transmit the neighborhood information to a next layer through learnable parameters.

In an embodiment of the invention, a Graph Convolutional neural Network (GCN) includes a plurality of Graph Convolutional layers, and through Graph Convolutional operation, the Graph Convolutional neural Network can extract neighborhood information and perform aggregation to obtain node hidden layer representation. The graph convolution operation is mainly divided into two steps, the first step is aggregation, namely aggregation is performed on information of a neighborhood node, and the second step is transmission, namely transmission is performed on the information of the node to a rear layer, and the method can be specifically expressed as follows:

wherein H ^l+1 Features the l +1 th layer of the graph convolution neural network, σ (-) represents a nonlinear activation function,

representing an adjacency matrix added to a self-loop, a represents an adjacency matrix,

representing degree matrix added to self-circulation, D representing degree matrix, I _N Represents a unit matrix, H ^l Features, θ, representing the l-th layer of the graph-convolved neural network ^l Parameters representing the l layer of the graph convolution neural network; when l =0, H ⁰ Represents the input of the graph convolution neural network, when l = K, K is the number of layers of the graph convolution neural network, H ^K The output of the graph convolution neural network is represented.

Wherein, the adjacency matrix is used for representing the interconnection relation between the nodes, if an undirected edge exists between the node i and the node j, the corresponding adjacency matrix element A _ij =1, otherwise A _ij =0; the degree matrix is used to represent the number of edges present at each node, and is a diagonal matrix,

m denotes the number of columns of the adjacency matrix.

Further, in an embodiment of the present invention, in order to improve the prediction accuracy and avoid the problem that the atlas neural network is excessively smooth, a batch normalization (BatchNorm) strategy and a Residual connection (Residual connection) form are added to the atlas neural network.

Specifically, after adding the strategy of batch normalization and the form of residual concatenation, the graph convolution operation can be expressed as:

wherein BN (-) represents batch normalization.

Further, after the atlas neural network is constructed, in order to improve the prediction accuracy of the atlas neural network and reduce the prediction error, the atlas neural network needs to be trained and updated by using training data.

In an embodiment of the present invention, a graph convolution neural network model constructed by training using a graph model, temperatures of temperature monitoring points in fourth training data, and temperatures of set attention points includes the following steps:

and taking the temperature of the temperature monitoring point in the fourth training data as the input characteristic of the corresponding node in the graph model, taking the graph model which comprises the node input characteristic and corresponds to the fourth training data as the input of the graph convolution neural network model, and taking the temperature of the set focus point in the fourth training data as the output of the graph convolution neural network model, thereby training the graph convolution neural network model.

Further, in an embodiment of the present invention, the training of the graph convolutional neural network model by using the temperature of the temperature monitoring point in the fourth training data as the input feature of the corresponding node in the graph model, using the graph model including the node input feature corresponding to the fourth training data as the input of the graph convolutional neural network model, and using the temperature of the set point of interest in the fourth training data as the output of the graph convolutional neural network model specifically includes the following steps:

step S531, taking the temperature of the temperature monitoring points in the fourth training data as the input features of the corresponding nodes in the graph model, determining a plurality of graph models which comprise the node input features and correspond to the fourth training data, and sequentially inputting the plurality of graph models which comprise the node input features into the graph convolution neural network model to obtain a temperature predicted value of the set attention points output by the graph convolution neural network model;

in an embodiment of the present invention, a graph model including node input features is input from an input end of a graph convolution neural network model, sequentially processed by parameters of each layer in the graph convolution neural network model, and output from an output end of the graph convolution neural network model, where information output from the output end is a predicted temperature value of a set focus corresponding to the graph model including the node input features. The initial image volume neural network model can be an untrained neural network or an untrained neural network, each layer of the model is provided with initialized parameters, and the parameters can be continuously updated and adjusted in the training process of the neural network.

Step S532, comparing the temperature predicted value of the set attention point with the temperature of the set attention point in the fourth training data, and calculating the prediction accuracy of the graph convolution neural network model;

in an embodiment of the present invention, the ratio of the difference between the predicted temperature value of the set attention point corresponding to each piece of training data and the temperature of the set attention point to the temperature of the set attention point may be calculated, and the average value of all the ratios is used as the prediction accuracy.

Step S533, judging whether the prediction accuracy obtained at least twice continuously is greater than a preset accuracy threshold, if so, taking the current graph convolution neural network model as the graph convolution neural network model after training is completed, if not, calculating a loss function, updating parameters of the graph convolution neural network model by using the loss function, and returning to step S531.

In an embodiment of the present invention, the average absolute error may be used as a loss function of the convolutional neural network model, and at this time, the loss function may be specifically expressed as:

wherein N represents the number of training data used in training, y _i Representing the temperature of the set point of interest in the ith training data,

and the temperature predicted value of the set attention point output by the graph convolution neural network model corresponding to the ith training data is represented, and the quantity of the training data can be determined according to the actual training requirement.

In an embodiment of the invention, the parameters of the graph convolution neural network model are optimized and updated in a gradient descent mode. Specifically, the parameters are derived by using a loss function by using a chain type derivation rule, and then the parameters are updated by using a derivation result and a preset learning rate.

Specifically, the following formula may be adopted to update the parameters of the graph convolution neural network model:

wherein, θ represents a parameter set of the graph convolution neural network model, η represents a learning rate, and the learning rate needs to be preset and is used for controlling the speed of parameter updating.

Specifically, after the determination of the mapping relationship or the training of the model is completed, the temperature sensors arranged on the temperature monitoring points are used for acquiring temperature information in real time according to different working conditions of the system in the spacecraft cabin, and the temperature of the concerned point or the temperature field of the distribution area in the spacecraft cabin is predicted and set by using the determined mapping relationship or the trained model according to the acquired temperature information of the temperature monitoring points, so that the reconstruction of the temperature field in the spacecraft cabin is completed.

Further, in an embodiment of the present invention, the benchmark method for researching the temperature field reconstruction task in the spacecraft cabin based on machine learning may further include:

and evaluating the performance of the built spacecraft cabin temperature field based on the set performance evaluation indexes.

Specifically, in an embodiment of the present invention, a plurality of different spacecraft temperature field reconstruction task performance evaluation indexes are provided for the layout characteristics of an spacecraft cabin system, the cabin component prediction performance, and the reconstruction performance at the boundary of a layout region, respectively, and include: the temperature field average absolute error, the maximum temperature field absolute error, the component area average absolute error, the component area maximum absolute error and the boundary average absolute error are 5 different performance evaluation indexes.

Specifically, the following are set: omega, omega _c 、Ω _b Respectively representing the layout area, component area and boundary area of the spacecraft inboard system, T (x) _i ,y _j ) Indicating spacecraft inboard (x) _i ,y _j ) True temperature value of (d), T' (x) _i ,y _j ) Representing (x) predicted by means of a mapping or model _i ,y _j ) The predicted temperature value of (a).

The average absolute error of the temperature field is used for measuring the average value of the absolute error of the temperature field, and can be represented as:

the maximum absolute error of the temperature field is used for measuring the maximum absolute error of the temperature field, and can be expressed as follows:

the component area mean absolute error is used to measure the average of the absolute errors of the temperature on the heat source component, and can be expressed as:

the maximum absolute error of the component area is used for measuring the maximum value of the temperature prediction absolute error on the heat source component, and can be expressed as:

the boundary mean absolute error is used to measure the mean of the absolute errors of the temperature on the boundary, and can be expressed as:

the spacecraft cabin temperature field reconstruction task research benchmark method based on machine learning provided by the embodiment of the invention can utilize the data of limited temperature monitoring points to quickly predict the temperature of other position points and/or the whole temperature field of the whole area, realize the real-time, quick and high-precision reconstruction of the spacecraft cabin temperature field, has fewer required temperature monitoring points, effectively reduces the demand of a spacecraft platform on related hardware resources, reduces the resource consumption and the equipment cost, can fully assist a spacecraft in temperature monitoring and control, and promotes the research of the spacecraft cabin temperature field reconstruction.

It is noted that, in this document, relational terms such as "first" and "second," and the like, are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation 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 may include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, "front", "rear", "left", "right", "upper" and "lower" in this document are referred to the placement states shown in the drawings.

Finally, it should be noted that: the above examples are only for illustrating the technical solutions of the present invention, and not for limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A benchmark method for researching a temperature field reconstruction task in an aircraft cabin based on machine learning is characterized by comprising the following steps:

determining boundary conditions of a layout area in an aircraft cabin and power distribution conditions and structures of heat source components;

determining the mapping relation from the position of any point in the layout area to the temperature of the point by an interpolation method by utilizing the first training data; or training a constructed traditional machine learning model or a neural network model by utilizing first training data to fit a mapping relation from the position of any point in the layout area to the temperature of the point; or training the constructed multilayer perceptron by utilizing second training data to fit the mapping relation from the temperature of the temperature monitoring points in the layout area to the temperature of the set attention point; or generating a corresponding temperature monitoring matrix according to the position and temperature information of the temperature monitoring points in the layout area in the third training data, and utilizing the temperature monitoring matrix and a deep neural network model constructed by the temperature field training in the third training data to fit the mapping relation from the temperature monitoring matrix to the temperature field; or a graph model is constructed according to the position relation between the temperature monitoring points and the set attention points, and a graph convolution neural network model constructed by training is trained by utilizing the temperature of the temperature monitoring points in the graph model and the fourth training data and the temperature of the set attention points so as to fit the mapping relation between the graph model and the temperature of the temperature monitoring points to the temperature of the set attention points;

2. The machine learning-based spacecraft indoor temperature field reconstruction task research benchmark method according to claim 1, characterized by setting: the layout area in the spacecraft cabin is provided with a plurality of heat source components, the power distribution of the ith heat source component is phi _i (x, y) M temperature monitoring points provided with temperature sensors are arranged in the layout area, and the position of the mth temperature monitoring point is (x) _sm ,y _tm )；

wherein T represents the temperature field of the reconstructed layout area,

in the temperature field representing the reconstruction

Temperature of the location, O _m The temperature monitoring value of the mth temperature monitoring point is shown, k represents the heat conduction coefficient, x and y represent the position coordinate of a certain point in the temperature field, T ₀ Denotes the temperature value at the isothermal boundary, n denotes the normal perpendicular to, h denotes the thermal convection coefficient, T = T ₀ Indicating the boundary conditions of a Dirichlet,

representing the conditions of the Neumann boundary,

representing the Robin boundary conditions.

3. The machine learning-based spacecraft indoor temperature field reconstruction task research benchmark method according to claim 1, wherein the selecting temperature monitoring points for arranging temperature sensors in a spacecraft cabin by using a cluster analysis algorithm comprises:

step S33, the number of the temperature monitoring points is specified, and K-Means clustering is utilizedAlgorithm for N with multiple dimensions ₁ ×N ₂ Clustering the temperature points in the spacecraft cabin to obtain a plurality of categories with the same number as the designated temperature monitoring points, calculating the generalized distance based on Euclidean distance measurement, selecting the temperature point closest to the clustering center of each category as a representative of the corresponding category, and taking the selected temperature point and the corresponding position as the temperature monitoring point and the corresponding position.

4. The machine-learning-based spacecraft intra-cabin temperature field reconstruction task research benchmark method according to claim 1, characterized in that the first training data is acquired in the following manner:

determining the position of a temperature monitoring point in a layout area;

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition aiming at each heat source component in the layout area, determining the temperature of each temperature monitoring point under the layout of the current specific working condition, obtaining first training data comprising the position and the temperature of the temperature monitoring point in the layout area, and repeating the process for multiple times until first training data with preset quantity are obtained;

acquiring the second training data in the following way:

randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component aiming at each heat source component in the layout area to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point and each set attention point under the layout of the current specific working condition to obtain second training data comprising the temperature monitoring points and the set attention points in the layout area, and repeating the process for multiple times until second training data with the preset number are obtained;

obtaining the third training data in the following manner:

acquiring the fourth training data in the following way:

and aiming at each heat source component in the layout area, randomly sampling a power value from the power distribution of each heat source component as the current power of the corresponding heat source component to obtain the layout of a specific working condition, determining the temperature of each temperature monitoring point and each set attention point under the layout of the current specific working condition, obtaining fourth training data comprising the position and the temperature of the temperature monitoring point in the layout area and the position and the temperature of the set attention point in the layout area, and repeating the process for multiple times until obtaining the preset number of fourth training data.

5. The machine learning-based spacecraft cabin temperature field reconstruction task research benchmark method of claim 1, wherein the interpolation method is one of a k-neighbor nonlinear interpolation method and a global Gaussian interpolation method;

when the interpolation method is a k-nearest neighbor nonlinear interpolation method, the mapping relationship from the position of any point in the layout area to the temperature of the point can be expressed as:

when the interpolation method is a global gaussian interpolation method, the mapping relationship between the position of any point in the layout area and the temperature of the point can be expressed as:

wherein, T (x) ₀ ,y ₀ ) Indicates the set point of interest (x) ₀ ,y ₀ ) Predicted temperature value of, S _k (x ₀ ,y ₀ ) Indicates a distance setting focus (x) ₀ ,y ₀ ) The most recent k temperature monitoring points are,

indicating the jth temperature monitoring point at which,

representing the temperature value of the ith temperature monitoring point.

6. The machine learning-based spacecraft cabin temperature field reconstruction task research benchmark method of claim 1, wherein the traditional machine learning model is one of a polynomial regression model, a random forest regression model, a gaussian process regression model, and a support vector machine regression model.

7. The machine-learning-based spacecraft inboard temperature field reconstruction task research benchmark method of claim 1, wherein the neural network model is one of a multi-layer perceptron, a constrained boltzmann machine, and a deep belief network;

8. The machine learning-based spacecraft in-cabin temperature field reconstruction task research benchmark method according to claim 1, wherein the multi-layer perceptron trained and constructed by utilizing second training data comprises:

9. The machine-learning-based spacecraft intra-cabin temperature field reconstruction mission research benchmark method of claim 1, wherein the deep neural network model is one of a full convolution neural network model, a feature pyramid network model, a SegNet network model, and a UNet network model;

10. The machine learning-based spacecraft intra-cabin temperature field reconstruction task research benchmark method according to claim 1, wherein the constructing of the graph model according to the position relationship between the temperature monitoring points and the set attention points comprises: