CN111259604A - High orbit satellite light pressure model identification method and system based on machine learning - Google Patents
High orbit satellite light pressure model identification method and system based on machine learning Download PDFInfo
- Publication number
- CN111259604A CN111259604A CN202010057017.6A CN202010057017A CN111259604A CN 111259604 A CN111259604 A CN 111259604A CN 202010057017 A CN202010057017 A CN 202010057017A CN 111259604 A CN111259604 A CN 111259604A
- Authority
- CN
- China
- Prior art keywords
- light pressure
- orbit satellite
- pressure model
- preset
- learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a high orbit satellite light pressure model identification method and system based on machine learning, and relates to the field of aerospace. The method comprises the following steps: determining an original characteristic; performing main feature analysis on all original features to obtain features for modeling; constructing a learning set according to the main characteristics, and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model; optimizing the hyper-parameters of the high orbit satellite light pressure model; and calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite. The method is suitable for complex light pressure perturbation force modeling in space flight, can improve the efficiency of subsequent calculation, generates a targeted light pressure model according to data of different satellites, has wide application range, can manage high-dimensional data sets in a short time and has high precision, and effectively solves the contradiction between the improvement of the calculation precision and the increase of the calculation amount requirement of a machine learning algorithm.
Description
Technical Field
The invention relates to the field of aerospace, in particular to a high-orbit satellite light pressure model identification method and system based on machine learning.
Background
The solar light pressure perturbation force of the medium and high orbit navigation satellite is the largest non-conservative perturbation force except the earth gravity and the sun and moon gravity, and is related to the perturbation force, a satellite attitude control strategy, the surface material property of a satellite body and the like. Due to the influences of solar activity, satellite attitude control errors, aging of materials on the surface of a satellite and the like, the photovoltage perturbation force is difficult to accurately model, and becomes a main error source in the determination of the precise orbit of the navigation satellite.
At present, a satellite light pressure model can be divided into an analysis type and an experience type, wherein the analysis type light pressure model divides the surface of a satellite into a plurality of parts according to the satellite star body structure and the reflection and absorption characteristics of the surface material of the satellite, respectively calculates the light pressure perturbation force components of each part, and then sums the results to obtain the light pressure perturbation force. The empirical light pressure model is obtained by searching optimal estimated parameters based on a large number of fitting polynomials for on-orbit data of the satellite for a long time, and requires years of actual on-orbit data of the satellite.
However, the light pressure radiation is closely related to the parameter states of the satellites, obvious individual differences exist, the existing analysis models cannot generate targeted light pressure models according to data of different satellites, and the application range is narrow.
Disclosure of Invention
The invention aims to solve the technical problem of the prior art and provides a high orbit satellite light pressure model identification method and system based on machine learning.
The technical scheme for solving the technical problems is as follows:
a high orbit satellite light pressure model identification method based on machine learning comprises the following steps:
determining n original characteristics for evaluating the light pressure of the high orbit satellite, wherein n is more than or equal to 2;
performing main feature analysis on all the original features through a preset unparameterized feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
constructing a learning set according to the m main features, and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
optimizing the hyper-parameters of the high orbit satellite light pressure model by a preset optimization method;
and calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite.
The invention has the beneficial effects that: the invention provides a high orbit satellite light pressure model identification method based on machine learning, which is suitable for complex light pressure perturbation power modeling in space flight.
Another technical solution of the present invention for solving the above technical problems is as follows:
a machine learning-based high-orbit satellite light pressure model identification system, comprising:
the acquisition unit is used for acquiring n determined original characteristics for evaluating the light pressure of the high orbit satellite, wherein n is more than or equal to 2;
the dimension reduction unit is used for carrying out main feature analysis on all the original features through a preset non-parametric feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
the modeling unit is used for constructing a learning set according to the m main features and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
the optimization unit is used for optimizing the hyper-parameters of the high orbit satellite light pressure model through a preset optimization method;
and the calculating unit is used for calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite.
The high orbit satellite light pressure model identification system based on machine learning is suitable for complex light pressure perturbation power modeling in space flight, the dimension reduction unit can improve the efficiency of subsequent calculation by performing dimension reduction processing on original characteristics, the modeling unit performs data fitting through a machine learning algorithm, the optimization unit optimizes hyper-parameters, and can generate a targeted light pressure model according to data of different satellites.
Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
FIG. 1 is a schematic flow chart of a method for identifying a light pressure model of a high orbit satellite based on machine learning according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of feature weight comparison results provided by other embodiments of the machine learning-based high-orbit satellite light pressure model identification method of the present invention;
fig. 3 is a structural framework diagram provided by an embodiment of the high-orbit satellite light pressure model identification system based on machine learning according to the present invention.
Detailed Description
The principles and features of this invention are described below in conjunction with the following drawings, which are set forth to illustrate, but are not to be construed to limit the scope of the invention.
The invention adopts a model-free estimation algorithm, and a known observation data set is assumed to be (x)1，y1)，(x2，y2)，...，(xN，yN) And f (X, y), where y is an implicit functional relationship, and the purpose of regression is to obtain the functional relationship. However, unlike the conventional method, the present invention does not require the form of the above function, such as linearity, polynomial, exponential, fourier series, etc., but directly uses it as a black box, and directly describes the input-output relationship of the black box through the existing observation data set, and the following detailed description is provided belowAnd (4) explanation.
As shown in fig. 1, a schematic flow chart is provided for an embodiment of a machine learning-based high-orbit satellite light pressure model identification method according to the present invention, and the high-orbit satellite light pressure model identification method includes:
s1, determining n original characteristics for evaluating the light pressure of the high orbit satellite, wherein n is more than or equal to 2;
for example, the raw features may be determined empirically and from data monitoring satellite motion, and may include, for example, time, sun position vector, satellite velocity vector, satellite-to-sun line heading angle, satellite attitude quaternion, and other parameters related to solar pressure perturbation force.
S2, performing main feature analysis on all original features through a preset non-parametric feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
it should be noted that, for the determined original features, the influences are different, and therefore, the features with larger influences can be used as main features by performing feature selection on the original features, and subsequent modeling is performed, so that the dimensionality of a training set is greatly reduced, and the training efficiency is improved.
For example, the main feature analysis may be performed by a method such as a domain component analysis.
S3, constructing a learning set according to the m main features, and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
for example, the learning may be performed by a machine learning algorithm such as a support vector machine, a neural network, or the like.
It should be understood that the input of the model of the light pressure of the high orbit satellite is the main feature, and the output is the light pressure information of the high orbit satellite, for example, the light pressure information may be the component of the light pressure acceleration in a certain coordinate system, such as the inertial coordinate system of J2000.
S4, optimizing the hyper-parameters of the high orbit satellite light pressure model by a preset optimization method;
and S5, calculating the light pressure information of the high orbit satellite according to the optimized light pressure model of the high orbit satellite.
The invention provides a high orbit satellite light pressure model identification method based on machine learning, which is suitable for complex light pressure perturbation power modeling in space flight.
Optionally, in some embodiments, the method includes performing principal feature analysis on all original features by using a preset non-parametric feature selection method to obtain m principal features for modeling, and specifically includes:
analyzing the weight of the light pressure influence of the high orbit satellite with all the original characteristics by a field component analysis method to obtain the weight of the light pressure perturbation acceleration of each original characteristic in a preset direction;
and determining m main features for modeling according to the weight of each original feature in the preset direction.
It should be noted that, the domain Component Analysis (NCA) is a non-parametric feature selection method, and the core of the method is to optimize the weight coefficients of each feature to reflect the influence of each feature on the output result. The optimized objective function is typically set to leave an error on the average of the training set.
Consider a training set of
S＝{(xi，yi)，i＝1，2，…，n}
Wherein xi∈Rp，y∈R。
Define any two points xiAnd xjThe weighted distance function between is:
consider the following regression stochastic model:
randomly selecting a sample from a training set S as a reference of x given an x, and recording as Ref (x); while the output of x is equal to the output of ref (x). It is assumed that the process of selection is related to the weighted distance. Then xjThe probability of a reference being selected as x is:
here, k is a kernel function, and is used to describe the similarity of two points, i.e. the smaller the weight distance is, the larger the value of the kernel function is.
Consider next the "leave one" operation on the above process. Let x be removediIs recorded as S-iThen xjIs selected as xiThe probabilities of the references of (1) are:
order to
For the above regression stochastic model pair xiPredicted value of response, yiIs a true response. The difference between the two is defined, i.e. the loss function is l: r2→ R. And may further relate to xiAverage loss function of (2):
wherein p isijRepresentative selection xjAs xiThe probability of the reference of (2).
Finally, we can obtain the objective function of the whole training set as:
the second term is a regular term, and represents the balance between the prediction accuracy and the complexity of the prediction function, so that overfitting is avoided.
To this end, the NCA method is actually an unconstrained optimization problem:
minwf(w)。
optionally, in some embodiments, the determining m main features for modeling according to the magnitude of the weight of each original feature in the preset direction specifically includes:
carrying out normalization processing on the weights of all the original features in the preset direction;
respectively comparing the weight of each original feature in the ith preset direction with the ith preset threshold;
and taking the original features of which the weight average in each preset direction is smaller than the corresponding preset threshold as main features.
It should be noted that the preset threshold may be set according to actual requirements.
The following is a description of a specific example.
Taking GPS satellite Block II as an example, the near-location altitude is around 300km, the far-location altitude is around 36000km, and the generated light pressure data is expressed in the J2000 inertial system by extrapolation for one year.
16 original features were selected as shown in table 1. It should be understood that the position and velocity vectors in these features are defined relative to the J2000 coordinate system, the pose is defined relative to J2000, and the illumination vector is defined relative to the orbital plane. These features are not completely independent, for example, the illumination line-of-sight angle can be calculated from the sun position and the satellite position. Therefore, the positions of these quantities in the light pressure model are also different.
TABLE 1
| Feature(s) | Numbering | Feature(s) | Numbering |
| Time of |
1 | High and low angle of illumination sight | 11 |
| X component of |
2 | Azimuth angle of illumination line of sight | 12 |
| Y component of sun position | 3 | Quaternion of satellite attitude | 13 |
| Z component of sun position | 4 | Quaternion of |
14 |
| Satellite position x component | 5 | Quaternion of satellite attitude | 15 |
| Y component of |
6 | Quaternion of |
16 |
| Satellite position z component | 7 | ||
| X component of satellite velocity | 8 | ||
| Y component of satellite velocity | 9 | ||
| Z component of |
10 |
The weights of the influence of the original features are estimated through a domain composition analysis method and normalized, and the obtained result is shown in fig. 2, wherein in fig. 2, the influence of each feature on the light pressure perturbation acceleration in three directions under a J2000 coordinate system is given, and in order to better represent the influence of different features, the median mean and the average mean of the weight coefficients are also given in fig. 2.
It can be seen from fig. 2 that the influence of the individual features is different. For example, the influence weight of the temporal light pressure model is almost zero, and the influence of the satellite velocity is small. In addition, the influence of the satellite attitude quaternion is also different, wherein the influence of the last component is the largest, and the influence of the yaw attitude of the GPS satellite is mainly reflected. Looking at the different perturbation acceleration components we can see that the effect of the features is similar, but there are some differences. The y-component of the sun's position has little effect on the perturbation acceleration in the z-direction and a greater effect on the x-and y-directions.
Having clarified these effects, the dimensionality of the training set can be greatly reduced. For example, the initial training set dimension is 16, while the reduced training set dimension may be 8-10. This may further improve the efficiency of the training.
Optionally, in some embodiments, a learning set is constructed according to the m main features, and a preset machine learning algorithm is learned through the learning set to obtain the high orbit satellite light pressure model, which specifically includes:
taking m main features as input, taking light pressure acceleration information as output, acquiring a preset number of learning data, and dividing the learning data into a training set and a verification set to obtain a learning set, wherein the learning data comprises the main features and the light pressure acceleration information;
and training the support vector machine through a training set to obtain a high orbit satellite light pressure model, and verifying through a verification set.
For example, the learning set may be:
inputting: sun position vector, satellite velocity vector, satellite-sun connecting line direction angle and satellite attitude quaternion;
and (3) outputting: the classification and amplitude of the light pressure acceleration in the J2000 coordinate system can be superposed with a random error of 15%.
Learning data of one year is obtained, and the learning data is randomly divided into two parts, wherein the first part is 20 weeks and is used as a training set, and the second part is 32 weeks and is used as a check set.
As the SVM can obtain a better result than other algorithms on a small sample training set, the storage and calculation resources can be saved, and the requirements on data scale and data distribution are reduced.
It should be noted that the support vector machine is also a model-free method. The support vector machine is essentially a linear classifier, and the invention mainly considers an epsilon-loss function. The goal of regression is to make the predicted and true observations deviate by no more than epsilon, while making the line more likely to be flat.
For regression of nonlinear functions, the idea of SVM is to project it into a high dimensional space, to perform a linear fit, i.e.:
here, the
A non-linear projection function.
Only the following are needed in the solution:
without the need to know the non-linear function
Itself. Therefore, it is only necessary to define the inner product in the high-dimensional space, and the commonly used inner products are:
G(x1，x2)＝x′1x2
G(x1，x2)＝exp(-||x1-x2||2)
G(x1，x2)＝(1+x′1x2)p
respectively, linear inner product, gaussian inner product and polynomial inner product.
And then a dual problem can be constructed and optimized, and the Lagrangian function is as follows:
the constraints are:
the KKT condition is:
after the optimization solution, a nonlinear fitting function can be obtained as follows:
optionally, in some embodiments, the method for optimizing the hyper-parameters of the high orbit satellite light pressure model by using a preset optimization method specifically includes:
determining k groups of hyperparameters, and obtaining alternative high orbit satellite light pressure models corresponding to each group of hyperparameters according to a Bayesian optimization method based on Gaussian process regression, wherein k is more than or equal to 2;
performing cross validation on all the obtained alternative high-orbit satellite light pressure models, and evaluating the performance of each alternative high-orbit satellite light pressure model;
and taking the alternative high-orbit satellite light pressure model with the performance meeting the preset condition as the high-orbit satellite light pressure model after the hyper-parameter optimization.
It should be noted that, for each set of parameters, calculating the SVM model requires solving an optimization problem, which requires a large amount of calculation. Therefore, the Bayesian optimization method is adopted when the hyper-parameters are optimized, the computational burden of the hyper-parameter optimization is reduced by adopting the Bayesian optimization idea, and the computational accuracy can be further improved.
Bayesian optimization methods aim at minimizing a scalar function f (x), where x is located within a bounded region. The objective function f (x) here may be a deterministic function or a random function, and the components of x may be continuous or discrete variables.
Bayesian optimization is mainly used to handle cases where a single calculation f (x) requires a higher computational/economic cost. For this purpose, Bayesian optimization models f (x) as a Gaussian process, with each iteration aiming to predict a new x*So that f (x)*) Lower than f (x). At the same time, f (x) is obtained*) The said gaussian process will be also corrected a posteriori probability.
An acquisition function a (x) and (acquisition function) are adopted in the process of predicting the new point to realize the balance of local and global exploration.
The specific algorithm steps are as follows:
1. calculating xiAnd corresponding yi＝f(xi)；
2. Updating the Gaussian process model to obtain a new posterior probability Q (f | x)i，yi)
3. Obtaining a new point x by maximizing the collection function a (x)i+1。
4. Repeat 1.
Bayesian optimization mainly depends on two factors, namely a Gaussian regression model of an objective function value; the second is the form of the acquisition function.
Assume that the prior distribution of the objective function is gaussian, the mean is μ (x, θ), and the covariance kernel is k (x, x, θ), where θ is a parameter of the kernel. Without loss of generality, the a priori mean is usually set to zero, while also taking into account certain observed noise, i.e. the a priori distribution covariance is K (X,X′，θ)+σ2I。
the acquisition function is combined with the posterior distribution of the target function to describe the quality degree of the new value taking point. The new point can be selected as the position with the maximum mean value in the posterior distribution of the objective function, which is called as 'development'; or selecting the new point as the position with the maximum variance in the posterior distribution of the objective function, which is called 'exploration'. The acquisition function represents the balance between the two. Commonly used acquisition functions are: maximize the expectation of improvement, maximize the probability of improvement, minimize the minimum confidence boundary, etc.
In Gaussian process regression, a random process can be represented by a random variable cluster { X (T, w), T ∈ T }. Arbitrarily extracting a limited number of indexes (e.g., n, t) from the random variable cluster1，...tn) The resulting vector of variables
If the joint distribution is a multidimensional Gaussian distribution, the stochastic process is called a Gaussian process.
Gaussian process regression is a kernel function-based non-parametric model and is suitable for processing complex regression problems such as high dimensionality, small samples, nonlinearity and the like.
Consider the training set as:
S＝{(xi，yi)，i＝1，2，…，n}
the objective of regression is to obtain the functional relationship y ═ f (x), where the distribution of this function is specified:
p(f|X，y)
the Gaussian process assumes p (f (x)1），…，f(xN) ) is gaussian, the mean and variance are:
the k function is a kernel function.
The gaussian process is recorded as:
f(x)～GP(m(x)，k(x，x′))
for a finite number of observations, a joint gaussian distribution can be defined:
p(f|X)＝N(f|μ，K)
wherein Kij＝k(xi，xj)，μ＝[m(x1)，…，m(xN)]T。
Due to the presence of observation noise, the functional form is assumed to be:
y＝f(x)+ε
first the covariance matrix of the observations is needed:
cov(y|X)＝K+σ2I
where σ is the distribution of observed noise.
The observed and predicted joint distribution is further constructed:
the a posteriori prediction can be found as:
it can be seen that from X to y, the intermediate variable f is involved, namely:
f～N(0，K)
y～N(f，σ2I)
the process from X to y can be expressed as p (y | X) -N (y | X, K + σ2I)。
The parameters are intended to maximize the edge likelihood function, i.e.:
the parameters herein mainly include parameters of σ and kernel functions.
Then, optimization methods such as conjugate gradient method, newton method and the like can be adopted to maximize the partial derivative so as to obtain the optimal solution of the hyper-parameter.
It is understood that some or all of the alternative embodiments described above may be included in some embodiments.
As shown in fig. 3, a structural framework diagram is provided for an embodiment of the high-orbit satellite light pressure model identification system based on machine learning according to the present invention, the high-orbit satellite light pressure model identification system includes:
the device comprises an acquisition unit 1, a calculation unit and a calculation unit, wherein the acquisition unit 1 is used for acquiring n determined original characteristics for evaluating the light pressure of the high orbit satellite, and n is more than or equal to 2;
the dimension reduction unit 2 is used for performing main feature analysis on all original features through a preset non-parametric feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
the modeling unit 3 is used for constructing a learning set according to the m main features, and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
the optimization unit 4 is used for optimizing the hyper-parameters of the high orbit satellite light pressure model by a preset optimization method;
and the calculating unit 5 is used for calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite.
The high orbit satellite light pressure model identification system based on machine learning provided by the embodiment is suitable for complex light pressure perturbation power modeling in space flight, the dimensionality reduction unit 2 can improve the efficiency of subsequent calculation by carrying out dimensionality reduction processing on original features, the modeling unit 3 carries out data fitting through a machine learning algorithm, the optimization unit 4 optimizes hyper-parameters, and can generate a targeted light pressure model according to data of different satellites.
Optionally, in some embodiments, the dimension reduction unit 2 is specifically configured to analyze, by a domain component analysis method, the weight of the light pressure influence of the high orbit satellite of all the original features, to obtain a weight of the light pressure perturbation acceleration of each original feature in a preset direction; and determining m main features for modeling according to the weight of each original feature in the preset direction.
Optionally, in some embodiments, the dimension reduction unit 2 is specifically configured to perform normalization processing on the weights of all the original features in the preset direction; respectively comparing the weight of each original feature in the ith preset direction with the ith preset threshold; and taking the original features of which the weight average in each preset direction is smaller than the corresponding preset threshold as main features.
Optionally, in some embodiments, the modeling unit 3 is specifically configured to take m main features as input, take light pressure acceleration information as output, obtain a preset number of learning data, and divide the learning data into a training set and a verification set to obtain a learning set, where the learning data includes the main features and the light pressure acceleration information; and training the support vector machine through a training set to obtain a high orbit satellite light pressure model, and verifying through a verification set.
Optionally, in some embodiments, the optimization unit 4 is specifically configured to determine k sets of hyper-parameters, and obtain a candidate high orbit satellite light pressure model corresponding to each set of hyper-parameters according to a bayesian optimization method based on gaussian process regression, where k is greater than or equal to 2; performing cross validation on all the obtained alternative high-orbit satellite light pressure models, and evaluating the performance of each alternative high-orbit satellite light pressure model; and taking the alternative high-orbit satellite light pressure model with the performance meeting the preset condition as the high-orbit satellite light pressure model after the hyper-parameter optimization.
It is understood that some or all of the alternative embodiments described above may be included in some embodiments.
It should be noted that the above embodiments are product embodiments corresponding to the previous method embodiments, and for the description of each optional implementation in the product embodiments, reference may be made to corresponding descriptions in the above method embodiments, and details are not described here again.
The reader should understand that in the description of this specification, reference to the description of the terms "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described method embodiments are merely illustrative, and for example, the division of steps into only one logical functional division may be implemented in practice in another way, for example, multiple steps may be combined or integrated into another step, or some features may be omitted, or not implemented.
While the invention has been described with reference to specific embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A high orbit satellite light pressure model identification method based on machine learning is characterized by comprising the following steps:
determining n original characteristics for evaluating the light pressure of the high orbit satellite, wherein n is more than or equal to 2;
performing main feature analysis on all the original features through a preset unparameterized feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
constructing a learning set according to the m main features, and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
optimizing the hyper-parameters of the high orbit satellite light pressure model by a preset optimization method;
and calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite.
2. The method for identifying the light pressure model of the high orbit satellite based on machine learning of claim 1, wherein the method for identifying the light pressure model of the high orbit satellite based on machine learning is characterized in that a preset unparameterized feature selection method is used for performing main feature analysis on all the original features to obtain m main features for modeling, and specifically comprises the following steps:
analyzing the weight of the light pressure influence of the high orbit satellite of all the original characteristics by a field component analysis method to obtain the weight of the light pressure perturbation acceleration of each original characteristic in a preset direction;
and determining m main features for modeling according to the weight of each original feature in a preset direction.
3. The method for identifying the light pressure model of the high-orbit satellite based on machine learning according to claim 2, wherein the determining m main features for modeling according to the weight of each original feature in the preset direction specifically comprises:
carrying out normalization processing on the weights of all the original features in a preset direction;
respectively comparing the weight of each original feature in the ith preset direction with the ith preset threshold;
and taking the original features of which the weight average in each preset direction is smaller than the corresponding preset threshold as main features.
4. The method for identifying the high-orbit satellite light pressure model based on machine learning according to any one of claims 1 to 3, wherein a learning set is constructed according to the m main features, and a preset machine learning algorithm is learned through the learning set to obtain the high-orbit satellite light pressure model, specifically comprising:
taking m main features as input, taking light pressure acceleration information as output, acquiring a preset number of learning data, and dividing the learning data into a training set and a verification set to obtain a learning set, wherein the learning data comprises the main features and the light pressure acceleration information;
and training a support vector machine through the training set to obtain a high orbit satellite light pressure model, and verifying through the verification set.
5. The method for identifying the light pressure model of the high-orbit satellite based on machine learning according to any one of claims 1 to 3, wherein the method for optimizing the hyper-parameters of the light pressure model of the high-orbit satellite by a preset optimization method specifically comprises:
determining k groups of hyperparameters, and obtaining alternative high orbit satellite light pressure models corresponding to each group of hyperparameters according to a Bayesian optimization method based on Gaussian process regression, wherein k is more than or equal to 2;
performing cross validation on all the obtained alternative high-orbit satellite light pressure models, and evaluating the performance of each alternative high-orbit satellite light pressure model;
and taking the alternative high-orbit satellite light pressure model with the performance meeting the preset condition as the high-orbit satellite light pressure model after the hyper-parameter optimization.
6. A high orbit satellite light pressure model identification system based on machine learning, comprising:
the acquisition unit is used for acquiring n determined original characteristics for evaluating the light pressure of the high orbit satellite, wherein n is more than or equal to 2;
the dimension reduction unit is used for carrying out main feature analysis on all the original features through a preset non-parametric feature selection method to obtain m main features for modeling, wherein m is less than or equal to n;
the modeling unit is used for constructing a learning set according to the m main features and learning a preset machine learning algorithm through the learning set to obtain a high orbit satellite light pressure model;
the optimization unit is used for optimizing the hyper-parameters of the high orbit satellite light pressure model through a preset optimization method;
and the calculating unit is used for calculating the light pressure information of the high-orbit satellite according to the optimized light pressure model of the high-orbit satellite.
7. The high orbit satellite light pressure model identification system based on machine learning of claim 6, wherein the dimensionality reduction unit is specifically configured to analyze the weight of the light pressure influence of the high orbit satellite of all the original features through a domain component analysis method to obtain the weight of the light pressure perturbation acceleration of each original feature in a preset direction; and determining m main features for modeling according to the weight of each original feature in a preset direction.
8. The system for identifying an optical pressure model of a high-orbit satellite based on machine learning of claim 7, wherein the dimension reduction unit is specifically configured to perform normalization processing on the weights of all the original features in a preset direction; respectively comparing the weight of each original feature in the ith preset direction with the ith preset threshold; and taking the original features of which the weight average in each preset direction is smaller than the corresponding preset threshold as main features.
9. The high orbit satellite light pressure model identification system based on machine learning of any one of claims 6 to 8, wherein the modeling unit is specifically configured to take m main features as input, light pressure acceleration information as output, obtain a preset number of learning data, divide the learning data into a training set and a verification set, and obtain a learning set, wherein the learning data includes the main features and the light pressure acceleration information; and training a support vector machine through the training set to obtain a high orbit satellite light pressure model, and verifying through the verification set.
10. The machine learning-based high-orbit satellite light pressure model identification system according to any one of claims 6 to 8, wherein the optimization unit is specifically configured to determine k sets of hyper-parameters, obtain alternative high-orbit satellite light pressure models corresponding to each set of hyper-parameters according to a Bayesian optimization method based on Gaussian process regression, where k is greater than or equal to 2; performing cross validation on all the obtained alternative high-orbit satellite light pressure models, and evaluating the performance of each alternative high-orbit satellite light pressure model; and taking the alternative high-orbit satellite light pressure model with the performance meeting the preset condition as the high-orbit satellite light pressure model after the hyper-parameter optimization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010057017.6A CN111259604A (en) | 2020-01-16 | 2020-01-16 | High orbit satellite light pressure model identification method and system based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010057017.6A CN111259604A (en) | 2020-01-16 | 2020-01-16 | High orbit satellite light pressure model identification method and system based on machine learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111259604A true CN111259604A (en) | 2020-06-09 |
Family
ID=70948974
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010057017.6A Pending CN111259604A (en) | 2020-01-16 | 2020-01-16 | High orbit satellite light pressure model identification method and system based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111259604A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112580819A (en) * | 2020-11-24 | 2021-03-30 | 中国西安卫星测控中心 | Low-orbit satellite precise orbit determination strategy supported by machine learning |
| CN114265432A (en) * | 2021-12-20 | 2022-04-01 | 中国科学院空间应用工程与技术中心 | Satellite cluster control method and system based on earth orbit |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104021241A (en) * | 2014-05-13 | 2014-09-03 | 北京空间飞行器总体设计部 | Earthlight light pressure perturbation modeling method for navigational satellite complicated model |
| WO2015184729A1 (en) * | 2014-06-05 | 2015-12-10 | Tsinghua University | Method and system for hyper-parameter optimization and feature tuning of machine learning algorithms |
| CN105203110A (en) * | 2015-08-28 | 2015-12-30 | 中国科学院空间应用工程与技术中心 | Low-orbit-satellite orbit prediction method based on atmospheric resistance model compensation |
| CN106156401A (en) * | 2016-06-07 | 2016-11-23 | 西北工业大学 | Data-driven system state model on-line identification methods based on many assembled classifiers |
| CN107516135A (en) * | 2017-07-14 | 2017-12-26 | 浙江大学 | A kind of automation monitoring learning method for supporting multi-source data |
| CN107798187A (en) * | 2017-10-24 | 2018-03-13 | 北京理工大学 | A kind of efficiently satellite constellation Multipurpose Optimal Method |
| CN108460176A (en) * | 2018-01-02 | 2018-08-28 | 佛山科学技术学院 | A method of it improving satellite orbit perturbation power model and indicates precision |
| CN108595150A (en) * | 2018-04-27 | 2018-09-28 | 北京航空航天大学 | The method that artificial intelligence program person writes digital satellite power thermal coupling source program |
| CN110245252A (en) * | 2019-06-10 | 2019-09-17 | 中国矿业大学 | Machine learning model automatic generation method based on genetic algorithm |
-
2020
- 2020-01-16 CN CN202010057017.6A patent/CN111259604A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104021241A (en) * | 2014-05-13 | 2014-09-03 | 北京空间飞行器总体设计部 | Earthlight light pressure perturbation modeling method for navigational satellite complicated model |
| WO2015184729A1 (en) * | 2014-06-05 | 2015-12-10 | Tsinghua University | Method and system for hyper-parameter optimization and feature tuning of machine learning algorithms |
| CN105203110A (en) * | 2015-08-28 | 2015-12-30 | 中国科学院空间应用工程与技术中心 | Low-orbit-satellite orbit prediction method based on atmospheric resistance model compensation |
| CN106156401A (en) * | 2016-06-07 | 2016-11-23 | 西北工业大学 | Data-driven system state model on-line identification methods based on many assembled classifiers |
| CN107516135A (en) * | 2017-07-14 | 2017-12-26 | 浙江大学 | A kind of automation monitoring learning method for supporting multi-source data |
| CN107798187A (en) * | 2017-10-24 | 2018-03-13 | 北京理工大学 | A kind of efficiently satellite constellation Multipurpose Optimal Method |
| CN108460176A (en) * | 2018-01-02 | 2018-08-28 | 佛山科学技术学院 | A method of it improving satellite orbit perturbation power model and indicates precision |
| CN108595150A (en) * | 2018-04-27 | 2018-09-28 | 北京航空航天大学 | The method that artificial intelligence program person writes digital satellite power thermal coupling source program |
| CN110245252A (en) * | 2019-06-10 | 2019-09-17 | 中国矿业大学 | Machine learning model automatic generation method based on genetic algorithm |
Non-Patent Citations (2)
| Title |
|---|
| 李利杰等: "一种改进的支持向量机模型优化算法", 《计算机技术与发展》 * |
| 陈秋丽等: "GPS卫星光压摄动建模发展与启示", 《航天器工程》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112580819A (en) * | 2020-11-24 | 2021-03-30 | 中国西安卫星测控中心 | Low-orbit satellite precise orbit determination strategy supported by machine learning |
| CN112580819B (en) * | 2020-11-24 | 2023-06-13 | 中国西安卫星测控中心 | Machine learning supported low-orbit satellite precise orbit determination strategy |
| CN114265432A (en) * | 2021-12-20 | 2022-04-01 | 中国科学院空间应用工程与技术中心 | Satellite cluster control method and system based on earth orbit |
| CN114265432B (en) * | 2021-12-20 | 2023-12-26 | 中国科学院空间应用工程与技术中心 | Satellite cluster control method and system based on earth orbit |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Seguy et al. | Large-scale optimal transport and mapping estimation | |
| Williams et al. | Bayesian classification with Gaussian processes | |
| US20160232637A1 (en) | Shipment-Volume Prediction Device, Shipment-Volume Prediction Method, Recording Medium, and Shipment-Volume Prediction System | |
| Wenzel et al. | Efficient Gaussian process classification using Pòlya-Gamma data augmentation | |
| KR20160041856A (en) | Systems and methods for performing bayesian optimization | |
| Luttinen et al. | Bayesian robust PCA of incomplete data | |
| US20130013538A1 (en) | Recovering the structure of sparse markov networks from high-dimensional data | |
| Senanayake et al. | Spatio-temporal hilbert maps for continuous occupancy representation in dynamic environments | |
| CN111259604A (en) | High orbit satellite light pressure model identification method and system based on machine learning | |
| CN109460065A (en) | Unmanned aerial vehicle cluster formation characteristic identification method and system based on potential function | |
| Karl et al. | Multi-Objective Hyperparameter Optimization--An Overview | |
| Nguyen et al. | A scalable hierarchical Gaussian process classifier | |
| CN116522143B (en) | Model training method, clustering method, equipment and medium | |
| WO2005122066A1 (en) | Support vector classification with bounded uncertainties in input data | |
| US20200285205A1 (en) | Policy improvement method, recording medium, and policy improvement apparatus | |
| Legendre et al. | Two line element accuracy assessment based on a mixture of Gaussian laws | |
| Lee et al. | Constructing sparse kernel machines using attractors | |
| Seligmann et al. | Beyond deep ensembles: A large-scale evaluation of bayesian deep learning under distribution shift | |
| Li et al. | Stochastic feature mapping for pac-bayes classification | |
| EP4124887A1 (en) | Methods and systems for predicting trajectory data of an object and methods and systems for training a machine learning method for predicting trajectory data of an object | |
| Einbeck et al. | Representing complex data using localized principal components with application to astronomical data | |
| US20210056352A1 (en) | Computational inference system | |
| Nugroho et al. | The improvement of C4. 5 algorithm accuracy in predicting Forest fires using discretization and AdaBoost | |
| Ceylan et al. | A Comparison of differential evolution and harmony search methods for SVM model selection in hyperspectral image classification | |
| CN112200488A (en) | Risk identification model training method and device for business object |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200609 |