Asteroid structure estimation model based on spacecraft flight data and prediction method thereof

Abstract

The invention discloses an asteroid structure estimation model based on spacecraft flight data and a prediction method thereof, and the method comprises the steps: directly predicting the internal structure of an asteroid through employing the position and speed sequence of a spacecraft, and combining a reversible neural network INNs, thereby constructing an asteroid structure estimation model, namely a DINNs model; the DINNs model is embedded in an on-board data processing center of the deep space exploration spacecraft, and when the DINNs model is applied to on-orbit real-time asteroid structure intelligent prediction, the spacecraft on-orbit asteroid structure prediction can be completed in a short time. According to the method, the asteroid structure can be predicted in real time, the detector can be helped to autonomously select high-value sampling points, and real-time decision making and autonomous control of the detector are served. The method provided by the invention has the characteristics of strong end-to-end fitting capability and relatively perfect framework and hardware foundation. According to the method, the end-to-end estimation of the parameters of the internal structure of the asteroid is realized by using DINNs for the first time, and a characterization extraction method for predicting the internal structure of the asteroid is provided.

Description

Technical Field

[0001] The present invention relates to a method for predicting an asteroid structure, and more particularly, to an asteroid structure estimation model and a prediction method based on spacecraft flight data (position and velocity), which are applied to deep space exploration spacecraft. Background Art

[0002] Small solar system bodies such as asteroids and comets are believed to be the remnants of the early formation of the solar system, and can provide important clues for major basic scientific research such as the origin of planets, the evolution of the solar system, and the origin of life. Regarding small celestial bodies, scientists have proposed cutting-edge scientific issues such as "the origin of organic matter, the distribution and source of water, the dynamic formation and evolution, and the collision threat of near-Earth asteroids", which strongly attract scientists' enthusiasm for exploration. The implementation of small celestial body exploration missions also involves a series of key technologies such as space propulsion technology, space energy technology, small celestial body surface attachment technology, and sampling technology under microgravity conditions. The long exploration mission cycle, long target distance, and moderate target size make small celestial body exploration a unique testing ground for these new technologies. The US "Deep Space 1" probe and the Japanese "Hayabusa" probe have both achieved important results in the demonstration and verification of new technologies. The verified technologies have also provided important technical guarantees for countries in subsequent deep space exploration missions.

[0003] At present, studies have shown that it is difficult to directly obtain information about the internal structure of asteroids through instruments, but astronomical observations can give possible a priori estimates of its internal structure through information such as the rotation of the asteroid. For example, it is estimated that the Itokawa asteroid has about 40% pores. In December 2018, the 27th volume of "Spacecraft Engineering" published "Research on Autonomous Operation Technology of Deep Space Probes", authored by Wang Datie et al., in Figure 1 A typical detector autonomous planning system is introduced in .

[0004] At present, the research on the estimation of the internal structure of asteroids at home and abroad mainly focuses on three points: the first is the modeling method of the internal structure of asteroids, the second is what kind of data to use for perception, and the third is what method to optimize. At present, the classic description methods for the internal structure of asteroids mainly include the Mascon method, the polyhedron method and the layered method. Among them:

[0005] The Mascon method uses astronomical observations and detector cameras to obtain a relatively accurate three-dimensional shape of the asteroid through image processing, and then fills the interior with mass units of the same scale without overlapping. The mass unit is usually a sphere or a cube. Researchers can choose the size of the mass unit independently based on factors such as computational complexity and degrees of freedom. The advantages of the Mascon method are large modeling fineness and strong descriptive ability, but the disadvantages are low computational efficiency and it is difficult to directly convert the spherical harmonic function expression corresponding to the external gravitational field.

[0006] The polyhedron method also requires the use of a 3D reconstruction model of the asteroid, with the asteroid center of mass as the anchor point, the outer contour as the base, and pyramid division, with the order of the pyramid as the center of mass. The advantage of the polyhedron method is that it can easily write the corresponding spherical harmonic function expression of the gravitational field, and the calculation efficiency is high, but the disadvantage is that the division fineness is not uniform.

[0007] The layered approach usually divides asteroids into two or three layers, each layer is modeled as an ellipsoid or sphere. The advantage of the layered approach is that the model is relatively basic and can be evolved into a complex model.

[0008] In summary, the prediction of asteroid structure adopts the process of deep space exploration spacecraft → ground probe → the classic description method of asteroid internal structure, and finally obtains the estimated asteroid internal structure. The problem of perceiving the internal structure of an asteroid through data is a recognized ill-posed problem. Whether it is through flight data (Position, Velocity) or gravity field data (Gravity field), it is necessary to deal with multiple solutions and instability. This requires the use of more prior information to constrain possible solutions, such as obtaining information on asteroid rotation and pores through astronomical observations and obtaining accurate geographic information on the outer surface of the asteroid through the probe camera. Summary of the invention

[0009] One of the purposes of the present invention is to design an asteroid structure estimation model based on spacecraft flight data, namely, the DINNs model. The DINNs model is completed on the ground, that is, only the parameters trained in the DINNs model need to be deployed on the onboard computer, and no gradient descent training is required on the onboard computer. When constructing the DINNs model, the present invention uses the hierarchical method as the dynamic model. In the process of constructing the DINNs model, the position and velocity of the spacecraft are selected as training data, and its fundamental purpose is to construct a sequence of different sampling points to pave the way for on-orbit real-time inversion in on-orbit missions. In addition, an end-to-end inversion algorithm is established, which can realize the on-orbit inversion of asteroid structures that all previous estimation algorithms cannot achieve, lay the foundation for the autonomous selection of high-value sampling points, and serve the interplanetary in-situ resource development; thirdly, the reversible neural network (INNs) model is improved according to the dynamic characteristics of the spacecraft flying around the asteroid, and the prior modeling based on physical characteristics is integrated into the INNs model architecture level to achieve accurate prediction of the internal structure parameters of the asteroid.

[0010] The second purpose of the present invention is to use the DINNs model in the spacecraft's on-orbit mission to complete the prediction of the asteroid structure. The difference between the characterization extraction method developed based on the invertible neural network (INNs) and the classical algorithm is that the training of the DINNs model of the present invention is completed on the ground, that is, only the parameters trained by the DINNs model need to be deployed on the onboard computer, and no gradient descent training or optimization is required on the onboard computer. Only reasoning is required, that is, after the instrument collects the flight status (position, speed) of the probe, the inverse mapping of the DINNs model can be used to obtain the prediction of the asteroid structure, which is the on-orbit real-time estimation that the traditional method cannot do.

[0011] The technical effects achieved by the present invention in constructing the DINNs model are:

[0012] (1) The present invention adopts the layered method to model asteroids based on the fact that asteroids of a certain scale are generally layered. The advantage is that the dimensions of the asteroid structural parameters are reasonably simplified, which is conducive to avoiding the situation where multiple solutions of parameters or the optimization effect is not obvious during the inversion process. The other two commonly used methods, the finite element method and the polyhedron method, both have obvious disadvantages. The disadvantage of the finite element method is that the mass units are independent and the degree of freedom is too large, resulting in the system being unobservable or poorly observable. The defect at the modeling level is that the correlation of mass units is not considered, and the fact that the density of celestial bodies changes little or remains unchanged within a certain continuous range is ignored; the advantage of the polyhedron method is that the asteroid can be completely divided and the shape characteristics can be clearly described, but the disadvantage is that the volume of the pyramid near the vertex side and the bottom side is greatly different, resulting in a large difference in the fineness of the division near the centroid side and the fineness near the surface side, and it cannot be solved at the modeling level.

[0013] (2) The advantage of selecting the flight state of the spacecraft as data instead of the classic gravitational field sampling is that the flight state is time-series. Using the flight state as the data source is conducive to the time-series modeling of the model, so that the probe can complete the real-time inversion of the internal structure of the asteroid during the data collection process. Selecting gravitational field sampling can directly impose gravitational potential constraints. The usual processing method is to use spherical harmonics to fit the asteroid's gravitational field distribution after sampling the asteroid's external gravitational field parameters, and use the coefficients of the spherical harmonics as the optimization target.

[0014] (3) The first advantage of the DINNs model based on the improved INNs architecture is that it can directly estimate the structural parameters of asteroids through end-to-end reversible nonlinear mapping. There is a nonlinear mapping between the structural parameters of asteroids and the flight state of spacecraft in physics that cannot be expressed explicitly. Since this mapping is irreversible, it is very difficult to solve this inverse problem through traditional algorithms such as least squares estimation and iterative root finding of nonlinear equations. The biggest feature of DINNs is that it can establish a reversible nonlinear mapping between two sets of data with unequal dimensions, and through model training, it can approximate the original irreversible physical process. Due to the bidirectional mapping nature of DINNs, the flight state of the spacecraft can directly obtain an estimate of the structural parameters of the asteroid through reverse mapping. The second advantage is that the design of the coding layer effectively avoids excessive reliance on flight data at a single moment. The flight status measured by the spacecraft through the instrument has inevitable errors. The original INNs model uses the flight status at a single moment, while DINNs uses the coding layer to first fuse the features of the flight status sequence and extract the joint representation of the flight status at different moments, so as to avoid excessive reliance on the flight status observation at a single moment, making the estimation of the asteroid structure parameters more stable and accurate. The third advantage is that through the combination of ground training and on-board reasoning, the trained DINNs model can be deployed directly on the computing equipment on the spacecraft before the spacecraft is launched. After the spacecraft arrives at the mission target asteroid and collects the flight status data, it can directly make real-time predictions on orbit, which improves the situational awareness of the spacecraft and lays the foundation for real-time and autonomous control of deep space spacecraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a block diagram of the detector autonomous planning system.

[0016] Figure 2 It is a schematic diagram of the processing signal flow in the deep space detection system structure of the present invention.

[0017] Figure 3 It is a ground training flow chart for constructing the DINNs model according to the present invention.

[0018] Figure 4 It is a deep space exploration flow chart of on-board reasoning using the DINNs model of the present invention.

[0019] Figure 5 This is a schematic diagram of the asteroid's body coordinate system.

[0020] Figure 6 It is a flow chart of coding representation in the kinetic model of the present invention.

[0021] Figure 7 It is a flow chart of the asteroid structure training model of the present invention based on invertible neural networks (INNs).

[0022] Figure 8 is the test set L of INNs model and DINNs model y (L_fwd_fit) loss comparison chart.

[0023] Fig. 9 is the test set L of INNs model and DINNs model z (L_fwd_mmd) loss comparison chart.

[0024] Fig.10 is the test set L of INNs model and DINNs model x (L_bwd_mmd) loss comparison chart.

[0025] Fig.11 This is a comparison chart of the parameter estimation accuracy of the DINNs model and the INNs model.

[0026] Fig.12 is the DINNs model parameter a mantle Estimation graph.

[0027] Fig.13 is the parameter b of the DINNs model mantle Estimation graph.

[0028] Fig.14 is the DINNs model parameter c mantle Estimation graph of .

[0029] Fig.15 is the parameter ρ of the DINNs model mantle Estimation graph of .

[0030] Fig.16 is the parameter ρ of the DINNs model crust Estimation graph. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The various parameter demonstrations listed are only preferred embodiments of the present invention. It should be noted that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention.

[0032] See also Figure 2 , Figure 3 , Figure 4As shown, the present invention uses the original spacecraft flight data (at least position and velocity information) to construct an asteroid structure estimation model (DINNs model) on the ground, and on the other hand, uses the DINNs model to invert and predict the internal structure characteristics of the asteroid in real time when the spacecraft performs an on-orbit mission. The DINNs architecture is an improvement on the INNs architecture, and the reversible neural network (INNs) architecture used is a known technology. However, after the INNs network is combined with the asteroid structure parameters, information processing is mainly performed according to the arrangement and combination of the left-end variable-reversible flow module-right-end variable.

[0033] In the present invention, the asteroid structure prediction model constructed based on the invertible neural network (INNs) architecture is referred to as the DINNs model.

[0034] Part 1: Ground-based modeling of asteroid structure estimation

[0035] The biggest feature of reversible neural networks (INNs) is that they can establish reversible nonlinear mappings between two sets of data with unequal dimensions. The internal structure of the INNs network is a stack of stream modules, the core of which is the Affine Coupling Layer, which can have built-in neural networks to enhance nonlinear fitting capabilities ("Modelling internal structure of differentiated asteroids via data-driven approach" published in Monthly Notices of the Royal Astronomical Society on November 24, 2022, author Liang Yuying. In Section 3, it is introduced that INNs are composed of continuous submodules Affine Coupling Layers are connected end to end. The principle is to divide the input data into two parts. The first part is directly output to the next layer without processing, similar to residual connection; the second part is output through a designable neural network, and then the two sets of data are exchanged in the next submodule, and nonlinear mapping is performed alternately. Based on the flow module, INNs networks that can solve practical problems in science and engineering can be developed. The design goal of the INNs network is to be able to infer the parameter space from the measurement space to solve the inverse problem in science or engineering. The inverse problem is a scientific or mathematical challenge, where the goal is to infer the potential causes of a specific phenomenon or data set. Unlike direct predictions from cause to effect, the reverse problem operates in the opposite direction, trying to infer the original conditions from the result observation. These problems are usually difficult due to non-uniqueness and instability, where multiple causes may lead to similar results and small changes in data can greatly change the solution. Since the flow module can only establish a reversible nonlinear mapping between two sets of data with equal dimensions, but it is difficult to meet in actual situations, the INNs network adds compensation variables to balance the dimensional differences of input and output for this problem.

[0036] In the present invention, the ground construction of the DINNs model includes the following steps:

[0037] Construction step 1, asteroid dynamics modeling;

[0038] Figure 5 O represents the center of mass of the asteroid, the Z axis is the rotation axis of the celestial body, and the O-XYZ setting satisfies the right-hand coordinate system. The direction of the X axis is consistent with the direction of the a1 axis of the asteroid's internal ellipsoid layer, the direction of the Y axis is consistent with the direction of the b1 axis of the asteroid's internal ellipsoid layer, and the direction of the Z axis is consistent with the direction of the c1 axis of the asteroid's internal ellipsoid layer. The geometric parameters of the asteroid's internal ellipsoid are expressed as {a1, b1, c1}.

[0039] The identification number of the asteroid is recorded as AES. The outline information of any asteroid AES observed by an astronomical telescope on the ground is recorded as AE 观测 . The AE 观测 It contains the information of the point set and face set of the AES.

[0040] Construction step 11, characterization of the two-layer ellipsoid model;

[0041] Based on the asteroid body coordinate system O-XYZ, a layered dynamic model of the asteroid is established. This dynamic model uses a two-layer ellipsoid model as a simplified dynamic model of the asteroid. Since the external images of the asteroid can be obtained through ground observation and probe cameras, and a relatively accurate asteroid shape (i.e., AE) is established using three-dimensional reconstruction technology 观测 ), therefore, in the present invention, the geometric parameters of the outermost ellipsoid are set to constants.

[0042] In the present invention, the asteroid dynamics model is modeled by a two-layer ellipsoid model. Its gravitational potential U AES (x, y, z) is obtained by adding the inner ellipsoid and the outer ellipsoid:

[0043] U AES (x,y,z)=U crust (x,y,z)+U mantle (x,y,z) (1)

[0044] U crust (x,y,z) represents the gravitational potential of the inner ellipsoid to any point in the space with coordinates O-XYZ.

[0045] U mantle (x,y,z) represents the gravitational potential of the outer ellipsoid to any point in the space with coordinates O-XYZ.

[0046] From the ellipsoid gravitational potential, we know that the gravitational potential energy of the ellipsoid core can be described by the density ρ of the parameter core and the geometric parameters {a1, b1, c1}, a total of 4 independent variables. The gravitational potential energy of the outer ellipsoid (shell ellipsoid layer) can be calculated by subtracting the gravitational potential energy of the inner ellipsoid (mantle ellipsoid layer), which also requires 4 independent variables ρ crust ,a crust ,b crust ,c crust , specifically expressed as:

[0047]

[0048] MU crust represents the gravitational potential energy of the ellipsoidal nucleus.

[0049] U crust (ρ crust ,acrust ,b crust ,c crust ) represents the gravitational potential energy of the shell ellipsoid.

[0050] U mantle (ρ mantle ,a mantle ,b mantle ,c mantle ) represents the gravitational potential energy of the mantle ellipsoid.

[0051] ρ crust represents the shell ellipsoid layer density.

[0052] a crust Represents the a1 axis in the main axis of the shell ellipsoid layer.

[0053] b crust , represents the b1 axis in the main axis of the shell ellipsoid layer.

[0054] c crust Represents the c1 axis in the major axis of the shell ellipsoid layer.

[0055] ρ mantle represents the density of the mantle ellipsoid layer.

[0056] a mantle Represents the a1 axis in the main axis of the mantle ellipsoid layer.

[0057] b mantle Represents the b1 axis in the main axis of the mantle ellipsoid layer.

[0058] c mantle Represents the c1 axis in the main axis of the mantle ellipsoid layer.

[0059] Construction step 12, obtaining asteroid physical parameter training set;

[0060] Considering U crust The geometric parameters (x, y, z) can be obtained through observation (probe camera and astronomical observation). In the two-layer ellipsoid structure, the AE of the asteroid 观测 can be used as known information, so the physical information of the asteroid can be expressed using five independent parameters (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle )express.

[0061] In the present invention, the asteroid physical parameter set is set to Φ, and Φ = {IS1, IS2, ..., IS h ,…,IS H}, the subscript h represents the identification number of the asteroid structure parameter set, and the subscript H represents the total number of asteroid structure parameter sets.

[0062] IS1 represents the first set of asteroid structure parameter sets.

[0063] IS2 stands for the second set of asteroid structure parameter sets.

[0064] IS h represents the hth set of asteroid structure parameters.

[0065] IS H Represents the last set of asteroid structural parameters.

[0066] For the sake of convenience, IS h Also known as any set of asteroid structural parameters. Belongs to information set IS h The structural parameters of the asteroid are expressed in the form of a set:

[0067] The subscript i indicates that it belongs to the information set IS h An identification number for the asteroid's structural parameters.

[0068] Indicates that it belongs to the information set IS h The first asteroid structure parameters.

[0069] Indicates that it belongs to the information set IS h The second asteroid structure parameters.

[0070] Indicates that it belongs to the information set IS h The structural parameters of the i-th asteroid.

[0071] Indicates that it belongs to the information set IS h The structural parameters of the j-th asteroid.

[0072] Indicates that it belongs to the information set IS h The 1024th asteroid structural parameter, also from the information set IS h The last asteroid structure parameters.

[0073] For the sake of convenience, Also known as information set IS h Any asteroid structural parameter. Contains the quintuple (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle ) information, in order to distinguish the description clearly, it is also expressed as

[0074] Similarly The five-tuple information contained in is recorded as Said The five-tuple information contained in is recorded as The is j The five-tuple information contained in is recorded as Said The five-tuple information contained in is recorded as

[0075] In order to show the structural parameters of asteroids belonging to information set IS1, the set form is represented as

[0076] In order to show the structural parameters of asteroids belonging to information set IS2, the set form is represented as

[0077] In order to show that the information set IS h The asteroid structural parameters are expressed in the form of a collection of

[0078] In order to show that the information set IS H The asteroid structural parameters are represented by

[0079] In the present invention, generally, 1024 (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle ) such independent parameters.

[0080] Construction step 2: Generate the orbital information training set required by the DINNs model;

[0081] Construct step 21, obtain the flight status training set of the asteroid;

[0082] In the present invention, all flight states of deep space exploration spacecraft flying around asteroids are recorded as spacecraft-asteroid-training set FSSO, and FSSO = {FSS1, FSS2, ..., FSS h ,…,FSS H}, the subscript h represents the identification number of the asteroid structure parameter set, and the subscript H represents the total number of asteroid structure parameter sets.

[0083] FSS1 represents the asteroid flight state set corresponding to the first group of asteroid structure parameter set IS1.

[0084] FSS2 represents the asteroid flight state set corresponding to the second group of asteroid structure parameter set IS2.

[0085] FSS h represents the hth group of asteroid structural parameter set IS h The corresponding asteroid flight status set.

[0086] FSS H Represents the last set of asteroid structural parameter set IS H The corresponding asteroid flight status set.

[0087] Flight Status Set (FSS) h The asteroid flight state parameters are expressed in set form as follows: The subscript f indicates the sampling point identification number, and the subscript S indicates the total number of sampling points. Generally, S is 1024.

[0088] Indicates that it belongs to the flight status information set FSS h The first flight state sampling.

[0089] Indicates that it belongs to the flight status information set FSS h The second flight state sampling.

[0090] Indicates that it belongs to the flight status information set FSS h The f-th flight state sampling. For the convenience of explanation, the Also known as Flight Status Set (FSS) h Any flight state sampling.

[0091] Indicates that it belongs to the flight status information set FSS h The s-th flight state sampling.

[0092] Indicates that it belongs to the flight status information set FSS h The 1024th flight status sample also belongs to the flight status information set FSS h The last flight status sampling.

[0093] Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component along the Z axis.

[0094] Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component along the Z axis.

[0095] Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component along the Z axis.

[0096] Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component along the Z axis.

[0097] Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis.

[0098] In order to show the flight status sampling of asteroids belonging to FSS1, the collection form is represented as

[0099] In order to show the sampling of the asteroid flight state belonging to FSS2, the set form is represented as

[0100] In order to show the FSS h The asteroid flight state sampling, the collection form is represented as

[0101] In order to show the FSS H The asteroid flight state sampling, the collection form is represented as

[0102] Construct step 22, pairing of parameter space;

[0103] The opposite of parameter space is state space, where the position and velocity parameters of the detector are Since the orbit is selected inside the gravitational sphere, the gravity of the central celestial body is dominant, and factors such as the gravity of other celestial bodies and the solar pressure can be used as perturbations to approximately model the error. Therefore, under given initial conditions, the probe orbit is determined by the gravitational field of the central celestial body, and the gravitational field of the asteroid is Φ = {IS1, IS2, …, IS h ,…,IS H}, so and There is a complex mapping based on physical rules. In the present invention, the rotation angular velocity ω of the central celestial body is not considered for the time being, and the complex mapping differential equation is formally expressed as:

[0104]

[0105] Represents the derivative of the flight data vector of a deep space exploration spacecraft.

[0106] It represents the time derivative of the X-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0107] It represents the time derivative of the Y-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0108] It represents the time derivative of the Z-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0109] It represents the time derivative of the X-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0110] It represents the time derivative of the Y-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0111] It represents the time derivative of the Z-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0112] V x It represents the X-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0113] V y It represents the Y-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0114] V z It represents the Z-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0115] α x It represents the acceleration component of the deep space exploration spacecraft on the X-axis in the O-XYZ coordinate system.

[0116] α y It represents the Y-axis acceleration component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0117] α z It represents the Z-axis acceleration component of the deep space exploration spacecraft in the O-XYZ coordinate system.

[0118] Represents the partial differential operator.

[0119] U(x,y,z,Φ) represents the gravitational potential at any position (x,y,z) in the O-XYZ coordinate system.

[0120] It means partial derivative with respect to x.

[0121] It means partial derivative with respect to y.

[0122] It means partial derivative with respect to z.

[0123] According to formula (3) and Mapping, get the mapped spacecraft flight status

[0124] According to formula (3) and Mapping, get the mapped spacecraft flight status

[0125] According to formula (3) and Mapping, get the mapped spacecraft flight status

[0126] According to formula (3) and Mapping, get the mapped spacecraft flight status

[0127] Construction step 23, addition of Gaussian error;

[0128] During the movement of deep space exploration spacecraft, the orbital information collected in real time will inevitably have errors due to instrument reasons. Therefore, in the training process of model construction, it is necessary to Add errors to enhance the robustness of the model building process, Figure 3 , Figure 6 shown.

[0129] After adding the Gaussian error GAU, the flight training set of asteroids is formed, denoted as

[0130] express The flight status after Gaussian error GAU. The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, It is the velocity component on the Z axis after adding the Gaussian error GAU.

[0131] express The flight status after Gaussian error GAU. The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, It is the velocity component on the Z axis after adding the Gaussian error GAU.

[0132] express The flight status after Gaussian error GAU. The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, It is the velocity component on the Z axis after adding the Gaussian error GAU.

[0133] express The flight status after Gaussian error GAU. The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, It is the velocity component on the Z axis after adding the Gaussian error GAU.

[0134] express The flight status after Gaussian error GAU. The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, It is the velocity component on the Z axis after adding the Gaussian error GAU.

[0135] In the present invention, FSSO = {FSS1, FSS2, ..., FSS h ,…,FSS H After the Gaussian error GAU, the coded spacecraft-asteroid-flight status code set is formed, which is denoted as GFSSO = {GFSS1, GFSS2, …, GFSS h ,…,GFSS H}.

[0136] Construction step three, based on the improvement of the INNs architecture and parameter training adapted to asteroid structures;

[0137] In the present invention, INNs refer to "Modelling internal structure of differentiated steroids via a data-driven approach" published in "Monthly Notices of the Royal Astronomical Society" on November 24, 2023, written by Liang Yuying.

[0138] Based on the INNs architecture As the left variable of the asteroid structure prediction network.

[0139] Based on the INNs architecture As the first variable on the right side of the asteroid structure prediction network.

[0140] Based on the INNs architecture, Gaussian distributed random sampling GSS is used as the second variable on the right side of the asteroid structure prediction network.

[0141] For the forward process, the asteroid structure parameter (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle ) is input from the left end to the INNs model. First, the dimensions are compensated. Each compensated dimension is a sample of the unit Gaussian distribution. A total of 5 dimensions are compensated, namely G1, G2, G3, G4, and G5. The left input after compensation is (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle , G1, G2, G3, G4, G5). Then it passes through 10 reversible nonlinear mapping modules with the same configuration and is output. For each module, the left input variable (vector) is first divided into two independent parts (vectors) from the middle, namely the first left input sub-part and the second left input sub-part. For the first layer, the first left input sub-part is (ρ crust ,ρ mantle ,a mantle ,b mantle ,c mantle), the second left-end input sub-part is (G1, G2, G3, G4, G5). Two operations are performed on the first left-end input sub-part, and one operation is performed on the second left-end input sub-part. First, the first left-end input sub-part is directly used as the first right-end output sub-part without any processing. Second, the first left-end input sub-part is added to the second left-end input sub-part through the fully connected layer to form the second right-end output sub-part. The fully connected layer consists of three linear mapping layers and two nonlinear activation layers (Gelu activation). The linear mapping layers and nonlinear activation layers are placed alternately. The first right-end output sub-part (vector) and the second right-end output sub-part (vector) are spliced ​​into the right-end output variable (vector), and then the right-end output variable is used as the left-end input variable of the next reversible nonlinear mapping module for the same operation. The above process is performed 10 times in total, and then the forward process output of INNs is obtained. The physical meaning of the output is the spacecraft flight state encoding sequence of the training process. Estimates.

[0142] For the reverse process, the spacecraft flight state sequence The Gaussian sampling vector sequence GSS is input from the right end to the INNs model. First, the spacecraft flight state and the Gaussian sampling vector sequence are encoded, and the encoded vector sequence is output as the spacecraft flight state encoding sequence of the training process. and Gaussian sampling code sequence GSS 训练 . Then, each vector in the sequence is sequentially passed through 10 reversible nonlinear mapping modules with the same configuration from the right end. The vector operations in the reversible nonlinear mapping module in the reverse process are inverse to the vector operations in the reversible nonlinear mapping module in the forward process. In the reverse process, for each module, the right-end input variable (vector) is first divided into two independent parts (vectors) from the middle, named the first right-end input sub-part and the second right-end input sub-part. Two operations are performed on the first right-end input sub-part, and one operation is performed on the second right-end input sub-part. First, the first right-end input sub-part is directly used as the first left-end output sub-part without any processing. Second, the first right-end input sub-part is passed through the fully connected layer and subtracted from the second right-end input sub-part to form the second left-end output sub-part. The fully connected layer consists of three linear mapping layers and two nonlinear activation layers (Gelu activation), and the linear mapping layers and nonlinear activation layers are placed alternately. The first left-end output sub-part (vector) and the second left-end output sub-part (vector) are spliced ​​into the left-end output variable (vector), and then the left-end output variable is used as the right-end input variable to enter the next reversible nonlinear mapping module, and then the right-end input variable is used as the right-end input variable of the next reversible nonlinear mapping module to perform the same operation. The above process is performed 10 times in total, and then the reverse process output of INNs is obtained. The physical meaning of the output is the asteroid structure parameter (ρcrust ,ρ mantle ,a mantle ,b mantle ,c mantle )'s reconstruction.

[0143] Constructing step 31, compensation of the neural network;

[0144] An obvious disadvantage of the original INNs model is that the input and output design of the INNs model is not completely consistent with the possible time solutions of the differential equations corresponding to the physical problem. Specifically, there is a lack of consideration of the time dimension. The existing INNs model fits the fixed structural parameters to the orbit sampling points at different times through a neural network. This design has certain loopholes. The mapping logic of the neural network is:

[0145]

[0146] Represents the forward mapping of INNs.

[0147] Indicates that Map to The flight status obtained above.

[0148] Said With the This is different in the original INNs model, which results in the original INNs model being unable to learn the physical constraints of a deep space exploration spacecraft flying around an asteroid. This defect is caused by the original INNs model not considering the temporal characteristics of the sampling points. Therefore, the present invention adopts a linear encoding method to perform a linear weighted mapping on the flight state sequence to model the temporal characteristics of the flight state of a deep space exploration spacecraft flying around an asteroid, which is recorded as the DINNs model.

[0149] The DINNs model of the present invention improves the output end in view of the shortcoming of the original INNs model that it cannot model the timing sequence, and encodes the flight state sequence to weaken or even eliminate the timing characteristics of the output end.

[0150] Constructing step 32, encoding of the forward mapping output;

[0151] To this end, the forward mapping output of the INNs model is improved by adding coding (such as Figure 6 The process shown in Figure 1 is for the flight state sequence Perform encoding processing.

[0152]

[0153] express Encoded flight status.

[0154] σ(·) represents a linear mapping. For elliptical motion under the central gravitational field of a two-dimensional plane, it can be known from the Lagrange coefficient that the flight states at different times can be associated using the state transfer matrix STM. Due to the influence of non-central gravitational perturbations, simple linear mapping relationships no longer hold, but they can be approximately represented by linear relationships after ignoring high-order terms under the assumption that the perturbations are small. Therefore, the linear correlation modeling of the attention mechanism (Attention) conforms to the physical background, and the representation obtained by modeling the attention mechanism has certain physical significance. After determining the coefficients, the representation is activated again. In the present invention, σ is selected as a linear activation.

[0155] β s,1 Represents the weight of the flight status of the sth sampling point to the flight status of the 1st sampling point.

[0156] β s,2 Represents the weight of the flight status of the sth sampling point to the flight status of the 2nd sampling point.

[0157] β s,f Represents the weight of the flight state of the s-th sampling point to the flight state of the f-th sampling point.

[0158] β s,s Represents the weight of the flight state of the s-th sampling point to the flight state of the s-th sampling point.

[0159] β s,S Represents the weight of the flight state of the sth sampling point to the flight state of the last sampling point.

[0160] The encoding process is linear combination + neural network activation. The linear combination weight is obtained through the attention mechanism and Softmax (translation: normalized exponential function). Weight β s,f Depend on and The dot product is then trained using Softmax normalization.

[0161] Construction step 33, linear normalization;

[0162] The essence of the correlation of Attention modeling is the linear mapping of linear normalized weighted combinations. From a physical perspective, the key to whether Attention modeling has physical significance lies in examining its weight coefficient β s,f The weighting coefficient β s,f Determination of flight status After linear transformation, dot product is realized. First, define two linear transformation matrices Q and K (Q and K are both learnable matrices, Q is the first learning matrix, and K is the second learning matrix). For the selected flight state Using the first learning matrix Q for learning, we can find Any subsequent flight state is learned using the second learning matrix K (select ), after linear transformation of the learning matrix Q, K, we can get:

[0163]

[0164] Represents the flight status after learning by the first learning matrix.

[0165] Represents the flight status after learning by the second learning matrix.

[0166] Construction step 4: training based on DINNs model;

[0167] Construction step 41, selection of training set and test set;

[0168] Set Φ = {IS1, IS2, …, IS h ,…,IS H} and GFSSO={GFSS1,GFSS2,…,GFSS h ,…,GFSS H} is divided into two parts from the middle, the first half is named training set (Φ Train , GFSSO Train ) and the second half is named the test set (Φ Test , GFSSO Test ), there is no overlap between the training set and the test set. Train , GFSSO Train ) is used to train the internal neural network parameters of the DINNs model and the mapping matrix of the encoding layer. Test set (Φ Test , GFSSO Test ) is used to test whether the trained model has a generalization effect. The test set does not participate in the training, that is, the neural network cannot receive the data of the test set during training, which ensures that the test set is meaningful for testing the performance of the DINNs model.

[0169] Constructing step 42, determining the hyperparameters of the DINNs model;

[0170] The hyperparameters of the DINNs model are determined. The hyperparameters that need to be confirmed in the DINNs model mainly include the batch size (Bs), the learning rate (Lr) and the maximum training round (Epoch). The present invention selects the batch size Bs = 1024 and the learning rate Lr = 10 -4 , the maximum training epoch is 300.

[0171] Construction step 43, DINNs model training and convergence judgment;

[0172] In the DINNs model training, the neural network parameters and the encoding layer parameters are randomly initialized. After each training round, the model obtained in this round of training will be saved. The models obtained from all training rounds are recorded as a set TMS = {DINNs1, DINNs2, ..., DINNs p ,…,DINNs 300}, record the loss function values ​​of the training set and test set corresponding to each round, and record them as the loss function set of the whole round training set TRLS = {TRL1,TRL2,…,TRL g ,…,TRL 300} and the set of trend functions of the full round test set TELS = {TEL1, TEL2, …, TEL k ,…,TEL 300}.

[0173] DINNs1 represents the model output after the first training round.

[0174] DINNs2 represents the model output after the second training round.

[0175] DINNs p Represents the model output after the pth training round.

[0176] DINNs 300 Indicates the model output after the last training round.

[0177] TRL1 represents the loss function value of the training set in the first training round.

[0178] TRL2 represents the loss function value of the training set in the second training round.

[0179] TRL g Represents the loss function value of the training set in the g-th training round.

[0180] TRL 300 Indicates the loss function value of the training set in the last training round.

[0181] TEL1 represents the loss function value of the test set in the first training round.

[0182] TEL2 represents the loss function value of the test set in the second training round.

[0183] TEL k Represents the loss function value of the test set in the g-th training round.

[0184] TEL 300 Indicates the loss function value of the test set in the last training round.

[0185] Then determine whether the model has converged. This requires examining two conditions, one of which is TRL. g As g increases, it approaches a constant CC; second, CC is less than TRL1. If the above two conditions are met, the DINNs model training is considered to have converged. If the above conditions are not met, the model is considered to have diverged. In this case, the neural network and encoding layer parameters are reinitialized and retrained until the convergence conditions are met. From the practical process, the model convergence is not sensitive to initialization. The model is robust and can converge under random initialization. However, the convergence effect of each training, that is, the value of the constant CC, is different. Generally speaking, the smaller the value of the constant CC, the better the training effect. Therefore, multiple rounds of training are often required for inspection.

[0186] Constructing step 44, obtaining the optimal solution of the DINNs model;

[0187] Compare TMS = {DINNs1, DINNs2, ..., DINNs p ,…,DINNs 300 The loss function value of each model in} on the test set, that is, the comparison set TELS = {TEL1, TEL2, …, TEL k ,…,TEL 300}, select the smallest value among them and record it as TEL Tar , and the optimal solution DINNs in the TMS set Tar Output as an estimate of the asteroid's structure.

[0188] In the present invention, the trained DINNs model is stored in the onboard computer of the deep space exploration spacecraft.

[0189] Part II: Application of DINNs model on deep space exploration spacecraft

[0190] See also Figure 2The structure diagram of the deep space exploration system shown in the figure, in which the asteroid structure prediction model (DINNs) first receives the on-orbit comprehensive information 11 output by the deep space exploration spacecraft; secondly, it outputs the estimated information 12 of the asteroid structure parameters to the deep space exploration spacecraft. The onboard computer of the deep space exploration spacecraft will combine its own flight status data obtained when it flies around the asteroid with the estimated information 12 of the asteroid structure parameters to perform Figure 4 Complete the acquisition of real-time asteroid structural parameter information 13 during the on-orbit mission and then send it to the ground measurement and control system.

[0191] The on-orbit comprehensive information 11 includes the position and velocity sequence of the on-orbit asteroid obtained in real time and output by the deep space exploration spacecraft.

[0192] Step 1: Fly around the asteroid in orbit and collect flight status information;

[0193] like Figure 4 The onboard instrument on the deep space exploration spacecraft collects all the flight states of the deep space exploration spacecraft flying around the asteroid in real time, which is recorded as the real-time flight state sampling set and set as RFSS 实时 , and RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S}, the subscript s represents the identification number of the spacecraft flight status sequence collected by the instrument in real time, and the subscript S represents the total number of spacecraft flight status collected in real time. Generally, the value of S is 1024.

[0194] Step 2: Using DNNs model optimization processing;

[0195] RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S} and the Gaussian sampling vector sequence GSS generated by the instrument are input from the right end of the DINNs model. Then the real-time collected spacecraft flight state sequence RFSS and Gaussian sampling vector sequence GSS are encoded, and the encoded vector sequence is output as the spacecraft flight state encoding sequence RFSS * and Gaussian sampling code sequence GSS * Then, each vector in the sequence is sequentially passed from the right end through 10 reversible nonlinear mapping modules with the same configuration.

[0196] The process of estimating the structural parameters of asteroids corresponds to the reverse process of the trained DINNs model. The vector operations in the reversible nonlinear mapping module in the reverse process are inverse to the vector operations in the reversible nonlinear mapping module in the forward process. For each module, the variables (vectors) are first input on the right side. The first layer on the right side is the real-time acquisition of the spacecraft flight status code sequence RFSS. * and Gaussian sampling code sequence GSS * , the middle part is divided into two independent parts (vectors), named as the first right-end input subpart and the second right-end input subpart. Two operations are performed on the first right-end input subpart, and one operation is performed on the second right-end input subpart. First, the first right-end input subpart is directly used as the first left-end output subpart without any processing. Second, the first right-end input subpart is subtracted from the second right-end input subpart through the fully connected layer to form the second left-end output subpart. The fully connected layer consists of three linear mapping layers and two nonlinear activation layers (Gelu activation), and the linear mapping layers and nonlinear activation layers are placed alternately. The first left-end output subpart (vector) and the second left-end output subpart (vector) are spliced ​​into the left-end output variable (vector), and then the left-end output variable is used as the right-end input variable to enter the next reversible nonlinear mapping module, and then the right-end input variable is used as the right-end input variable of the next reversible nonlinear mapping module for the same operation. The above process is performed 10 times in total, and then the internal structure parameters (ρ) of the asteroid obtained by the DINNs model after obtaining the real-time flight status data collected by the instrument are obtained. crust ,ρ mantle ,a mantle ,b mantle ,c mantle ) is a direct estimate of .

[0197] RFSS represents a set of flight status sequences collected by the spacecraft in real time.

[0198] RFSS * Represents the set of flight status sequences collected by the spacecraft in real time after being encoded.

[0199] RFS1 represents the first flight state sample of the flight state sequence set RFSS collected by the spacecraft in real time.

[0200] RFS2 represents the second flight state sampling of the flight state sequence set RFSS collected by the spacecraft in real time.

[0201] RFS s Represents the sth flight state sample of the flight state sequence set RFSS collected by the spacecraft in real time.

[0202] RFS SRepresents the Sth flight state sample of the flight state sequence set RFSS collected by the spacecraft in real time.

[0203] Step 3: encoding of the forward mapping output;

[0204] use About RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S} Carry out encoding processing to obtain the encoded real-time flight status

[0205] Step 4: Linear normalization;

[0206] right Perform linear mapping of linear normalized weighted combination to obtain the normalized real-time flight status of the asteroid, that is, the real-time asteroid structure parameters.

[0207] Step 5: Hyperparameter setting of DINNs model;

[0208] The present invention selects batch size Bs=1024 and learning rate Lr=10 -4 , the maximum training epoch is 300.

[0209] Step 6: Output the optimal solution after optimization by the DINNs model;

[0210] Record the value of each element in the real-time sampling point, and select the smallest value in each sampling point; combine the minimum values ​​of all sampling points, and select the minimum value again as the optimal solution, which is the estimated value of the asteroid structure based on the spacecraft flight data.

[0211] Example 1

[0212] This example uses Python to write a program, conduct a series of experiments, and draw graphs of the experimental results, but there are differences in hardware, floating-point number types, and whether the program is run in parallel.

[0213] The data-driven method based on deep learning is developed based on the Python deep learning library PyTorch. The floating-point number type is float32, which means 32-bit floating-point number. The gradient solution is based on the built-in calculation graph of PyTorch tensor (Tensor) calling the library function for solution. The hardware uses both CPU and GPU. The CPU of the series of experiments uses an Intel Core i7-13700K processor, and the GPU uses an NVIDIA GeForce RTX 4080 16G graphics card. The program can support parallel computing.

[0214] The DINNs model is an improvement on the INNs model, so it is meaningful to compare the experimental results of DINNs with those of the INNs model. Therefore, the purpose of this experiment is to compare the convergence effect, convergence speed and relative error of the reconstructed structural parameters of the DINNs model and the INNs model during the training process.

[0215] In this experiment, the reference ellipsoid is selected as ELI = {a mantle =0.6,b mantle =0.52,c mantle =0.46,ρ mantle =0.4,ρ crust =0.3}. First, generate the asteroid structure parameters Φ = {IS1, IS2, …, IS h ,…,IS H} combination, a total of 8192 groups, generated by random sampling of the reference ellipsoid parameter set ELI±20%, that is, H = 8192, corresponding to each group of structural parameters IS h Generate an orbit of 102.4 time units. The initial conditions of the flight state are the same. The orbit is uniformly sampled every 0.1 time unit. Each group obtains an orbit sampling sequence of length 1024, and adds Gaussian error to finally obtain, Where S=1024.

[0216] For the two sets generated above, Φ={IS1,IS2,…,IS h ,…,IS H}and Divide it into training set and test set evenly from the middle, and make the training set (Φ Train , GFSSO Train ) and the test set (Φ Test , GFSSO Test ).

[0217] Since DINNs modeling is related to timing, each set of structural parameters corresponds to a flight state sequence of equal length, so the batch size BS is set according to its physical meaning. A－D－INNs = 1024. For the INNs model, since there is no modeling time series, the neural network treats each set of data independently, so there is no physical constraint on the batch size. The best batch size BS is selected. INNs =4096. And record the loss function values ​​of each round of the INNs model and the DINNs model of the present invention and the reconstruction results of the asteroid structure parameters. In order to ensure fairness, the training and test data of the INNs and DINNs models are guaranteed to be exactly the same.

[0218] First, we compare the DINNs model and the INNs model in the test set (Φ Test , GFSSO Test). The loss functions of the two models are consistent, namely L y (L_fwd_fit, Mean Square Error, forward mean square error), L for independent learning of hidden layer Z z (L_fwd_mmd, Maximum Mean Discrepancy, forward maximum mean difference), and L to measure the accuracy of the reverse mapping x (L_bwd_mmd, Maximum Mean Discrepancy, reverse maximization of mean difference).

[0219] Experimental results (such as Figure 8 , Fig. 9 , Fig.10 As shown) compared to the L y , L z , L x The convergence values ​​of DINNs on the three losses are all lower than those of INNs. y It can be increased by two orders of magnitude, from 10 0 The magnitude is reduced to 10 -2 This shows that the temporal properties of the flight state sequence are weakened after the Encoder. Extracting representations from the flight state sequence can help improve the accuracy of the forward mapping, and can alleviate to some extent the problem of the INNs model lacking temporal modeling, leading to excessive forward MSE, and the problem that INNs is too dependent on a single moment.

[0220] In this problem, the forward mapping essentially provides physical information, and the reasoning process for solving the problem is the reverse mapping, that is, L x Loss. x The size of L determines whether the combination of structural parameters of the asteroid interior can be inverted from the flight state sequence or characterization. x In the comparison of (L_mmd_back), the convergence effect of DINNs model is slightly better than that of INNs, with an improvement of about 0.3. Since there is no clear linear relationship between MMD loss and relative error, additional experiments are still needed to improve the relative error of the estimate.

[0221] Experimental Results Fig.11 The DINNs model is shown to have good performance in three asteroid structural parameters (ρ crust ,b mantle ,c mantle ) is significantly better than the INNs model in terms of the average relative error. crust The parameter estimation error is reduced by 65.3% compared with the original INNs model. mantle Compared with the original INNs model, the estimation error is reduced by 25.2%, cmantle Compared with the original INNs model, the estimation error is reduced by 11.4%. In the analysis of this work, the dynamic model is a two-layer ellipsoid, with the mantle as the inner layer and the crust as the outer layer. The DINNs estimation results show that the DINNs model can help estimate the geometric shape of the mantle layer ellipsoid and the density of the crust layer. In terms of the other two physical parameters, the INNs model and the DINNs model proposed in the present invention have similar effects.

[0222] Experimental results ( Fig.12 , Fig.13 , Fig.14 , Fig.15 , Fig.16 ) is the DINNs model using the test set flight state Gaussian error sequence data GFSSO Test For the corresponding asteroid structural parameter Φ Test The color in the figure represents the density of points, bright yellow is dense, dark purple is sparse, and the vertical axis represents the density of points in the test set Φ Test The relative error of the estimated results of asteroid structural parameters is shown above. The horizontal axis represents the distribution range of asteroid structural parameters. The black dotted line is the regression line, which shows the trend of the estimated error.

[0223] The experimental results show that in all Φ Test The upper bound of the upper estimation error is 5%, for parameter ρ crust The estimation result of is the best, and its upper bound is 1%.

[0224] In summary, DINNs learns the flight state sequence Linear representation of (encoding layer) The estimation accuracy of asteroid structural parameters can be significantly improved. It also shows that the INNs model can make up for the lack of temporal sequence in modeling flight states by learning representation (coding layer).

Claims

1. A method for predicting the structure of an asteroid flying around an asteroid in orbit using an asteroid structure estimation model based on spacecraft flight data, firstly using spacecraft flight data to train and construct an asteroid structure estimation model on the ground, and then using the asteroid structure estimation model to estimate the structure of an asteroid in orbit, characterized by: The asteroid structure estimation model based on spacecraft flight data includes the following steps: Construction step 1, asteroid dynamics modeling; Construction step 11, characterization of the two-layer ellipsoid model; Based on the asteroid body coordinate system O-XYZ, a layered dynamics model of the asteroid is established. The dynamics model uses a two-layer ellipsoid model as a simplified dynamics model of the asteroid. The asteroid dynamics model is based on a two-layer ellipsoid model; its gravitational potential U AES (x, y, z) is obtained by adding the inner ellipsoid and the outer ellipsoid: U AES (x,y,z)=U crust (x,y,z)+U mantle (x,y,z) (1) U crust (x, y, z) represents the gravitational potential of the inner ellipsoid to any point in the space with coordinates O-XYZ; U mantle (x, y, z) represents the gravitational potential of the outer ellipsoid on any point in the space with coordinates O-XYZ; The gravitational potential energy of the outer ellipsoid can be calculated by subtracting the gravitational potential energy of the inner ellipsoid, which also requires 4 independent variables ρ crust ,a crust ,b crust ,c crust , specifically expressed as: MU crust represents the gravitational potential energy of the ellipsoidal core; U crust (ρ crust ,a crust ,b crust ,c crust ) represents the gravitational potential energy of the shell ellipsoid; U mantle (ρ mantle ,a mantle ,b mantle ,c mantle ) represents the gravitational potential energy of the mantle ellipsoid; ρ crust represents the density of the shell ellipsoid layer; a crust represents the a1 axis in the main axis of the shell ellipsoid layer; b crust , represents the b1 axis in the main axis of the shell ellipsoid layer; c crust represents the c1 axis in the main axis of the shell ellipsoid layer; ρ mantle represents the density of the mantle ellipsoid layer; a mantle represents the a1 axis in the main axis of the mantle ellipsoid layer; b mantle represents the b1 axis in the main axis of the mantle ellipsoid layer; c mantle represents the c1 axis in the main axis of the mantle ellipsoid layer; Construction step 12, obtaining asteroid physical parameter training set; The asteroid physical parameter set is set to Φ, and Φ={IS1,IS2,…,IS h ,…,IS H }, the subscript h represents the identification number of the asteroid structure parameter set, and the subscript H represents the total number of asteroid structure parameter sets; IS1 represents the first set of asteroid structure parameter sets; IS2 represents the second set of asteroid structural parameter sets; IS h represents the hth set of asteroid structural parameters; IS H represents the last set of asteroid structural parameters; Belongs to information set IS h The structural parameters of the asteroid are expressed in the form of a set: The subscript i indicates that it belongs to the information set IS h The identification number of the asteroid structure parameters; Indicates that it belongs to the information set IS h The first asteroid structure parameters; Indicates that it belongs to the information set IS h The second asteroid structure parameters; Indicates that it belongs to the information set IS h The structural parameters of the ith asteroid; Indicates that it belongs to the information set IS h The j-th asteroid structure parameters; Indicates that it belongs to the information set IS h The 1024th asteroid structural parameter, also from the information set IS h The last asteroid structural parameters; Also known as information set IS h Any asteroid structural parameter; The five-tuple contained in Construction step 2: Generate the orbital information training set required by the DINNs model; Construct step 21, obtain the flight status training set of the asteroid; All the flight states of deep space exploration spacecraft flying around asteroids are recorded as spacecraft-asteroid-training set FSSO, and FSSO = {FSS1, FSS2,…, FSS h ,…,FSS H }, the subscript h represents the identification number of the asteroid structure parameter set, and the subscript H represents the total number of asteroid structure parameter sets; FSS1 represents the asteroid flight state set corresponding to the first group of asteroid structure parameter set IS1; FSS2 represents the asteroid flight state set corresponding to the second group of asteroid structure parameter set IS2; FSS h represents the hth group of asteroid structural parameter set IS h The corresponding asteroid flight state set; FSS H Represents the last set of asteroid structural parameter set IS H The corresponding asteroid flight state set; Flight Status Set (FSS) h The asteroid flight state parameters are expressed in set form as follows: The subscript f indicates the sampling point identification number, and the subscript S indicates the total number of sampling points; in general, S is 1024; Indicates that it belongs to the flight status information set FSS h The first flight state sampling; Indicates that it belongs to the flight status information set FSS h The second flight state sampling; Indicates that it belongs to the flight status information set FSS h The f-th flight state sampling; for the convenience of explanation, the Also known as Flight Status Set (FSS) h Any flight state sampling; Indicates that it belongs to the flight status information set FSS h The s-th flight state sampling; Indicates that it belongs to the flight status information set FSS h The 1024th flight status sample also belongs to the flight status information set FSS h The last flight state sampling; Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis; Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis; Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis; Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis; Said The information contained is position and velocity, that is, In flight state location, In flight state The velocity component on the X-axis, In flight state The velocity component on the Y axis, In flight state The velocity component on the Z axis; Construct step 22, pairing of parameter space; The opposite of parameter space is state space, where the position and velocity parameters of the detector are The gravitational field of the asteroid is Φ = {IS1, IS2, …, IS h ,…,IS H }, so and There is a complex mapping based on physical rules. Without considering the rotation angular velocity ω of the central celestial body, the complex mapping differential equation is formalized as: Derivative of the flight data vector representing a deep space exploration spacecraft; It represents the derivative of the X-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system with respect to time; It represents the derivative of the Y-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system with respect to time; It represents the time derivative of the Z-axis coordinate of the deep space exploration spacecraft in the O-XYZ coordinate system; It represents the time derivative of the velocity component of the X-axis of the deep space exploration spacecraft in the O-XYZ coordinate system; It represents the time derivative of the Y-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system; It represents the time derivative of the Z-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system; V x It represents the velocity component of the X-axis of the deep space exploration spacecraft in the O-XYZ coordinate system; V y It represents the Y-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system; V z It represents the Z-axis velocity component of the deep space exploration spacecraft in the O-XYZ coordinate system; α x It represents the acceleration component of the deep space exploration spacecraft on the X-axis in the O-XYZ coordinate system; α y It represents the Y-axis acceleration component of the deep space exploration spacecraft in the O-XYZ coordinate system; α z It represents the acceleration component of the deep space exploration spacecraft on the Z axis in the O-XYZ coordinate system; Represents the partial differential operator; U(x,y,z,Φ) represents the gravitational potential at any position (x,y,z) in the O-XYZ coordinate system; It means to find the partial derivative of x; It means to find the partial derivative of y; It means to find the partial derivative of z; According to formula (3) and Mapping, get the mapped spacecraft flight status According to formula (3) and Mapping, get the mapped spacecraft flight status According to formula (3) and Mapping, get the mapped spacecraft flight status According to formula (3) and Mapping, get the mapped spacecraft flight status Construction step 23, addition of Gaussian error; In the training process of model construction, any Add errors to enhance the robustness of the model building process; After adding the Gaussian error GAU, the flight training set of asteroids is formed, denoted as express The flight state after Gaussian error GAU; The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, is the velocity component on the Z axis after adding the Gaussian error GAU; express The flight state after Gaussian error GAU; The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, is the velocity component on the Z axis after adding the Gaussian error GAU; express The flight state after Gaussian error GAU; The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, is the velocity component on the Z axis after adding the Gaussian error GAU; express The flight state after Gaussian error GAU; The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, is the velocity component on the Z axis after adding the Gaussian error GAU; express The flight state after Gaussian error GAU; The information included is To add the Gaussian error GAU, is the velocity component on the X-axis after adding the Gaussian error GAU, is the velocity component on the Y axis after adding the Gaussian error GAU, is the velocity component on the Z axis after adding the Gaussian error GAU; FSSO={FSS1,FSS2,…,FSS h ,…,FSS H After the Gaussian error GAU, the coded spacecraft-asteroid-flight status code set is formed, which is denoted as GFSSO = {GFSS1, GFSS2, …, GFSS h ,…,GFSS H }; Construction step three, based on the improvement of INNs architecture and parameter training adapted to asteroid structures; Based on the INNs architecture As the left variable of the asteroid structure prediction network; Based on the INNs architecture As the first variable on the right side of the asteroid structure prediction network; Based on the INNs architecture, the Gaussian distributed random sampling GSS is used as the second variable on the right side of the asteroid structure prediction network; Constructing step 31, compensation of the neural network; The mapping logic of the neural network is: Represents the forward mapping of INNs; Indicates that Map to The flight status obtained above; Constructing step 32, encoding of the forward mapping output; To this end, the forward mapping output of the INNs model is improved to increase the encoding process for the flight state sequence Perform encoding processing; express Encoded flight status; σ(·) represents a linear mapping; β s,1 represents the weight of the flight status of the sth sampling point to the flight status of the 1st sampling point; β s,2 represents the weight of the flight status of the sth sampling point to the flight status of the second sampling point; β s,f represents the weight of the flight state of the s-th sampling point to the flight state of the f-th sampling point; β s,s represents the weight of the flight state of the s-th sampling point to the flight state of the s-th sampling point; β s,S represents the weight of the flight state of the sth sampling point to the flight state of the last sampling point; Construction step 33, linear normalization; After learning the linear transformation of matrices Q and K, we can get: Represents the flight status after learning by the first learning matrix; Represents the flight status after learning by the second learning matrix; Construction step 4: training based on DINNs model; Construction step 41, selection of training set and test set; Set Φ = {IS1, IS2, …, IS h ,…,IS H } and GFSSO={GFSS1,GFSS2,…,GFSS h ,…,GFSS H } Divide it into two parts equally from the middle, the first half is named training set and the second half is named test set, and there is no overlap between the training set and the test set; Constructing step 42, determining the hyperparameters of the DINNs model; Determine the hyperparameters of the DINNs model. The hyperparameters that need to be confirmed in the DINNs model mainly include the batch size Bs, the learning rate Lr, and the maximum training round Epoch; Select batch size Bs=1024 and learning rate Lr=10 -4 , maximum training epoch = 300; Construction step 43, DINNs model training and convergence judgment; In the DINNs model training, the neural network parameters and the encoding layer parameters are randomly initialized. After each training round, the model obtained in this round of training will be saved. The models obtained from all training rounds are recorded as a set TMS = {DINNs1, DINNs2, ..., DINNs p ,…,DINNs 300 }, record the loss function values ​​of the training set and test set corresponding to each round, and record them as the loss function set of the whole round training set TRLS = {TRL1,TRL2,…,TRL g ,…,TRL 300 } and the set of trend functions of the full round test set TELS = {TEL1, TEL2, …, TEL k ,…,TEL 300 }; DINNs1 represents the model output after the first training round; DINNs2 represents the model output after the second training round; DINNs p represents the model output after the pth training round; DINNs 300 Indicates the model output after the last training round; TRL1 represents the loss function value of the training set in the first training round; TRL2 represents the loss function value of the training set in the second training round; TRL g Represents the loss function value of the training set in the g-th training round; TRL 300 Indicates the loss function value of the training set in the last training round; TEL1 represents the loss function value of the test set in the first training round; TEL2 represents the loss function value of the test set in the second training round; TEL k Represents the loss function value of the test set in the g-th training round; TEL 300 Indicates the loss function value of the test set in the last training round; Then determine whether the model has converged; this requires examining two conditions, one of which is TRL g As g increases, it approaches a constant CC; second, CC is less than TRL1; if the above two conditions are met, the DINNs model training is considered to have converged; if the above conditions are not met, the model is considered to have diverged, so the neural network and encoding layer parameters are re-initialized and re-trained until the convergence conditions are met; Constructing step 44, obtaining the optimal solution of the DINNs model; Compare The loss function value of each model in the test set, that is, the comparison set TELS = {TEL1, TEL2, …, TEL k ,…,TEL 300 }, select the smallest value among them and record it as TEL Tar , and the optimal solution DINNs in the TMS set Tar Output as an estimate of the asteroid's structure.

2. The method for predicting the structure of an asteroid flying around an asteroid in orbit using the asteroid structure estimation model based on spacecraft flight data according to claim 1, characterized in that: The prediction of asteroid structure by a spacecraft flying around an asteroid in orbit includes the following steps; Step 1: Fly around the asteroid in orbit and collect flight status information; The onboard instruments on the deep space exploration spacecraft collect all the flight status of the deep space exploration spacecraft flying around the asteroid in real time, which is recorded as the real-time flight status sampling set and set as RFSS 实时 , and RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S }, the subscript s represents the identification number of the spacecraft flight status sequence collected by the instrument in real time, and the subscript S represents the total number of spacecraft flight status collected in real time. In general, the value of S is 1024; Step 2: Use DNNs model optimization processing; RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S } and the Gaussian sampling vector sequence GSS generated by the instrument are input from the right end of the DINNs model; then the spacecraft flight state sequence RFSS and the Gaussian sampling vector sequence GSS collected in real time are encoded, and the encoded vector sequence is output as the spacecraft flight state encoding sequence RFSS * and Gaussian sampling code sequence GSS * ; Then, each vector in the sequence is sequentially passed from the right end through 10 reversible nonlinear mapping modules with the same configuration; The process of estimating the asteroid structure parameters corresponds to the reverse process of the trained DINNs model; The vector operation in the reversible nonlinear mapping module in the reverse process is inverse to the vector operation in the reversible nonlinear mapping module in the forward process; it is performed at least 10 times, and then the internal structure parameters (ρ) of the asteroid are obtained by the DINNs model after the real-time flight status data of the instrument is acquired. crust ,ρ mantle ,a mantle ,b mantle ,c mantle ) is a direct estimate of RFSS represents the set of flight status sequences collected by the spacecraft in real time; RFSS * It represents the flight status sequence set collected by the spacecraft in real time after being encoded; RFS1 represents the first flight state sampling of the flight state sequence set RFSS collected in real time by the spacecraft; RFS2 represents the second flight state sampling of the flight state sequence set RFSS collected in real time by the spacecraft; RFS s It represents the sth flight state sample of the flight state sequence set RFSS collected by the spacecraft in real time; RFS S It represents the Sth flight state sample of the flight state sequence set RFSS collected by the spacecraft in real time; Step 3: encoding of the forward mapping output; use right RFSS 实时 ={RFS1,RFS2,…,RFS s ,…,RFS S } Carry out encoding processing to obtain the encoded real-time flight status Step 4: Linear normalization; right Perform linear mapping of linear normalized weighted combination to obtain the normalized real-time flight status of the asteroid, that is, the real-time asteroid structure parameters. Step 5: Hyperparameter setting of DINNs model; Select batch size Bs=1024 and learning rate Lr=10 -4 , maximum training epoch = 300; Step 6: Output the optimal solution after optimization by the DINNs model; Record the value of each element in the real-time sampling point, and select the smallest value in each sampling point; combine the minimum values ​​of all sampling points, and select the minimum value again as the optimal solution, which is the estimated value of the asteroid structure based on the spacecraft flight data.

3. The method for predicting the structure of an asteroid flying around an asteroid in orbit using the asteroid structure estimation model based on spacecraft flight data according to claim 1, characterized in that: The trained DINNs model is stored in the onboard computer of the deep space exploration spacecraft.

4. The method for predicting the structure of an asteroid flying around an asteroid in orbit using the asteroid structure estimation model based on spacecraft flight data according to claim 1, characterized in that: The asteroid structure prediction model first receives the on-orbit comprehensive information output by the deep space exploration spacecraft; secondly, it outputs the estimated information of the asteroid structure parameters to the deep space exploration spacecraft; the onboard computer of the deep space exploration spacecraft will obtain its own flight status data while orbiting the asteroid and combine it with the estimated information of the asteroid structure parameters to perform real-time asteroid structure parameter information acquisition during the on-orbit mission, and then send it to the ground measurement and control device.

Images